The AI Buildout Has to Happen Somewhere

Who gets the next AI dollar if demand keeps rising but power, permits and financing cannot keep up? At All-In Summit, we followed that question from the stage into six off-stage conversations. This report separates what the CEOs put on the record from what still needs a filing or a number.

At The Shrine in Los Angeles, Monday’s debate followed Dario Amodei’s September 12 essay, “We Must Pace the Frontier,” and the Washington Post’s headline about calls to slow AI down. Then Jensen took the stage, and Donald Trump called him. On Tuesday, the room went quiet when Elon joined the Gwynne Shotwell session by video.

The keynotes are going up on YouTube. Our subscribers had updates from the room in the chat. The additional work was outside the sessions: talking to people who build, finance and use the infrastructure, then asking what would make their claims investable. People still do business.

Below, we work through the public evidence on neocloud demand and the power bottleneck. Paid subscribers get the six anonymized conversations, the evidence needed to test each one, the bear case and how it connects to our book.

Two CEOs Put the Neocloud Hypothesis in Their Own Words

Jensen Huang told the room on Monday that the neoclouds’ first customers were the hyperscalers, and Satya Nadella, the same day, said Microsoft is renting capacity because it is short on supply. Both sessions are on the All-In Podcast’s channel, and the quotes here are checked against the published video. Both land on a call we made in March.

Jensen, on why the neoclouds exist: “all the early customers were the hyperscalers. And the reason for that is because the hyperscalers plan once a year. But the market dynamics is so volatile right now that they’re always almost wrong.” The regional clouds, he said, are now “a large scale distributed network of companies that are building, securing land, power, shell for us,” and he framed NVIDIA’s own power and site investments as a “downstream” version of its upstream supply-chain work.

As we wrote in The NeoCloud Hypothesis: “in a supply-constrained environment, speed of deployment is the scarce resource.” Our mechanism in March was hyperscalers diluting their deployment speed with their own silicon. Jensen’s is simpler: the big buyers plan annually, the market moves faster, and whoever can move a shell and a substation inside a quarter gets the first racks.

Nadella, on Microsoft’s own capacity: “right now we’re even renting quite a bit because we kind of were short on supply. But the overall goal is to build more, lease some and then if really need to surge we will even rent.”

Four days earlier, in the subscriber chat, we wrote up Bloomberg’s Microsoft road map as rationing at two gigawatts, about 2 gigawatts of AI chips inside a 12 gigawatt estate. Build, lease, rent is that road map from the CEO’s side, and its 38 gigawatts by 2032 excludes what Microsoft rents from neoclouds.

Two lines from the same day cut against us. A host put the labs’ economics to Nadella as token compression, roughly $50 per million output tokens today against an estimated floor near $0.15, which is the algorithmic-efficiency short, the one risk our clusters mark as changing direction rather than slope. And Nadella, on whose silicon Azure runs: “we have Jensen’s stuff which is I think our primary thing, we have our own, OpenAI is building their” silicon, “AMD in there.” That is our March custom-silicon bear, restated by a hyperscaler CEO.

One of the hosts said, as fact, that NVIDIA bought Hugging Face. Agreed is the word: a definitive agreement on September 3 at about $12.93 billion, filed on an 8-K, with the close expected in the first half of 2027.

We published The Only Company That Can Afford Free that same morning on press reports, and wrote that we were “not resizing a position we have held publicly since 2016 on a press report.” The 8-K printed the same day, so that piece’s first gate is confirmed, and our ledger carried it as open until this note.

Gerstner Cut the Gigawatt Forecast, and the Room Clapped

Brad Gerstner’s keynote on Tuesday cut next year’s AI compute build from SemiAnalysis’s 43 gigawatts to about 25, and three of the four reasons on his slide have been in our power file since December. The All-In Podcast published the talk on September 17 as “Brad Gerstner: No AI Bubble, Semis Eat the Nasdaq & AI’s Take Off Problem,” and his words here are from that video. His slide, sourced to SemiAnalysis and Dylan Patel on the Dwarkesh Podcast, August 25, counts global installed AI compute: 19 gigawatts added this year, 7 of them to OpenAI and Anthropic, and 43 next year, 14 to the two labs. On stage he called that base the country’s. His own number: “I think the total amount we’re actually going to stand up is somewhere closer to 25 gigawatts,” half of it for Anthropic and OpenAI. His “Atoms and Energy Are Hard” slide gives the reasons: permitting and local opposition, grid interconnection delays, skilled labor shortages, power equipment sold out.

We wrote this in the subscriber chat on Tuesday, the morning of the keynote: Gave us like 30 slides on the current state of AI investing, too numerous to summarize, just two things stood out: He’s predicting “only” 25 GW of new data center capacity coming online next year, half being used by Anthropic and OpenAI, as opposed to Dylan Patel’s 43 GW prediction. If the 10 year govt bond goes to 5.5% it’s a problem for massive capital investment (debt will dry up).

This is the mechanical constraint we laid out in The Watt Asymmetry in May: “The grid will not save the hyperscalers.” Interconnect queues and transformer backorders were that piece’s Layer 6, permitting and local opposition goes back to our December Bloom piece, and skilled labor is the box we never filed.

When Gerstner compared the treatment of data centers to nuclear plants, the room applauded. You could feel the agreement. That tells us something about the room’s conviction; it does not tell us whether the next county will approve a project.

The number he is watching is monthly lab revenue: “If the monthly AI lab revenues are closer to that $8 billion number, I think it’s takeoff.” He put the top three labs, Anthropic, OpenAI and SpaceX, at about a $100 billion collective run rate coming out of July, needing “at least $180 billion by the end of the year... just to keep the AI trade intact.” The IPO that traded down the day before on worry about a halt is Anthropic’s, and he does not expect one.

The nine minutes our recording missed were the valuation case: “This is no bubble like it was in 2000,” with Nvidia at 14 times next year’s fully taxed GAAP earnings, and semiconductors at 70% of the Nasdaq’s return. “The makers of the tokens are making the money and the buyers of the tokens are basically going along for the ride.” The slide for our Token Dollar cluster is “What has to be true? 2027 and Beyond”: hyperscaler capex of $788 billion this year and $1,086 billion next against lab AI revenue of $105 billion and $300 billion, his estimates. That gap is the carry question, drawn by a bull.

His third risk was rates. His “over 90%” on a hike the next day sat beside a slide showing about 88% market-implied odds, and the Federal Reserve delivered it the next morning, 25 basis points to 3.75 to 4.00 percent, with 16 of 18 participants expecting another this year. The ten-year answered by crossing 5 percent, 5.02 on Wednesday, and sat at 4.93 on Thursday. His level for trouble is five and a half on the ten-year, where the carry in the Token Dollar loop stops being a footnote.

Upgrade to read the full field report: six anonymized conversations, our table of the evidence that would test each claim, the polling and local-tax charts, the bear case, and how it all connects to our book. These are leads to investigate, with each source’s incentives made explicit. Paid subscribers can follow the open questions in our subscriber chat.

Read more

The Headline and the Dashboard Disagree

On September 7 published a headline that says Google’s Ironwood TPU beats NVIDIA’s Blackwell by up to 50% on performance per dollar. SemiAnalysis’s own dashboard, with each side running the way its maker actually ships it, says NVIDIA wins by 8 to 12x. Same shop, same benchmark, one day apart. We pulled the dashboard’s own numbers, and this piece is what they say.

The article is “TPU Inference Externalization Full Steam Ahead,” and the InferenceX site carries a free copy. Its section title is “TPU is King on Performance per Dollar.” The comparison under it puts Ironwood, which the dashboard labels TPU7x, against a single box of eight NVIDIA chips, running the slower of NVIDIA’s two number formats, writing one token at a time, on one pool of chips that reads and writes together instead of the split setup the rack runs. SemiAnalysis calls that apples to apples. It calls NVIDIA’s full 72-chip rack against the TPU “apples to bananas,” and keeps the rack out of the headline.

That is the part worth sitting with. The dashboard is one click from the headline. The data that contradicts the headline is SemiAnalysis’s own, run by SemiAnalysis’s own methodology, priced on SemiAnalysis’s own cost model, and in the default table the rack’s curve is sorted to the top. A benchmark shop is the referee in this fight. This time the referee published the halftime score as the final.

The question we kept asking this week: what would you do to find that common denominator for both, to make it a fair comparison?

The answer a buyer uses is simple to say. Tokens are the word-pieces a model reads and writes. How many of them do you get for a dollar, at the speed a user sees the answer come back, with each machine set up the way its maker recommends. The bill starts in that number. Read that way, on the same model and the same workload the article uses, SemiAnalysis’s dashboard has GB300 NVL72 delivering 9.7x the tokens per dollar of TPU7x when each user gets 100 tokens a second, and 11.9x at the fastest speed the TPU reaches. The discount is on the chip. The bill is per token.

What the Headline Switched Off

Four things are off on the NVIDIA side of the “apples-to-apples” chart, and all four are visible in SemiAnalysis’s own text and in the dashboard’s configuration labels. The test runs NVIDIA in eight-bit math, FP8, instead of the four-bit math, NVFP4, that Blackwell was built around, which does twice the arithmetic per second on Blackwell and three times on Blackwell Ultra, the chip inside GB300; SemiAnalysis notes Ironwood cannot run four-bit natively and the next TPU will. It turns off multi-token prediction, the trick where the chip drafts several tokens ahead and checks them in one pass, the way a good autocomplete does. It uses a box of eight chips instead of the 72-chip rack NVIDIA sells as one computer. And it serves the model aggregated, reading the question and writing the answer on the same chips, instead of disaggregated, which splits those two jobs across different chips the way a kitchen splits prep cooks from the line. That split is a serving choice, not a software brand: vLLM and SGLang can run it too, and the article’s setup does not.

Each switch has a price on the same dashboard. The multi-token trick alone is worth about 2.3x on the rack when each user gets 100 tokens a second. Going from a single box serving aggregated to the rack serving disaggregated, same number format and the same engine, is worth about 2.7x. The engine itself, TensorRT-LLM against SGLang on the same rack with the same split, about 3x. Four-bit over eight-bit on the same box, about 1.9x. The ratios do not multiply cleanly, because the curves were run on different dates, and the table says so. Together they are where the 8 to 12x comes from. None of this is hidden. NVIDIA has said in public for two years that the rack, the number format and the serving software are the product, SemiAnalysis’s own methodology says which of them it switched off, and if you turn a lot of that stuff off it doesn’t work out.

Then there is what the benchmark leaves out. It runs one model, Qwen3.5 397B, released in February; on Artificial Analysis’s intelligence ranking it now sits 52nd of 113 with a score of 19, while Qwen3.8 Flash Next, released August 26, scores 40, ranks fifth, and was not run. Neither were DeepSeek V4 Pro, at 1.6 trillion parameters, or Kimi K3, at 2.8 trillion, both of which are already on the dashboard for NVIDIA chips. It runs one workload of three: a long question, 8,000 tokens in, with a short answer, 1,000 out. It runs nothing from AgentX, the multi-step agent benchmark, which SemiAnalysis says comes for TPU “later in the year.” It carries zero accuracy results for any TPU on a page that held 2,746 of them as of September 11, so nobody has checked that the TPU’s answers are as good as the GPU’s on this test. And it carries a caption under the TPU curve: “TPU7x results are an official preview and may change as validation and publication continue.” The GB300 NVL72 curve carries no such label. It is a published result, and every point on it links to the public run that produced it.

The objection here is about the reader, not the authors. Why come out with such a report if things are so skewed, when the average person doesn’t really understand all these things that are missing?

We are calling this the Peeled Apple: an apples-to-apples comparison that only comes out even after one apple has been peeled. SemiAnalysis’s label for rack against TPU is “apples to bananas.” The buyer does not get a bananas column. The buyer pays for tokens at the speed the user wants them, on whatever each vendor ships.

Apples to apples is the whole point of a benchmark, and it cuts the other way from how the article uses it. Strip the rack, the number format, the read-write split and the guessing trick off one side and the comparison is no fairer; the other side just looks like a contender. Fair is each chip at its best, on the same model, at the same speed, on the same accuracy test, and the dashboard can show exactly that. The headline is what you get when you stop one dropdown short of it.

Best Against Best Is 8 to 12x

We ran a commercial lighting company from 2005 to 2019, and every fixture we sold came with an LM-79 report. LM-79 is the lighting industry’s test for a product as sold: the whole fixture, LEDs, driver and optics together, sent to a third-party lab that measures the light coming out in lumens and the power going in at the wall, and reports the number the buyer actually cares about, lumens per watt. Nobody bought on the LED chip’s datasheet. The chip’s number on a test bench at room temperature is not what the building pays for; the fixture’s system wattage is, every hour, for a decade. So buyers compared products on the tested number, because the electricity bill, the utility rebate and the payback were all computed on it. Efficiency matters. These are decisions that buyers are making, and the ROI is determined by the actual performance. Headlines cannot be misleading, because in that business a misleading headline sat on the customer’s power bill for ten years.

The chip-hour is the fixture price. The token is the lumen. Interactivity, tokens per second per user, is the light level the design has to hit, and a serving fleet is tuned to a target speed the way a lighting layout is tuned to a target brightness. So a buyer counts tokens per dollar at the speed the user expects, with each machine running the way its maker ships it, the way we counted lumens per watt on the fixture as tested. Nobody priced a retrofit on the sticker of the fixture, and nobody should price a fleet on the chip-hour. We said the same thing to AMD in July: move the wattage and you move the ratio, so publish the number you can defend.

On SemiAnalysis’s cost model, the TPU chip-hour costs $1.21 against $2.31 for GB300 NVL72, or $1.03 if you are Google and pay the fab yourself. The chip-hour is about half the price. Now count lumens. When each user gets 100 tokens a second, GB300 NVL72 delivers 9.7x the tokens per dollar, 8.2x if you give Google its own lower cost. At 50 tokens a second the gap is about 6x. At 130, the fastest the TPU goes on this chart, it is about 12x. Turn it sideways and the same money buys speed: at the tokens per dollar the TPU delivers at 100 tokens a second, the NVIDIA rack delivers about 430, 4.3x faster. The TPU curve ends at about 130 tokens per second per user. The rack’s runs to about 600.

Two things about the chart’s accounting belong in print. The dashboard charges every chip, the ones reading the question and the ones writing the answer, so the rack’s per-chip number is honest. But it is a rack: the winning NVIDIA points run 17 to 42 chips working together, while every TPU preview point runs on four. A buyer who wants the rack’s economics buys the rack. And the metric counts tokens read and tokens written together, which for a long question and a short answer is nine read for every one written. Count only the written ones and the ratio barely moves: 9.6x at 100 tokens per second per user, 11.9x at 130.

Anyone can rebuild that chart in two minutes. Open the dashboard, pick Qwen3.5 397B, the 8K / 1K scenario, FP4 and FP8, Total Tokens per $1 TCO on the y-axis and Interactivity on the x-axis, then flip the TCO basis between External and Internal and switch on Optimal Only. The table view gives the numbers behind every ratio in this piece, and every point links to the public GitHub run that produced it. Play with it yourself. That is the whole point of a public benchmark, and it is why the dashboard is worth more than the headline.

Tokens per watt is the number the trap was written in, and the preview publishes no power figure for TPU7x. Using SemiAnalysis’s own power assumption for each chip, 2.12 kilowatts for GB300 and 1.207 for TPU7x, the gap is about 10.5x at 100 tokens per second per user and 13x at 130. That is a model, not a measurement.

As we wrote in The Cheap Chip Trap: “The trap is not the chip price. It is the gigawatt that has to feed the chip for the next five years.” That piece was a spec-sheet call built on a Morgan Stanley FLOPs-per-watt chart, and we said so at the time. This is its first mark against a delivered TPU curve, on a public dashboard, on SemiAnalysis’s own cost model.

Jensen said it at Stanford in March 2024, and it read as bravado then: NVIDIA’s total cost of ownership is “so good that even when the competitor’s chips are free, it’s not cheap enough.” Run it on this chart. Google’s own $1.03 an hour only moves the ratio from 9.7x to 8.2x. Halve the TPU’s entire chip-hour, silicon and power and building together, and the gap is still 4x. Quarter it and the gap is still 2x. No chip price on this chart closes it.

The Apples-to-Apples Numbers Moved the Next Day

Even inside SemiAnalysis’s own framing, the dashboard moved the day after the article. The article’s B200 number at the operating point it headlines was 8,903 tokens per second per chip, and $0.222 per million tokens when each user gets 100 tokens a second, with the TPU “approximately 19% lower cost than B200.” The dashboard’s September 8 run of the same B200 setup, on a newer release of the open-source software, shows 10,389 tokens per second per chip and $0.149 per million.

On that run, eight-bit against eight-bit, a single B200 box beats the TPU by 24 to 44% across the speeds a user would notice, 50 to 130 tokens a second. The TPU leads only at the slow end, under about 24 tokens per second per user, which is the point the article headlines, and there its lead is 29% on SemiAnalysis’s cost model and 51% on Google’s own. The B300 curve in the comparison dates from May 21, a four-month-old software build, and matches the article’s number exactly. The authors used the runs they had. One of them moved the next day, the other had not moved since May, and on September 10 the B200’s four-bit curve moved again, which is less a knock on SemiAnalysis than the point of the piece: on NVIDIA hardware the software moves every week, and a benchmark that runs continuously moves with it.

The point where the TPU wins is also the point where the user waits. Time to first token is how long you stare at a blank screen before the answer starts. At the article’s headline operating point the TPU’s median is 3.6 seconds and its average 5.4, which the article reports. B200 is 1.7 and 3.0. B300 is 0.8 and 2.4.

Read the other way, the wait favors the TPU. At its own operating points, 48 to 132 tokens per second per user, the TPU’s first token arrives in about 0.3 seconds. The rack’s cheapest points at the same speeds make the user wait 1 to 3 seconds, because they are serving hundreds to thousands of people at once. A buyer who needs a snappy first token at small scale is reading a different chart.

One thing cuts the other way, and it should be in print. NVIDIA’s rack in eight-bit math without the multi-token trick, a curve dated July 13, trails the TPU at 75 to 130 tokens per second per user, by 10 to 30%. That is the “apples to bananas” comparison SemiAnalysis did make, where it found “about a 30% perf per dollar advantage for GB300 against TPUv7 agg” mid-curve and expects TPU software to close it “within a few months.” On a July build at the wrong number format, the rack is beatable. That is a statement about a July build: the same curve sits below a single B200 box on the September 8 run, which is how stale it is.

The article also asserts that “When serving models using FP4 on NVIDIA GPUs, there is quality loss versus FP8.” That is a testable claim, and SemiAnalysis’s own site tests it. Its accuracy page runs the model on a set of grade-school math problems, GSM8K, the only test listed for this model, and scores the four-bit and eight-bit versions the same: 0.9685 against 0.9700 on B300, 0.9682 against 0.9694 on B200. That test is easy and we would not hang a precision verdict on it. The claim is asserted; the measurement on the same site does not show it.

One more mark against ourselves. In All Roads and Rockets Lead to NVIDIA in August we wrote that “the buyers stopped paying for tokens some time ago, and the two numbers do not rank the same,” meaning a buyer ranks silicon on what a finished task costs, not on tokens. This dashboard has no task number, and for the TPU it has no accuracy number either, which is the other reason the empty evals column matters. Tokens per dollar is where the bill starts. A finished task is where the buyer’s comparison ends.

Where the Data Cuts the Other Way

The first risk is that this is a software snapshot, not a silicon verdict. The TPU curve is a preview: open-source software, one token at a time, reading and writing on the same chips, one run. The two software features that would help it most, the read-write split and the multi-token trick, are not on the public dashboard yet, and SemiAnalysis expects them within months. Our gate: on InferenceX’s Qwen3.5 chart, if Google’s best public TPU7x setup, with both features on, comes within 2x of GB300 NVL72’s best at 100 tokens per second per user by the end of March 2027, the 12x was a software-maturity snapshot and the Peeled Apple was a snapshot too. Anyone with a browser can grade it.

The second is one model, one workload. One model on one long-question, short-answer workload is one point. No long-answer workload, no agent workload, no accuracy submission. A TPU result on agentic workloads could land anywhere, and until it lands the 8 to 12x is a statement about one chart. The second half of the gate: a TPU7x accuracy-eval submission and an AgentX TPU result on the public dashboard. Both, or the preview stays a preview.

The third is the one we have carried since May. The live bear on the Cheap Chip Trap is that models keep learning to do the same work with far fewer tokens, faster than the world adds work. A 12x gap in tokens per dollar matters less in a world that needs a tenth of the tokens. The volume leg of the custom-silicon bear is also untouched. In The Neocloud Hypothesis in March we called that bear “more credible today than it was six months ago,” because Claude and Gemini run mostly on TPUs and Trainium. This chart says nothing about where Anthropic’s tokens run, only what they cost when they run there. The cost leg now has measured data against it. The volume leg does not.

The first two change the slope of the gap. Only the third changes the direction of the call.

So What?

SemiAnalysis’s own text says why a public dashboard matters this year: “Ironwood (TPUv7) is the first generation in which Google is competing for others’ inference workloads with chips that can be purchased outright or rented through its own cloud.” It also dovetails with the Google TPU Blackstone effort to basically sell this to other companies as well, the WSJ-reported joint venture we covered in May with $5 billion of Blackstone equity behind it. The buyer of an externalized TPU is a stranger with a spreadsheet. The stranger can read the dashboard.

That is also the answer to the “Nvidia tax.” VentureBeat’s Sam Witteveen put the take plainly in April: “Google pays fab, packaging and engineering costs on its TPUs. It does not pay that margin.” True, and it shows up on the dashboard as a chip-hour half the price. The margin the buyer pays is per token, and on this chart the tax runs the other way.

There is a second reason the gap on this chart matters, and Jensen gave it yesterday at Goldman’s Communacopia conference. General-purpose compute, he said, “gives you fungibility, durability, versatility, rentability, and very importantly today, because of capital constraints, everybody’s balance sheets, investability. The NVIDIA Compute is finally a computer that could be asset backed. We can use it to secure loans.” The plain version: a GPU can be used by different customers. The tenant who leaves gets replaced by the next one, running a different model, on the same software everyone already uses. A custom ASIC or TPU cannot be ported as easily. Its workloads are tuned to it, its software is young outside its owner’s walls, and if the one customer who runs it walks away there is no line of second customers waiting. That is why the GPU is the easier asset to finance: an asset anyone can use is an asset a lender can take back and re-rent. We wrote in August, off the $500 billion of financing platforms NVIDIA announced with six firms, that compute is the new collateral. This chart is the other half of that story. Financeability rests on the asset being usable by anyone, and usable means the buyer can see what it delivers, on public data, before the money moves. You need the data to show that, so customers can make the right decision. The dashboard is the first place that data exists for a TPU sold to outsiders, and it says what it says.

In April, after Google Cloud Next, we wrote that Camp Two’s efficiency advantage “does not come from having a better accelerator than NVIDIA. It comes from having a better system”. That still reads right, and it cuts both ways: the system on the dashboard is a preview running open-source software. Google’s own claim for its next inference chip, TPU 8i, is about 80% better performance per dollar than Ironwood. Eighty percent does not close 8 to 12x. Google’s published figure for TPU 8i is 10.1 petaflops of four-bit math per chip against Ironwood’s 4.6 at eight-bit, a little over 2x per generation; NVIDIA’s published rack figure for Vera Rubin is about 3.3x GB300 NVL72 on the same four-bit math. The comparison to watch is the next TPU against NVIDIA’s next rack, four-bit against four-bit, on a dashboard neither company runs.

The question we keep landing on is whether “TPUs are just best served horizontally for Google type workloads,” and whether a buyer with a choice would spend the money on NVIDIA GPUs and be better off. The dashboard is the first public place that question has a number.

For NVDA, which we have been publicly long since 2016, the Nvidia-tax leg of the bear just lost its data, and the share-loss story needs a new number. It is one data point for the sixth gate in The Only Company That Can Afford Free, open inference staying on NVIDIA iron; that gate grades on serving mix over two quarters, not on one chart.

For Google, no position, the dual-fleet path from Cloud Next, Vera Rubin A5X next to TPU, looks like the right hedge, because the TPU pitch to outsiders now lands on a public chart. For Broadcom and Marvell, no position, the custom-chip programs they are paid to build through 2027 and 2028 are not the question; how much of the accelerator market those chips end up with is what a measured 8 to 12x squeezes.

We wrote this down before pulling a single number, and it still holds. Either way SemiAnalysis stays the reference. One shop runs these numbers, so its assumptions become the market’s assumptions, holes and all. The fix is not a better rebuttal. It is a second, independent set of numbers.

Coming Up

• We re-pull the InferenceX dashboard the morning this publishes and any morning a TPU configuration changes. If Google’s disaggregated or MTP curves land, the update runs in the subscriber chat the same day with the gate regraded.

• The next MLPerf inference round is due; we will grade it against this chart when the results are public.

• We Will be at The All In Conference this week in Los Angeles

Related BEP Research

• The Cheap Chip Trap Jensen Just Confirmed At Dell World (May 18): the spec-sheet call this chart marks

• The Third Announcement Most People Missed at Google Cloud Next (April 25): two camps, one dual-fleet cloud

• The NeoCloud Hypothesis (March 11): where the custom-silicon bear was named

• The Only Company That Can Afford Free (September 3): the sixth gate this chart feeds

• The Fourth Piece Ships (March 17): the first InferenceX number we printed, off a vendor slide

• AMD Showed a 15% Lead. Its Own Slide Showed Why It May Not Last. (July 27): reading a benchmark against its footnotes

Disclosure: BEP Research’s principal is long NVDA, LITE, CRDO, TSEM, ALAB, WOLF, NOW, SMCI, BE, NBIS, and ORCL (2027 LEAPS), and has been publicly long NVDA since 2016. This is investment research, not investment advice. Do your own work.

Free Models Have a Compute Bill

NVIDIA is reportedly in talks to buy Hugging Face for around $13 billion. The number is a scoop, not a filing. Insider had talks above $13 billion that can still die. The Information had $12.9 billion agreed. Bloomberg had about $14 billion with a $1 billion retention package, then on Tuesday evening moved it to $12.9 billion plus the retention pool, an agreement possible as soon as this week, no final terms, both companies declining to comment. It is September 3. Nobody has filed.

Days earlier came Poolside: a $6 billion non-exclusive license to its Model Factory, $1 billion of equity, and jobs for more than 100 of its engineers. The two moves are one strategy.

The context for both is a chart. NVIDIA is already the biggest corporate publisher on Hugging Face, and it pulled away from the field this year: by AI World’s count of repositories with real traction, more than 500 added since September 2025, ahead of Alibaba Cloud, Hugging Face itself, and Tencent.

We were on ETN the morning after the print, and put the problem this way, on tape:

“Poolside was making a really good open source coding model, but the challenges of open source is like if you’re spending all this money on making this model, but you’re kind of giving it out for free, how do you pay for all this compute?”

That question is the whole deal, more than the headline number. An open model is a product with billions of dollars of cost of goods and a price of zero. For three years the answer to who pays was venture capital, and venture capital wants the loop to close eventually.

It closes a different way. As we said on the same tape: for a lot of the open source guys the business model didn’t work, and you needed someone like an NVIDIA to come in, fund it with the compute, and take care of the talent.

An open model cannot charge more for its tokens than the cheapest stranger hosting its weights, so the durable owner of open source is whoever gets paid when the tokens run.

That is the Compute Patron. To make all these models you need compute, and compute is expensive. How do you expect to have something open source and not be able to pay for it? What is your end game? You can’t be a charity. The patron’s answer: if you create more demand for compute through open source, you still need to sell more compute, which is NVIDIA compute, and a lot of these models will be optimized for NVIDIA general-purpose compute and CUDA.

The Loop Breaks at Step Four

The business model of open source did not fail on quality. It failed on a missing step. In The Token Dollar we mapped AI compute finance as a loop: borrow dollars, buy GPUs, mint tokens, sell tokens for dollars, service the carry, roll the principal. An open-weights release gives step four away. The tokens still get minted, on the same GPUs, and anyone who downloads the file can sell them. DeepSeek and Mistral run price sheets on their own weights; Together and Fireworks run them on everyone’s. The publisher keeps whatever margin the cheapest host leaves, which rounds to nothing. That is why an open lab’s runway is somebody else’s balance sheet, always.

Jensen said the cost side plainly in July: “Frankly, I think closed models are cheaper.” His list of what open costs the builder ran through fine-tuning, guardrails, evaluation, and building the computers to host it, and he landed on “there’s nothing cheap about doing that.” If open models are not cheap, someone has to pay for them, on purpose, indefinitely.

Run the arithmetic on one lab. Poolside’s training capacity was a 40,000-plus GPU commitment at CoreWeave, estimated at $5 to $6 billion over five years, with no token margin on the other side. They then missed a six-week window to raise about $2 billion for the GB300 cluster due in January, and sold the factory to the company that sells the compute. The letter to investors is the instrument: $6 billion non-exclusive license, $1 billion of equity at $12 billion pre-money, founders stay. Not an acquisition. The structure is the tell.

We tried to get in contact with Eiso before and after the talks and never got through. Laguna was a promising frontier coding model and the team behind it was executing well. Maybe the investors didn’t see a path to revenue or profitability versus the competitors that were out there, and this is the best case for them: the founders stayed, and the employees left. Licensing is a faster path to getting people into the office to work on the product, just like Groq, announced on Christmas Eve with Jonathan Ross and his team going to NVIDIA. We don’t think it could possibly be a rescue, and it also benefits the open source ecosystem. A missed raise and a rescue are different things: the equity came in at $12 billion pre-money, which is not a distress price, and a non-exclusive license leaves Poolside free to sell the same factory again.

Now look at who carries that kind of bill today, model by model. Meta pays for Llama out of the ad business. Alibaba pays for Qwen out of the cloud business. Zhipu and Moonshot pay for GLM and Kimi out of funding rounds and, increasingly, out of Beijing’s interest in them existing. Our theory, and it is only a theory, is that somehow they are being subsidized by the Chinese government; there is no way they can do all this at the cost they are saying. Every open model of consequence is subsidized by a business that is not tokens.

The Patron Gets Paid in CUDA

NVIDIA is the one patron whose subsidy is structural, because the free model is a demand generator for the thing NVIDIA actually sells. If you buy the box, you cannot put frontier OpenAI or Anthropic weights on it. Those stay in their clouds. What you can run on your own iron is Qwen, GLM, Kimi, Llama, and Nemotron. Most of the open-model inference in the world, on-prem, at a neocloud, on a DGX Spark under a desk, clears through NVIDIA silicon today, and Nemotron is the executor tuned to that silicon on purpose. Most, not all, and we have already published the other side. On August 4, in All Roads and Rockets Lead to NVIDIA, we wrote that “inference also sells behind an API, so a provider can swap hardware under an OpenAI-compatible endpoint without telling anyone,” and open weights make the swap easier, because the same file runs on TPU, Trainium, and MI. What the patron structure changes is which iron the free model is tuned for first, and NVIDIA is hoping that more compute in every form factor it sells, Jetson Thor, the DGX Spark, the datacenter GPUs, captures one of these open source models. Whether that holds the volume is the sixth gate at the end of this piece. We would probably see the drift in the neoclouds first, or on TPUs. Up until now it hasn’t happened, and on our read NVIDIA is increasing its share of inference.

The demand side showed up too, and it is personal for a lot of buyers now, including us. We put a DGX Spark under a desk this summer, the machine we called The Edge Token Factory in June, and we had a hard time setting it up. Connect to Wi-Fi, then command prompt, then a bunch of updates, and we had to use ChatGPT to ping it and find it on the network. When you can’t find it on your network and it’s hard to connect, it gets a little scary, and then you’re like, okay, how do I get open-source models? Where do I get them from? Everyone was pointing towards Hugging Face as the place to be. We had to create a Hugging Face account and get a token and do all that stuff, and those are barriers to the process to get open source models. The initial setup took an hour and a half. In the end we had it all connected, with Claude Code basically running the Spark for us on the Qwen models, NVIDIA’s own build of them, and we were able to do some image generation. With NVIDIA owning the shelf, it gets more seamless and easy to set up and less guessing, because it is all coming from one trusted provider. NVIDIA shows up on the screen now instead. On the Blockspace tape we put it the way an owner would: run a good enough model at home and your landlord isn’t Dario or Sam. You own your own tokens.

Open-source share of enterprise LLM spend fell from 19 percent to 11 percent in a year while Decagon still runs about 90 percent of its workloads on open models, and open-weight models went from a third of the tokens routed through OpenRouter in late 2025 to a majority by mid-2026, per the July State of Open Source AI report. The work stayed. The invoice left.

The CUDA Playbook, One Layer Up

Poolside and Hugging Face are the first rows of NVIDIA’s deal ledger that touch the model and its distribution. In January, in NVIDIA’s Inference Stack Depth Strategy, we called the pattern the CUDA playbook replicated at the inference layer. That ledger ran past $30 billion in announced deals, licenses, and reported talks, every row below the model. Add $7 billion for Poolside and a reported $13 billion for Hugging Face and it is past $50 billion, still mixing checks and term sheets.

Hugging Face is the distribution row, and distribution is where this stops being an M&A story. Every executor in the enterprise agent stack lives there, and so does NVIDIA, more than any other company, as the chart at the top shows: Nemotron, Cosmos, GR00T, and optimized versions of other labs’ models. The biggest corporate publisher on the shelf is the one reportedly buying it, and a good share of what it publishes is other people’s weights tuned to its iron. America needs an open source leader, and no one else is really stepping up to the plate. It is funny which company it is: people think, why would you want to give away something for free? This is NVIDIA. Trusted distribution matters, and putting all the pieces together, the full stack story we have written about all year, this is the next full stack solution.

As we wrote in The Open Source War Is Over: “Nemotron was built to be the layer under the frontier, not to beat it, and to make sure every token in the system, orchestrator and executor alike, gets minted on NVIDIA hardware.” If the report is right, the company that ships the American executor would also own the shelf that every executor, American and Chinese, ships from. We said as much on ETN the morning after the print, before any of the numbers firmed: NVIDIA is “building a full stack solution for open source all the way from Hugging Face to the Nemotron models that they’re creating.”

The Bear Case: A Patron Is Not a Neutral

Four ways this read gets hurt. The first three are in rising order of damage. The fourth is the one we lived.

1. The deal is a report. Talks die and sizes shrink, and a $13 billion headline can end as a commercial partnership with a press release. If it dies, the patron economics still exist; Nemotron and Poolside already prove them. What dies with it is the depot, and the depot is what the $13 billion was for.

2. Neutrality was the asset. Hugging Face works because it belongs to no lab and no silicon vendor, and ownership by NVIDIA tests that instantly. Our own answer as an uploader: we’re okay with it, at least it is a trusted source, and they have other models hosted there as well. People might pull weights, but others might keep it; it’s not for everyone. The observable: where new model uploads and download share go in the two quarters after a close. Migration is the miss.

3. The frontier closes the seam. Carried forward unchanged from July: the moment Anthropic or OpenAI sells a first-party small model, fine-tunable on your task, served in your perimeter, the two-tier stack becomes a one-vendor product and the executor tier loses its reason to exist. This is the one that changes the conclusion rather than the slope, and it was the one that changed the conclusion in July too.

4. Open source stays a hobby. What if open source is just like a hobby and doesn’t really take off, and everyone just wants closed source? Based on the experience we had, it wasn’t really easy to set up. For open source to be a real winner it needs an iPhone or Mac type of experience, where you get up and running out of the box simply and don’t have to be a developer or data scientist to put it together. It is the bear the reported deal is built to fix.

We don’t think you can look at revenue numbers on this one. It is all about scarcity, moat, and user value, and the talent behind the team, and possibly other bidders. Stripe buying OpenRouter for more than $7 billion last month is the same type of concept. The one public anchor is the $4.5 billion valuation from its 2023 round; the reported price is roughly three times that, on revenue the company does not disclose. What we can underwrite is the stack, and the stack does not need Hugging Face to earn its keep. It needs open models to keep improving and keep running on NVIDIA.

So What?

Ask who pays for the next open model release, because the answer determines whether it is a durable layer of the stack or a runway with weights attached. Venture patience and Beijing’s interest both run out. A patron who books the subsidy as ecosystem spend, against the highest-margin hardware business on earth, does not run out. People didn’t believe in CUDA, and NVIDIA was just spending, hoping developers would come and use it. Eventually it found its use case. When it’s early, people never see it. We honestly didn’t see it then either. We bought the stock in 2016 because NVIDIA was using its GPUs to solve hard math problems for crypto mining, and our thought was that if you can solve hard math problems you can probably solve other problems, and these systems have a value for creating some sort of intelligence. We see it now.

The instinct runs all the way up the house. On September 2, Open Athena, a nonprofit, announced a 535-billion-parameter Marin training run paid for with what it called “a generous gift of compute from The Jen-Hsun and Lori Huang Foundation.” Jensen’s line in the announcement: researchers “need access to frontier-scale models they can build, study, and understand.” A nonprofit cannot sell tokens either. It is just interesting to see Jensen and his family spending money on a nonprofit to make other open-source models. That is something he believes in. He doesn’t have to do this. He could have taken this money and built a museum, or donated to some university campus to be named after him. The patron paid anyway.

It should deflate all the token prices. Open source tokens should be much cheaper than frontier, but then you are paying upfront hardware and maintenance costs to host them yourself, and that invoice goes to the patron. For the closed labs, the floor just got harder: Anthropic and OpenAI now price against an executor tier maintained by a competitor that never needs token revenue.

For NVDA itself, we read the reported deal as CUDA spend, and we are not resizing a position we have held publicly since 2016 on a press report. Still holding core strong. It deepens the moat, and these are wise investments that pay off in the long run. It is hard to look at a $5 trillion company and think there is still room to grow, but the market isn’t fully penetrated; the TAM is so big that there are still opportunities.

Six gates, all observable from outside, in the table below. The Hugging Face talks confirm near the reported size, which grades the depot. Nemotron stays free with no paid token API, which is the patron gate. The depot stays neutral and its access never gets tied to NVIDIA hardware, both live only after a close. The frontier does not close the seam with a first-party small model in the customer perimeter, which changes the conclusion as it did in July. And open-model inference keeps clearing mostly on NVIDIA silicon, where two straight quarters of non-NVIDIA share rising grades the ticker, not the patron. It would break the call for us, probably in a few years, but not now. Miss the first gate and NVIDIA is still the patron. Crack the second and the patron was just a shopper.

The model is free; the compute never was.

Underwrite the patron, not the model.

Coming Up

• The paid follow-up: what Hugging Face is actually worth as a business, and what a closed deal does to our coverage universe. Ships when the deal status hardens.

• GTC Berlin is October 20, and we will be there. If you are going, reply to this email; last GTC produced two on-record interviews that became pieces.

• One more piece of housekeeping. BEP's principal runs occasional SPVs in private companies across this stack, offered to verified accredited investors under Rule 506(c). If you want to see those deals when they open, join the list at tools.bepresearch.com/syndicate.

Related BEP Research

• The Open Source War Is Over (X, July 7): frontier orchestrates, open executes, and NVIDIA armed both sides

• The Token Dollar (April 4): the loop this piece breaks at step four

• The Slide in NVIDIA’s RTX Spark Keynote (June 1): The Edge Token Factory

• All Roads and Rockets Lead to NVIDIA (August 4): the Palantir deployment that closed the July gate

Disclosure: BEP Research’s principal is long NVDA, LITE, CRDO, TSEM, ALAB, WOLF, NOW, SMCI, BE, NBIS, and ORCL (2027 LEAPS), and has been publicly long NVDA since 2016. This is investment research, not investment advice. Do your own work.

Good morning. It’s been a busy week of Hot Chips and NVIDIA earnings and all the action. Pour yourself a coffee and dig in, this is a long one!

This is a free report from Ben and Rick.

Co-authored with Rui “Rick” Xie (rickxie.cn), a memory-systems researcher with a PhD from Rensselaer Polytechnic Institute whose work runs from DRAM to SSDs. This is the third installment in our memory series. Rick leads the engineering and standards analysis; Ben Pouladian leads the investment framing.

The first two installments lived on the DRAM side of Rick’s research. The Bandwidth Tax argued that the real cost of HBM is levied on the system, not the stack. The Tax Dodge tested the first commercial architecture built to owe less of that bill, and it closed with a promise: the next installment crosses to flash, and the honest question there is not whether NAND can deliver more bandwidth. It is which AI bytes actually belong on NAND. This is that piece.

At its August investor day, SanDisk put a number on a slide that deserved more attention than the guidance did. Of the 1.2 zettabytes of data-center NAND demand the company forecasts for 2030, 420 exabytes is KV cache. I was not on the call. I read the deck, and my first reaction was the useful one: that was a big number. That’s crazy how big that is.

It’s always a problem, KV cache, because you just never have enough of it, and it slows things down. It is the live working memory of inference, the scratchpad a model writes to with every token it generates, and constantly writing it and deleting it is exactly what reduces the endurance of a flash drive. So a third of the demand forecast for the storage tier is the one workload flash is worst at serving. The deck prices those bytes. It does not price the writes.

NAND is priced by the byte and worn out by the write, and the biggest new line on SanDisk’s demand slide is a write stream.

Call it the write problem. It decides which part of the zettabyte is real.

What SanDisk Actually Told Investors

SanDisk’s investor-day slide is a forecast of how much paper the AI factory will buy. Three piles by 2030, all inside the data center, adding to 1.2 zettabytes.

480 exabytes is staging. The unread homework piled up for the next training run. 300 exabytes is fast data lakes. The filing cabinet of work already finished. 420 exabytes is KV cache. The notebooks. That last pile is a third of the forecast, and it is the pile flash is worst at holding, because notebooks get written the whole time the session runs.

The rest of this section is their supply story. Same factory, more pages per wafer, they still choose to ship fewer pages than they could. They have already pre-sold about half of next year’s paper and two-thirds of the year after to eight data-center customers. They want to sell it like a scarce good, not like a commodity. Then, near the end of the roadmap, they propose stacking that paper next to the GPU. That product is HBF. This piece is not about whether the slide is big. It is about which pile on the slide is real.

The demand claims first, and they are large. SanDisk expects industry bit shipments to reach 1.2 zettabytes in 2026, up from 960 exabytes in 2025, with data centers now roughly half of bit demand. By 2030 the company forecasts 1.2 zettabytes of data-center demand alone, split three ways: 480 exabytes of staging for training data, 420 exabytes of KV cache, and 300 exabytes of fast data lakes. The mix maps onto cell types: TLC at roughly two-thirds of bits, serving staging and hot KV cache, with QLC at a third, serving fast data lakes and cold KV cache. Zephyr, whose recap on the Citrini Research chat is the best public read of the day, called the forecast conservative, and separately estimated that persistent agent memory could add another 500 to 600 exabytes per year on top of it. That framing is theirs; the deck numbers are SanDisk’s own. My read on the buckets: obviously it’s data centers, but most personal compute devices will probably need bigger amounts of data to run onboard compute and these bigger models, and then you have this whole token explosion of robotics and edge AI cases. People move a lot of their AI close, on premise. You just keep on training bigger and bigger models. Maybe you don’t need that much for fast data lakes. I don’t know about that.

The supply strategy is the more disciplined half of the deck. SanDisk plans to grow bits through wafer productivity rather than new fabs: each BiCS generation adds about 54 percent more bits per wafer, roughly 27 percent a year, and the company targets only mid-to-high-teens bit growth. A vendor that can grow bits 27 percent a year through layer count alone, and chooses to ship mid-to-high teens, is telling you it intends to sell scarcity, not volume.

The commercial structure says the same thing louder. SanDisk has signed NAND bit agreements, long-term contracts in all but name, with eight data-center customers, covering about 50 percent of FY2027 bits and about two-thirds of FY2028 bits. On top sits a FY28 through FY30 target of mid-to-high-teens revenue growth and a non-GAAP gross margin of roughly 80 percent, which is less a stretch than a hold: SanDisk printed 84.6 percent non-GAAP in its fiscal fourth quarter. A commodity flash vendor is now writing multi-year contracts that price forward a constraint, and guiding to keep the margin the constraint produces, which is the same behavior we documented in the HBM oligopoly one tier up.

Then, near the end of the roadmap, High Bandwidth Flash: NAND stacked and interfaced like HBM, the product this series flagged in July.

Weights Read. KV Cache Writes.

When you open Claude, ChatGPT, or Grok and type a question, two different piles of bytes go to work.

The first pile is the model itself. Fable 5, GPT-5.6 Sol, Grok 4.6. That brain got written to storage when the lab shipped it. Then it gets read, over and over, by every session on earth until the next version drops. Think of a textbook. They print a new edition a few times a year. Every student reads the same copy. You do not reprint the book for each class. That is weights. Read mostly. Flash, the cheap chips inside a data-center SSD, is good at this job. It hates being rewritten. It loves being read.

The second pile is this conversation. Every sentence you send, and every sentence the model sends back, has to be remembered so the next answer still knows what you asked. Paste a 10-K and ask fifteen follow-ups. Let Cursor chew a repo for an afternoon. Run a research agent for an hour. That memory grows for as long as the session runs. It is a notebook you keep adding lines to, not a textbook on a shelf. That is KV cache. It is a write stream. And a write stream is exactly what wears flash out.

A one-line haiku barely fills a page. A hundred-thousand-token agent session fills a notebook. SanDisk’s demand slide adds the textbooks and the notebooks together and calls it one market.

Weights change on deployment cadence, not nanosecond cadence, so paying HBM prices to store them is architecturally wasteful; this newsletter has been making that case since April. A read-mostly workload is the best case NAND ever gets. Endurance barely enters the equation, access patterns are large and predictable, and the only hard questions are bandwidth, latency, and heat.

KV cache is the opposite animal. Putting that write stream on NAND means paying costs that never appear on a demand slide, and this is where Rick’s triage begins. What follows is his arithmetic, worked on public sources.

Endurance. Think of endurance as how many times you can fill and toss the notebook before the paper wears through. DWPD, drive writes per day, is that number for a flash drive. The arithmetic turns the write problem from a metaphor into an endurance budget. Take a representative conventional attention configuration with 80 layers, 8 KV heads, a head dimension of 128, and BF16 storage. The KV state created by each token is:

KV bytes per token = 2 × 80 × 8 × 128 × 2 bytes = 320 KiB

A 100,000 token session therefore creates approximately 32.8 GB of logical KV state. For latent attention, quantized KV, or compressed representations, the serialized bytes per token replace this value while the rest of the calculation remains unchanged. The physical write rate is:

Physical DWPD = Tokens per day × KV bytes per token × Persisted fraction × WAF / Provisioned capacity

On a 1 PB flash tier, persisting 10,000 sessions of this size per day produces 328 TB of logical writes, equal to 0.328 logical DWPD. A WAF of 1.5 raises the physical rate to 0.492 DWPD. A WAF of 2.0 raises it to 0.656 DWPD. A WAF of 3.0 raises it to 0.984 DWPD.

The result exposes the variables that control the endurance budget. Session length sets the KV bytes created by each request. Request volume sets the daily write stream. Admission policy determines how much of that stream reaches flash. Placement determines how many physical writes the media ultimately absorbs. SanDisk’s TLC and QLC mapping identifies the intended capacity tiers. The operating economics depend on how quickly each tier consumes its write budget. At the scale of the 420 EB forecast, every 0.1 physical DWPD corresponds to 42 EB of media writes per day. The value of the forecast therefore depends on the lifetime of the bytes as much as their total capacity.

Write amplification. Flash never writes just the notebook you asked for. To add a line, it often has to copy a whole page. That extra copying is write amplification, WAF. Write amplification connects logical KV demand to physical NAND wear: physical writes equal logical writes times WAF. In the 1 PB example, the logical workload writes 328 TB per day, equal to 0.328 logical DWPD. Using WAF values of 1.5, 2.0, and 3.0 as illustrative sensitivity cases, the physical write rate becomes 492 TB, 656 TB, and 984 TB per day, respectively. Each additional 0.5 of WAF adds 164 TB of media writes per day, equal to 0.164 physical DWPD on a 1 PB tier. Moving from WAF 1.5 to WAF 3.0 doubles the daily physical write rate and consumes the same endurance budget twice as quickly.

The investment implication is direct. A capacity forecast becomes an endurance forecast only after logical write intensity and WAF are included. For the same installed capacity, physical DWPD determines replacement cadence, sustained media traffic, and the useful lifetime of the tier. The deck measures how many KV bytes the market may store. WAF determines how many NAND writes those bytes generate.

Reload cost. Reuse is the only reason to keep the notebook. If thirty employees ask “what’s our refund policy” every morning, you might keep yesterday’s notes and reread them instead of re-reading the whole manual. That reread is a reload. Recomputing from scratch is prefill. A one-off haiku never pays. A company bot answering the same prompt ten thousand times a day might. Reuse determines whether persisted KV creates economic value. Live decode KV is appended and reread on the token path, so active state remains in HBM or another upper memory tier. Persisted KV enters flash when its expected reuse value exceeds the cost of the admission write and subsequent reloads. Using time as a first-order resource proxy, persistence is favorable when:

T_write + E[N_hits] × T_reload < E[N_hits] × T_prefill

Here T_prefill is the prefill time avoided by each successful reuse. When T_reload is less than T_prefill, the break-even reuse count is:

E[N_hits] > T_write / (T_prefill − T_reload)

The 100,000 token example contains approximately 32.8 GB of KV state. Its idealized transfer time is 1.31 seconds at 25 GB/s, 0.33 seconds at 100 GB/s, and 0.082 seconds at 400 GB/s. Consider an illustrative case in which the initial write and each reload take 0.33 seconds. If one reuse avoids 5 seconds of prefill, the first hit creates a net time benefit: 0.66 seconds to write and reload against 5.0 seconds of prefill avoided. If one reuse avoids 0.5 seconds of prefill, the break-even reuse count becomes 0.33 divided by 0.17, or 1.94, and the second hit crosses the threshold.

Context length determines how much prefill computation can be avoided. Batch size determines the economic value of the accelerator time recovered. The storage and interconnect path determine the reload cost. The reuse distribution determines how many times that value can be collected before the object expires. Direct GPU I/O reduces CPU involvement, memory copies, and software overhead within T_reload. Effective path bandwidth acts on a separate component of the same term. For a 32.8 GB object, increasing effective bandwidth from 25 GB/s to 100 GB/s reduces idealized transfer time from 1.31 seconds to 0.33 seconds. Lower transfer and software costs move more prefixes across the reuse threshold. Discounted context caching is most valuable for frequently reused prefixes, where avoided prefill accumulates across repeated hits. Its economic value depends on identifying those prefixes before committing the admission write.

The demand slide counts KV bytes. The investable opportunity is the subset that earns its place on flash.

The one-line version of Rick’s boundary, from his framing of this piece, is the line to keep: HBF looks strongest for read-heavy weights, while KV cache may drive much larger NAND demand but remains constrained by writes, endurance, placement, and reload cost. The decisive question is no longer whether NAND can deliver more bandwidth. It is whether the tier actually improves inference once placement, reuse, writes, thermal limits, and the remaining HBM tier are included.

HBF Is the Test Case, Not the Story

High Bandwidth Flash is the product name for a simple idea: take those cheap flash chips and stack them next to the GPU the way HBM, the expensive fast memory, is stacked today. A bigger shelf on the same desk, not a warehouse across town. SanDisk wants that shelf to hold the textbooks, and maybe some of the notebooks. The rest of this section is whether the notebooks survive on paper that wears out when you write.

High Bandwidth Flash has moved from a concept into a standards and sampling story, and the record on that trajectory is partly ours to correct. In January, The Hierarchy Rewrites mapped HBF as the first of Patterson’s four research directions, quoting his paper’s claim of “10× memory capacity with HBM-like bandwidth,” and said the SanDisk and SK Hynix standardization effort would decide whether it became a real product or vaporware. In July, the Bandwidth Tax migration map gave HBF a row on the strength of that standardization and carried the then-current timeline: samples this half, first inference devices in 2027.

That timeline did not hold, and the update goes on the record here: on SanDisk’s own investor-day roadmap, accelerators with HBF now arrive late 2028 or early 2029 at the earliest. We printed 2027 in July. The company’s own deck has since moved the date by 18 months to two years. Adjust the row accordingly. I can’t get the timing right on everything. I was early. I’m more of the direction of the thesis, and the direction held.

The scope narrowed along with the schedule, and two independent reads converged on the same boundary. Zephyr’s read of the investor day is that HBF is a weights tier, because KV-cache write traffic would consume NAND endurance, and that real deployments will be hybrid, HBM for the write-hot state and HBF underneath it. Rick’s read of the public FMS and OCP materials lands in the same place with the skepticism intact: NAND endurance is not solved, HBF may work well for read-mostly data such as model weights, and KV cache is much less clear because the actual write stream depends on placement, eviction, reuse, and write amplification. Zephyr also flagged that SanDisk’s deck misstates the HBM specs it compares against (their X thread), which is worth keeping in frame when reading any HBF-versus-HBM chart from the company.

The spec sheet will throw numbers at you. Stacks 8 or 16 dies high. Up to 512 GB on the shelf. Bandwidth grades that sound like HBM. SanDisk’s 1.6 TB/s is one product target, not a promise the notebooks will keep up. Version 0.7 is a draft interface, not a qualified product. Peak speed on the pin is the highway speed limit. Useful speed is what you get in traffic, with stops, heat, and paper that wears. Read-mostly textbooks keep the useful speed high. Tiny writes, erasing old notes, and shuffling pages are what kill it for KV. The formula a few lines down is just that sentence in symbols.

HBF mechanics. The public OCP v0.7.0 high-level base-die specification describes UCIe-based xPU connectivity, 8-high and 16-high stacks, capacity up to 512 GB, and bandwidth grades of approximately 0.384, 1.536, and 3.072 TB/s. SanDisk’s 1.6 TB/s number is therefore a first-generation product target near one public grade, not the only bandwidth defined by the architecture. Version 0.7.0 is a pre-1.0 high-level interface specification, not a qualified product specification. It establishes an interface and system contract, but it does not establish sustained application throughput or workload qualification.

At system level, the useful quantity is service rate under the actual access mix, not peak pin bandwidth. A compact model is B_useful = B_peak × η_protocol × η_scheduling × η_media, subject to tail latency, power, and thermal constraints. Read-mostly weights are favorable because static placement and parallel reads can keep η_media high. Small writes, invalidation, and reclamation couple KV throughput to wear and latency variance, so the same peak interface number does not imply the same usable bandwidth for the two byte classes.

The qualification questions are therefore workload specific. For weights, test sustained random-read service rate, p99 latency, thermals, retention, read disturb management, and error recovery within the package power envelope. For persisted KV, add program rate, reclamation behavior, endurance, and recovery from bursty admission. Public evidence does not yet identify one universal failure mode or publish enough data to close those gates.

So the test case is narrower than the pitch, and that is the honest shape of it. A weights tier with HBM-class packaging and a 2029 accelerator date is a real product thesis, but it is not 420 exabytes of KV cache. The story that carries the zettabyte is the broader one this piece is actually about: an industry trying to move the working set of inference down the hierarchy, one byte class at a time, and discovering which classes survive the move.

The Room Attacked the Write First

Hot Chips is the chip conference. Sunday’s memory tutorial had a live Slack channel for attendees. They did not let the bandwidth pitch land. They asked the notebook question first.

If you take notes on cheap paper, does the paper wear out?

Flash can be rewritten far fewer times than the fast memory sitting on the GPU. One person in the room said a million times fewer. That is not a lab spec. It is an engineer smelling the problem. Another assumed 100,000 rewrite cycles, the kind of number cheap flash is often rated for, and said a real agent session would burn the drive. His fix: only put the textbooks on it. Never the notebooks.

Someone asked for a tiny expensive notepad on the same chip, so you do not scribble on the cheap paper every second. That is a coalescing cache.

The Oxmiq speaker said new notes per batch are small and can be bundled. He did not give a wear number. Energy, he said, is not in the spec. Samsung and Micron answered a different question, heat on the expensive stack. Nobody answered whether flash survives the notebook.

The rest of this section is that Slack, almost as it happened.

Oxmiq Labs presented “HBF in AI Compute” just after noon. One attendee asked how HBF can serve KV cache if write endurance is more than 1E6 lower than HBM, and whether the simulations assumed a read-only case. That 1E6 figure is an attendee characterization: HBM DRAM has no analogous program-and-erase cycle rating, so it does not establish a like-for-like endurance ratio. A researcher at UC Merced ran the same attack with an SLC assumption of 100K program/erase cycles and said KV writes would burn the flash. The 100K P/E figure is his illustrative SLC assumption, with no corresponding SanDisk product rating or OCP requirement, and P/E count alone cannot establish KV-tier lifetime without admitted writes, physical write amplification, provisioned capacity, temperature, and retention. His read: stripe it for read-only data. Another asked for an SRAM or DRAM coalescing cache on the HBF base die. Others asked whether read and write bandwidth even match, and for pJ/B including the controller.

Anurag Agrawal of Oxmiq answered later Sunday afternoon. On decode writes: “batch size number of new entries. Its little and can be aggregated.” No P/E number. Aggregation can reduce transaction overhead and some physical amplification, while the logical KV write volume remains. On energy: “Its not in the spec.” The simulations, as he described them in the channel, assume the same pJ/bit as HBM, and that equal pJ/bit assumption is a simulation input, not a measurement of system energy including the controller and media path. On a coalescing cache: “I am not sure how it will evolve.”

The speaker who did answer memory questions was Samsung’s Sangwook Han, and he was answering a different problem. He said it would be difficult to have more than about 2K IOs on the interposer. On zHBM, the C-die stack sits on the xPU with no B-die, so HBM PHY and D2D are not necessary; only vertical TSVs exist, and power is “extremely low.” Heat is the catch: “DRAM dies will sit directly on the xPU.” On aHBM: not HBM PIM. Processing elements live in the B-die, not the C-die, and in one worked example, moving some attention units cut D2D bandwidth to “about half,” an example rather than a spec. That is still the HBM wall one tier up, not NAND writes. He did not answer HBF endurance. Micron’s Raghu Sreeramaneni called stacked DRAM important and blocked by thermals, packaging, and architecture. He did not answer HBF either.

The channel is not a 420-exabyte forecast. It is a room that already knows the KV write stream is the part of the slide that has to survive contact with NAND.

The Same Overflow, Contested From Two Sides

The notebooks have to go somewhere when they no longer fit in your hand. NVIDIA already built the overflow system. Dynamo is the software that decides which notes stay on the desk, which go to a nearby shelf, which go down the hall, which get thrown out. BlueField is the clerk who moves them. That stack is already running in production clouds.

SanDisk’s 420-exabyte slide is the same claim, restated by a paper company. NVIDIA also has a huge pile of cheaper memory at the rack, LPDDR, 384 terabytes in a full Vera setup. We asked if that pile becomes addressable notebook space. They would not say. If the overflow lands on that rack memory, SanDisk’s KV forecast deflates. If it lands on flash, the 420 exabytes gets real, and the wear problem becomes the industry’s problem. Both routes go around expensive HBM. They pay different tolls. The write problem is the toll when the route is flash.

The biggest platform in the industry has already built the plumbing for the KV rows. NVIDIA’s Dynamo serving stack manages a four-tier KV hierarchy, with rack-scale storage on BlueField hardware purpose-built for cache overflow. When the market sold memory stocks on a compression paper in March, our response piece read that software roadmap as the tell that the industry is planning for “the future involves more cached state, not less.” The cached-state forecast SanDisk just published is that same claim, restated in exabytes by a flash vendor.

That deferral is on the record in the Vera piece, where we put the question to NVIDIA directly. SanDisk’s investor day is, in effect, a bid for the same bytes one tier further down.

You route around a wall and you pay for the route. The wall here is HBM capacity cost per byte of context; the routes are LPDDR at the rack, NAND in the storage tier, and eventually HBF under the HBM; and the toll changes form at each stop, from interposer and packaging at the top to endurance and placement at the bottom. The write problem is what the tax looks like when it reaches flash.

Two Routes Into the Same Bytes

Days before SanDisk’s investor day, the buyer side published its own version of the slide. At FMS in Santa Clara, NVIDIA argued in a company blog post that storage has to become an active component of the compute pipeline, with microsecond-class access, because the memory-versus-storage tradeoff “now plays out in microseconds.” The post names a product tier for exactly the bytes on SanDisk’s slide: CMX Context Memory Storage, NVIDIA’s storage tier for agentic AI inference context.

The plumbing NVIDIA named is how the library works. cuFile lets the GPU fetch the notebook itself instead of asking a CPU librarian to walk it over. SCADA is the rules for those trips. BlueField-4 STX is the clerk NVIDIA says is faster than a generic x86 clerk at packing and locking the files. Those speedup numbers are NVIDIA’s own, untested by us. Storage-Next is the vendor list trying to fill that library, more than 40 names. The investable split is still the same: library down the hall versus shelf on the desk.

Around it sits the plumbing to make a drive behave like a slow, enormous memory: the cuFile APIs open-sourced so that “GPUs, not just CPUs, read from and write to storage directly” with hundreds of thousands of GPU threads, a data-access framework called SCADA, and a Vera BlueField-4 STX storage processor that NVIDIA claims delivers up to 3.21x the throughput of x86 in compression and encryption pipelines. Those are NVIDIA’s own numbers, vendor-run and untested by third parties, and the Storage-Next initiative behind them lists more than 40 vendors, DDN, KIOXIA, and Micron among them.

In the second installment, NVIDIA’s Vera validated datacenter LPDDR as a tier while Qualcomm pitched HBC at the same wall from inside the accelerator package. Here, NVIDIA’s CMX validates the microsecond context tier while SanDisk pitches HBF into the same bytes. The platform owner keeps building the socket; the challenger keeps proposing a different part to fill it.

But the two vendors reach those bytes by different routes, and the distinction is investable.

Picture two ways to store the notebooks. NVIDIA’s CMX is a shared library down the hall. Any GPU on the floor can check a notebook out, so you buy fewer copies, and you wait a little to walk the hall. SanDisk’s HBF is a bigger shelf on the desk. Faster to grab, still the same paper, still wears out when you write. Neither is the live conversation in your hand. That still lives in HBM, the fast memory on the GPU. These are overflow shelves for notes you might need again.

NVIDIA’s public CMX model adds a network-attached flash tier between local SSD and shared storage and stages reusable KV toward system memory and HBM. SanDisk moves flash into the accelerator package through HBF. CMX can pool capacity and reduce part of T_reload. HBF can reduce physical distance and raise local interface bandwidth. Neither turns flash into a latency-equivalent substitute for active KV.

Both routes are governed by the same admission inequality, but they move different terms. The network route primarily changes reload latency, sharing, and queueing. The packaged route primarily changes locality and bandwidth while retaining NAND write, retention, and endurance constraints. The addressable flash demand is therefore gross KV bytes multiplied by the fraction admitted, the probability of reuse before eviction, and the capacity provisioned for the required service level. Pricing every generated KV byte as durable flash demand overstates the opportunity.

None of this arrived unannounced, and the January record is worth re-reading now. The Hierarchy Rewrites carried NVIDIA’s context-memory platform plans, BlueField-4 DPUs managing KV cache lifecycle across HBM, NVMe, and the rack, and put the mechanism on the page: “Runtime policies that intelligently evict and reload from SSD tiers enable workloads that would otherwise be impossible.” We drew the tier in January. NVIDIA gave it a product name in August.

CMX and Storage-Next were on the same Hot Chips agenda. The tutorial room’s fight over the write is in the section above; the fabric route’s turn comes as the CMX qualification lists fill in.

Who Collects in the Tier Below

If the tax moves down the memory stack, someone still collects it. Not by selling the cheapest paper. China can already print that. YMTC is the proof. What is scarce is everything that turns cheap paper into a shelf the exam room will accept: the contracts, the stacking, the software that decides which notes get kept, and the engineering that keeps the paper from wearing through.

Four collectors.

The contract is already collecting. SanDisk pre-sold the paper. Two-thirds of FY2028 bits are spoken for by eight customers, before a single HBF stack ships.

The stacking layer collects only if HBF becomes a real part. Flash wearing HBM’s clothes: stacked dies, tiny vertical wires, a logic chip, the same footprint as the expensive memory. That work pays the bonding and test shops, the same class of names that got paid when HBM got stacked.

The placement layer is the sleeper. Whoever decides which notes stay on the desk and which get thrown out owns how fast the paper wears. That is NVIDIA’s Dynamo, CMX, and BlueField. NVIDIA gets paid at least once or twice. They own the hallway. SanDisk gets paid once, when it sells the part.

The equipment layer collects on thinner paper. Fifty-four percent more bits per wafer per generation is a tools story. The deposition and etch vendors get paid no matter which flash brand wins share.

The Tax Dodge promised that if the tax moves down the hierarchy, we map who collects there. Rick placed the feasibility; the monetization read is mine.

As we wrote in The Bandwidth Tax: “the investable question is where the next unit of value accrues while qualified DRAM dies and stacks remain scarce.” Swap DRAM for NAND and the question survives intact, because bits are not the scarce thing in flash. YMTC’s rise, covered in the bear case below, is the demonstration that raw bits are the most contestable commodity in memory. What is scarce is everything that turns a cheap bit into a qualified tier: the contracts, the packaging, the placement software, and the endurance engineering.

The contract layer collects first, and it is already collecting. The NAND bit agreements are the receipt. Two-thirds of FY2028 bits sold forward to eight data-center customers is a constraint priced at signing. If the margin structure lands anywhere near the deck’s ambition, the LTA is where the tier-below tax shows up in a financial statement first.

The stacking and packaging layer collects if HBF scales. HBF is NAND in HBM’s clothing: stacked dies, through-silicon vias, a logic interface, an HBM-matching footprint. Every dollar of that integration lands on the same class of suppliers the first installment mapped for DRAM: bonding, assembly, and test. The tax dodging HBM’s interposer at the top of the hierarchy re-arrives as stacking complexity at the bottom.

The placement layer is the sleeper. If bytes tier across HBM, LPDDR, TLC, and QLC, then the software and controllers that decide which byte lives where hold the margin lever for the whole hierarchy. That is NVIDIA’s Dynamo, CMX, and BlueField at rack scale, and controller and firmware players below it. Whoever manages the churn owns the endurance budget, and the endurance budget is the cost of goods of a KV tier. In English: the teacher who decides which notebooks get kept decides how fast you have to buy new paper. And note the asymmetry between the two routes. NVIDIA gets paid at least once or twice; they control the whole stack and data flow. SanDisk collects once, when it sells the part. If NVIDIA’s fabric route wins the KV bytes, the collection point is the platform layer NVIDIA already owns and the flash vendor sells drives. If HBF wins them, the collection moves to qualification and packaging. Either way, the bit stays the cheapest part, and the bit side of this is gonna turn into a huge volume game where whoever can absorb the lowest margins at scale will win.

The equipment layer collects on the layer-count race. A supply strategy built on 54 percent more bits per wafer per generation is a deposition and etch intensity story, and it accrues to the NAND tool makers regardless of which flash vendor wins share.

The honest read on timing: I am on the fence about when, and whether, HBF ships in accelerator volume. Late 2028 on the vendor’s own deck means 2029 in practice, and the graveyard of memory-adjacent architectures we walked in The Tax Dodge argues for humility about all of it. But the contract layer does not wait for HBF. The scarcity-forward selling is happening now, on read-mostly workloads that are already proven, and that is the part of this piece that does not depend on a 2029 product date.

And the positioning read, which is mine and not Rick’s: I’ve always been a fan of complex things with compute that have a moat and have a software layer. Obviously that’s NVIDIA, and maybe some of the optics, which are harder to replicate. I wouldn’t try to buy commodity storage, flash, and maybe some memory. It’s just too volatile. Demand is real and healthy and it’s going, but I kinda feel like it’s an extended, bigger super cycle. It’s still a cycle, and people need to understand that.

Where This Breaks

1. Bits drown the tier. NAND overcapacity is structurally easier to create than DRAM overcapacity. Cheap paper is easy to overproduce. DRAM, the expensive memory, needs EUV machines almost nobody else has. Flash does not. China’s YMTC is already the number-three NAND shipper. If the world can print paper this fast, an 80 percent gross-margin target is a claim that contracts beat commodity physics. That bear hits the cell vendors. It actually helps the qualification story. Abundant cheap bits make the accepted shelf the only scarce thing left. Hybrid bonding via Xtacking lowers the barrier further. YMTC is now the number-three NAND player by bit shipments as of Q2 2026 per TrendForce, at 14 percent share, having passed Kioxia, though still fifth by revenue on a consumer-heavy mix, with Fab 4 and Fab 5 online by 2028. Zephyr’s read is that the incumbents’ share loss continues from here, and the history of this industry says supply discipline rarely survives a demand slide this good. This is the bear that attacks the margin ambition directly: an 80 percent gross margin target, held from an 84.6 percent print, published into a rising YMTC supply curve is a claim that contracts beat commodity physics. It changes the slope for every cell vendor. It does not change where the tax collects, and it arguably strengthens the qualification layers, because abundant cheap bits make the qualified tier the only scarce thing left.

2. Compression outruns the pull. Labs got better at taking shorter notes for the same conversation. DeepSeek cut the notebook size about eightfold in a year. If the notes get that much smaller, you need that much less paper. Context windows still got longer, so people are stuffing more into each session even as each page got cheaper. The race is whether shorter notes, or longer sessions, win. SanDisk’s 420-exabyte KV line is a bet that the sessions win. DeepSeek cut KV cache per million tokens from roughly 35 gigabytes in R1 to 4.4 gigabytes in V4, per Zephyr’s recap, and DeepSeek’s own V4 paper corroborates the magnitude, reporting about a tenth of V3.2’s KV cache at million-token context. Every frontier lab is pushing context length up while pushing KV bytes per token down. This is the standing algorithmic-compression short our memory coverage has carried all year, and it lands inside the compression window we flagged as the thesis-breaker. The market is continuously scared of the oversupply of memory and storage, and the fear is that companies just use less and optimize, and that whole bull thesis goes away. SanDisk fell 11 percent in a day on a compression paper in March; we defended memory demand through that selloff, and the defense has held every print since. We’re always gonna Jevons. The cheaper things get, the more you’re gonna use. There’s always gonna be memory optimizations, but it’s not gonna kill stuff. But a 420-exabyte KV forecast is the largest single bet yet that cached state keeps outgrowing compression, and we will not pretend the race is settled. This is the bear that changes direction, not slope: if KV bytes per useful task fall faster than tasks grow, the KV rows of the demand slide deflate and the NAND story shrinks back to staging and data lakes.

3. HBF never ships in volume. HBF slips, ships as a weights-only curiosity, or joins the graveyard of near-memory ideas. The textbooks stay on ordinary SSDs and on LPDDR. The staging pile, the filing cabinet, the bit contracts, and the tool vendors all survive. The 420-exabyte notebook forecast does not, unless some other controller figures out the writes. The timeline has already slipped two years against what we printed in July, the deck’s HBM comparisons have a flagged accuracy problem, and the endurance question forces a hybrid design that keeps HBM in the socket anyway.

Two dated tests. By the end of 2027, a real HBF spec with wear ratings, plus one named chip that will actually use it. That turns the stacked shelf from a slide into a product. By the end of 2028, a hyperscaler or a frontier lab saying out loud that they keep chat notebooks on flash, with the economics attached, and that fetching the notes beats starting the textbook over. Until those two print, the 420-exabyte line is a forecast, not a receipt.

What would change our mind. First, by end-2027: a 1.0-or-later HBF specification carrying endurance and qualification requirements out of the SanDisk and SK Hynix standardization work, plus at least one named accelerator design win, would move HBF from roadmap to product and activate the stacking-layer leg. Second, by end-2028: a hyperscaler or frontier lab disclosing production KV-cache offload on flash with economics attached, reload beating recompute at disclosed context lengths, would validate the KV rows. If neither arrives on schedule, the verdict in plain words is the one I’d give you on a call: things are delayed, it’s not working out exactly how we thought it was gonna happen in theory, and we’re just gonna use more of the same memory. We will write that follow-up, and the verdict will print in the series’ own terms: the tax stayed where it was.

So What?

For two pieces we watched the industry climb around expensive HBM from above: delete the interposer, stack SRAM, promote cheap LPDDR. SanDisk’s investor day is the same move from below. The storage floor reaching up for bytes that used to live in memory. Both sides count bytes. The constraints are not bytes. At the top of the stack the limit is watts. At the bottom it is writes. Which AI bytes belong on NAND has a narrower answer than a zettabyte. The textbooks already belong, and they are already paying. The notebooks are the open question.

Placement, endurance, and compression will settle the notebook question in public over the next two years. The layers that make the tiering decision, contracts, packaging, placement, and tools, are where the tier-below tax collects. The commodity bit is the one thing here we would not underwrite, and YMTC is the reason. That is not a contradiction with the contract layer collecting first: the NBMs price the committed, qualified tier, not the merchant bit.

Long the constraint, not the cell. In the tier below, the constraint is the write.

Like the first two installments, this piece is free and citable on purpose. The thesis is the marketing. The expression, the names and entries across the four collection layers, including the under-covered controller and equipment names, goes to paid subscribers in the follow-up.

Coming Up

The rest of the Hot Chips field notes, beyond the tutorial room reported above, go to paid subscribers this week. The paid follow-up to this piece, the collection map for the tier below with names and entries, follows. Rick and I are also planning a joint Substack chat Q&A on this piece, engineering feasibility and monetization side by side, like the one we ran for The Bandwidth Tax.

One more piece of housekeeping. BEP's principal runs occasional SPVs in private companies across this stack, offered to verified accredited investors under Rule 506(c). If you want to see those deals when they open, join the list at tools.bepresearch.com/syndicate.

What to Watch

Watch the spec, not the slide. A published wear and interface spec is the single most important disclosure in this story. Watch who fills NVIDIA’s named context shelf, ordinary drives over the fabric or HBF in the package. Watch whether a lab admits, in public, that flash is holding live agent notes. Until then, long the constraint, not the cell. In the tier below, the constraint is the write.

• CMX design-ins and the Storage-Next qualification lists. The Hot Chips memory track was the first public round of that argument; the tutorial section above reports it.

• NAND bit agreement expansion beyond the current eight data-center customers, and whether competitors sign equivalents. The contract layer is the tell that collects first.

• The next frontier-model KV disclosure. Another DeepSeek-scale compression step would pressure the 420EB row directly.

• YMTC Fab 4 and Fab 5 progress and quarterly bit share. The overcapacity bear runs through Wuhan.

• SK Hynix’s dual position: the HBM leader co-authoring the flash tier below it is either a hedge or a confession, and its own HBF commentary will say which.

• Answers to the two questions this piece puts on the record: which workload is HBF’s first qualification target, and whether SanDisk or SK Hynix will publish sustained application bandwidth, p99 latency, package power, and the resulting reduction in GPUs required; and from NVIDIA, at what model size or topology near-package flash reduces the minimum GPU count, and which model state must remain in HBM.

Sources

• SanDisk Investor Day materials, August 13, 2026 (bit shipment forecast, 2030 demand breakdown, BiCS productivity, NAND bit agreements with eight data-center customers, non-GAAP gross margin target, HBF roadmap); primary deck and press releases reviewed for the gross margin and customer-coverage figures

• OCP High Bandwidth Flash Architecture Specification v0.7.0 (high-level base-die specification: UCIe connectivity, 8-high and 16-high stacks, capacity and bandwidth grades)

• Zephyr (Citrini Research), investor-day recap and analytical framings: the conservative-forecast call, the persistent-agent-memory estimate, the YMTC share-loss read, and the HBF weights-only read, credited throughout; X thread flagging the deck’s HBM spec comparisons

• NVIDIA blog, FMS 2026 (Jason Hardy, Aug 4, 2026): CMX Context Memory Storage, cuFile open-sourcing, SCADA, Vera BlueField-4 STX claims, Storage-Next vendor list

• SanDisk, “Memory-Centric AI” (company blog)

• SanDisk, “Scaling Beyond the Wall: Inside SanDisk’s High Bandwidth Flash for AI” (company blog)

• SanDisk and SK Hynix, HBF global standardization (Feb 25, 2026)

• DeepSeek-V4 technical report (arXiv, 2026): roughly 10 percent of V3.2 KV cache at million-token context; the R1 35 GB and V4 4.4 GB per-million-token working figures are per Zephyr’s recap

• David Patterson et al., January 2026 paper on memory-hierarchy research directions (HBF characterization)

• TrendForce Q2 2026 NAND data as reported (Caixin, Electronics Weekly): YMTC number three by bit shipments at 14 percent, fifth by revenue

• Hot Chips 2026 memory tutorial Slack, #t1-memory_technology, August 23, 2026 (Oxmiq “HBF in AI Compute” session Q&A; Samsung and Micron speaker responses); attendee assertions marked as such in the text

Related BEP Research

• The Bandwidth Tax

• The Tax Dodge: Qualcomm’s 133 TB/s Isn’t What You Think

• The Hierarchy Rewrites

• TurboQuant and Dynamo: Why the Market Sold the Wrong Story

• NVIDIA Vera: When CPU Latency Becomes GPU Economics

About the Authors

Rui “Rick” Xie is a researcher in computer systems, memory architecture, and AI infrastructure (rickxie.cn). He holds a PhD from Rensselaer Polytechnic Institute, where his research focused on memory architecture and computer systems, with a particular interest in DRAM and SSDs. For this piece he set the framing boundaries, built the endurance, write-amplification, and reload arithmetic through public sources, and read the OCP HBF v0.7.0 specification and SanDisk’s primary investor-day materials. His views and analysis are his own and do not represent his employer.

Ben Pouladian is the publisher of BEP Research and CEO of BEP Holdings. He led the investment framing and the collection-map analysis. Framings are credited to their originators throughout. More at www.bepresearch.com

Disclosure: BEP Research’s principal is long NVDA, NOW, LITE, CRDO, TSEM, ALAB, WOLF, SMCI, BE, NBIS, and ORCL (2027 LEAPS). No position in SNDK or MU. This is investment research, not investment advice. Do your own work.

I sat in on their Hot Chips pre-brief Friday, and this note publishes the minute that embargo lifts.

Groq LPX is in full production. We said late Q3; it is shipping now. Spectrum X Multiplane is in silicon. SpaceX AI is putting Vera into production at gigawatt scale behind Grok. Wednesday grades the rest.

Five Beats, Read Honestly

NVIDIA has beaten its own revenue guide five quarters in a row: $1.1 billion over, then $1.7, $3.0, $3.1, and $3.6 billion last quarter. The beat rate peaked at 5.6 percent five quarters ago and has compressed three straight. Either the company is learning to guide closer to its own reality at enormous scale, or, in a supply-constrained year, shrinking upside is just what a supply cap looks like.

And May is the tape warning: $3.6 billion over its own guide, $2.8 billion over street, and a $4.8 billion raise, and the stock still lost roughly 23 points to the SMH in twenty sessions, three months of sideways since, $223.47 on print day to $214.72 on Friday. Our read is Rubin timing, first production shipments moving to Q3.

NVIDIA reports Wednesday: results after the close, call at 2pm Pacific, middle of Hot Chips week. This is the first Audit: dated claims, marked to what NVIDIA has printed since, and what Wednesday can still break. The Token Dollar loop, the optics read, the collateral thesis, the roadmap and the margin line all report this week.

Everyone is already expecting another blowout number. Our own group chats went into the week at portfolio highs, calling new all-time highs before earnings, revenue-is-exploding as the base case, while the indices sat red. Sentiment like that is its own bar; the published consensus sits near $91.9 billion on the 42-analyst compilation.

Run the cadence on the $91.0 billion guide and you get $94 to $96 billion. That is a BEP cadence estimate, not a claim about what the marginal buyer carries. A $92 or $93 billion print would beat consensus on every headline and still read light against the pattern.

At $214.72, the close on Friday, August 21, NVIDIA is a $5.2 trillion company at roughly 33 times trailing earnings, about 24 times this fiscal year’s (FY27) consensus earnings. You’re still paying for a lot of untapped revenue potential: the street’s own compilation has revenue near $394 billion this fiscal year, up more than 80 percent, with next year’s estimates centered near $560 billion and running as high as $710 billion. At that price and multiple it still looks inexpensive to us for a business producing this much free cash flow. What breaks it is not Wednesday’s beat. It is a break in margins, meaningful competition, or problems shipping products. Wednesday reads on all three.

That is the free half: the cadence math and the tape. Behind the paywall is the audit itself: the board, every claim dated and marked to what NVIDIA has printed since; the shell put, the 8-K’s $105 billion read at the megawatt and our threshold on the minimums that land with Wednesday’s 10-Q; the three Hot Chips answers marked against the lines they hit; the scorecard, including the one line that can grade against us; Thompson’s bear graded on the line he actually named; and the three states we will call the print by on Wednesday night, stated before the fact.

Read more

Co-authored with Rui “Rick” Xie (rickxie.cn), a memory-systems researcher with a PhD from Rensselaer Polytechnic Institute whose work runs from DRAM to SSDs. This is the second installment in our memory series. Rick leads the engineering and standards analysis; Ben Pouladian leads the investment framing.

Six weeks ago Rick and I published The Bandwidth Tax and argued that the real cost of HBM is levied on the system, not the stack: the interposer, the base die, the packaging, the cooling, everything a platform pays around each cube to turn a faster stack into delivered bandwidth. We said the durable trade was one level up from the DRAM cell: long the constraint, and long whoever gets paid to relieve it.

On June 24, at its Investor Day, Qualcomm put a commercial roadmap behind an architecture built to reduce its dependence on the conventional HBM interposer. It stacks the accelerator under the memory, and Qualcomm says the silicon interposer is no longer needed.

Rick’s verification and my framing point in different directions, so let me be exact about what we are underwriting. Not HBC working: Rick’s read, below, is that the headline numbers are unproven and the base case is a niche. We are underwriting something narrower and more actionable: the packaging premium is currently priced as though nobody is trying to route around it, and that is no longer true.

What Qualcomm Actually Announced

The announcement that moved the stock was the guidance: non-handset revenue expectations for FY29 nearly doubled, from $22 billion to $40 billion, with data center as the major driver. The market is processing that number through a familiar template, which is whether Qualcomm can pull off the MediaTek trade, out of the handset doghouse and into a data-center re-rating. Citrini Research made that case in July in “All Along the AI Watchtower,” and described the hardware side as “taking detours around the HBM toll booth.” The metaphor is theirs; the detour is worth mapping.

It has a name: High Bandwidth Compute. Instead of an accelerator die connected sideways to HBM stacks across a silicon interposer, HBC takes the AI accelerator out of the SoC and places it directly beneath a DRAM stack, connected vertically by through-silicon vias. Bandwidth flows up through the whole die area rather than out through the edges. Qualcomm says multiple HBC stacks can be deployed using standard packaging, without the HBM silicon interposer. The AI200 rack ships this year without HBC. HBC debuts with AI250 in 2027, ahead of an AI300 and a server CPU in 2028.

One thing Qualcomm has not said is what that memory actually is. The company has never identified HBC’s DRAM as LPDDR, and the coverage has assumed it, including our own first draft of this piece. Hold that thought, because the entire investment leg below turns on it.

The claimed numbers, from the Investor Day: 6x the bandwidth per watt of HBM. 200x the capacity per watt of SRAM. 4–8x better decode performance per watt and on a TCO basis. A headline figure of 133 TB/s. Tony Pialis, who runs Qualcomm’s data center business, describes the result as SRAM-like performance with stacked-memory density, built out of everything a mobile-chip designer was forced to get good at: low power, small area, no exotic packaging.

Every one of those numbers is Qualcomm’s own.

The Claims That Don’t Survive Contact

Qualcomm describes 133 TB/s as effective memory bandwidth, and explains that HBC makes internal memory bandwidth available to local computation while the main accelerator keeps the orchestration-heavy work. Read it as workload bandwidth consumed inside HBC, not raw bandwidth exposed to the main accelerator.

No defensible sustained number can be reconstructed from LPDDR5X or LPDDR6 pin rates, which is the first thing we tried and the first thing that failed. Qualcomm has disclosed neither the HBC memory protocol nor its interface width, channel structure, or stack configuration. Pin-level arithmetic would add precision the public architecture does not support.

Qualcomm defines the sixfold bandwidth-per-watt figure as an internal estimate against competing published specifications, normalized at card level. The two-hundredfold SRAM comparison normalizes at rack level. Different boundaries, so the two numbers were never a common benchmark and cannot be quoted as one. The baseline will also move before HBC samples: Micron currently specifies more than 2.8 TB/s for HBM4 and Samsung up to 3.6 TB/s for HBM4E, so any efficiency claim should be retested against the HBM systems shipping when AI250 becomes available.

Classification is where the investment case and the engineering case meet. HBC is near-memory compute, not processing-in-memory. Qualcomm describes a compute die handling data-bound operations beneath DRAM bonded directly onto logic, which puts a separate logic layer under the memory rather than computation inside the DRAM array. That distinction is not pedantry, and it is one this newsletter has been on the record about since January, quoting Patterson in The Hierarchy Rewrites: “better suited for LLM inference than processing-in-memory (PIM) because shards can be 1,000× larger.”

Which makes software the central test, not bandwidth. Qualcomm has not detailed operator coverage, compiler partitioning, or fallback behavior. HBC creates system value only when the traffic it eliminates locally exceeds the coordination cost of getting work to it. And the dodge is partial by construction: it may remove the large silicon interposer, but it still needs direct DRAM-to-logic bonding and 3D integration, which relocates packaging complexity into the vertical stack rather than deleting it.

The decisive evidence will be sustained end-to-end tokens per joule under disclosed workload and thermal conditions. Until Qualcomm publishes the bandwidth boundary, the workload methodology, and measured system results, the architecture is credible and the headline numbers are not.

Why Every Previous Attempt Died

Processing-in-memory has produced good benchmark numbers for a decade without producing a platform. Samsung reported almost 2.5x higher system performance and more than 60% lower energy from HBM-PIM. SK Hynix was still demonstrating AiMX on H100s with vLLM in 2025. UPMEM shipped programmable DRAM modules and a real software stack. None of it reached a named hyperscale deployment, which shows technical viability alone was never enough to build a platform.

It was ownership and software. The economics asked a memory vendor to sell compute it had no ecosystem to support, and the programming model asked customers to restructure their code for the memory’s benefit. HBC addresses part of the ownership problem: one vendor now designs the memory-side compute, the accelerator, the runtime, and the rack, which is a genuinely different starting position than SK Hynix bolting AiMX onto someone else’s H100. That integration reduces friction. It does not remove the software, execution, and ecosystem constraints that limited the earlier systems.

The second problem is untouched. CUDA is not a compiler. It is a decade of optimized libraries, profilers, and deployment tooling that customers have already built their models, their reliability processes, and their hiring around. ROCm is a thinner version of the same thing and it has taken AMD years to get that far. Against that, an incremental bandwidth or energy advantage does not clear the migration bar, and HBC still splits every real model across two compute domains. The variable that matters is the fraction of the token path that stays inside HBC after partitioning, because high internal bandwidth buys little when control keeps returning to the main accelerator.

Then there is heat, and this is where the architecture is most exposed. HBC parks active logic in direct three-dimensional proximity to temperature-sensitive DRAM, which is precisely the separation the interposer was providing. Prior die-stacked research used an 85°C DRAM threshold, modeled refresh doubling for each additional 10°C, and found allowable logic power ranging from about 8.5 W under passive cooling to about 55 W under active cooling. Those numbers do not predict HBC. They do establish that cooling and stack orientation set a ceiling on how much compute can run under a DRAM tower, and Qualcomm has disclosed no logic power, no DRAM junction temperature, and no sustained performance after thermal steady state.

Rick’s base case: HBC becomes a specialized decode or offload tier inside selected Qualcomm systems while GPU and HBM platforms keep the broader inference market. Qualcomm has built a more integrated version of the PIM proposition, and most of the conditions that stopped earlier systems from becoming platforms are still standing.

Everyone Is Routing Around Something

That base case is survivable for the trade, because the trade does not depend on Qualcomm winning. It depends on Qualcomm not being alone, and it isn’t.

HBC is not the first architecture built to avoid the HBM bill, and this newsletter documented an earlier one under its own byline. From The Fourth Piece Ships in March: “The Groq V3 uses on-chip SRAM on what appears to be a standard packaging process. No CoWoS interposer. No HBM stacks.” Positron got there with commodity LPDDR before that. Qualcomm’s version is the largest and best-funded instance of a move already in progress, which is a stronger fact for the thesis than being first would have been. One company routing around advanced packaging is a science project. Three is a pattern.

The pattern is the one this newsletter has been tracking all year. As I wrote in The Memory Wars: “NVIDIA Feynman (2028) won’t fight SRAM physics. It routes around them: 3D-stacked SRAM using AMD X3D-style hybrid bonding.” Memory vendors are routing around thermal compression bonding with hybrid bonding; NVIDIA is routing around SRAM scaling with stacked cache. Qualcomm’s version is the most aggressive, and it takes on a whole new cost structure in logic-memory integration to get there.

Four weeks after the Investor Day, NVIDIA supplied the data point the LPDDR argument was missing. Vera pairs 88 custom cores with SOCAMM2 LPDDR5X at up to 1.2 TB/s per socket, 256 GB to 1.5 TB per socket, field-replaceable, multi-vendor sourcing explicitly enabled. As I wrote in NVIDIA Vera: When CPU Latency Becomes GPU Economics: “Datacenter LPDDR is now a named product line.” At maximum configuration that is 384 TB of LPDDR5X in one rack.

Vera is not itself a dodge, and it matters not to file it as one. It is a CPU, and the Rubin GPUs beside it still run HBM4. What it establishes is narrower and more useful: the largest accelerator vendor in the world has qualified datacenter LPDDR as a real memory tier, with multi-vendor sourcing, at rack scale. The question NVIDIA left open is whether that capacity becomes addressable KV cache. If it does, the tier stops being one company’s bet.

The Bandwidth Tax gave the pattern its rule: you route around a wall rather than push through it, and you pay for the route. Hybrid bonding pays in packaging complexity, stacked SRAM in bonding yield, and both bills come due at the same suppliers. So the question for HBC is not whether the tax disappears, since our thesis says it never does, but where it gets collected instead.

Who Collects If It Works

HBC is bearish the large-interposer premium and neutral-to-bullish the memory oligopoly. Those are not the same trade, and the market is pricing them as one.

The layer it disintermediates is the layer collecting today: the large interposer, CoWoS capacity, and the 30%-plus HBM4 manufacturing premium we documented in the parent piece. An architecture that tiles accelerator-memory stacks without a large silicon interposer is a direct attack on that premium, arriving exactly as the premium is at its highest.

But whatever DRAM ends up inside an HBC stack, it is high-density memory, and high-density memory is sold by Micron, Samsung, and SK Hynix. The same three vendors. The bear case for HBM packaging is not a bear case for DRAM. It is a mix-shift within the oligopoly’s own product lines, and the incumbents have already told us how that story goes: as I wrote in Micron Just Proved the Memory Thesis, “Sanjay confirmed that non-HBM margins are currently higher than HBM margins.” If the next bandwidth architecture consumes non-HBM DRAM at data-center scale, the same vendors sell the dodge with one hand while they sell the tax with the other.

Now the part that keeps this from being a clean short. Qualcomm has not identified the memory technology, named a DRAM supplier, or said who performs final logic-and-memory integration, so treat the mix-shift as a scenario with a supplier list still missing from it. And the dodge may partially refund what it avoids: even skipping large interposers and some CoWoS capacity, HBC still consumes wafer thinning, bonding, known-good-die infrastructure, and advanced test. The pressure falls on the scarcity premium attached to large interposers and conventional accelerator-to-HBM integration. It does not fall on packaging as an activity.

Which is why this is a migration, not a disappearance, and the destination is one we have already named. The Hierarchy Rewrites put the candidates for exactly this direction on the page in January: TSMC SoIC, Intel Foveros, and the hybrid-bonding equipment suppliers, Besi, EVG and SUSS. Direct DRAM-to-logic bonding is their process, not CoWoS’s, which makes them plausible beneficiaries rather than confirmed ones, since Qualcomm has disclosed neither its assembly flow nor its suppliers. If HBC and the architectures chasing it scale, value moves from interposer area toward bonding tools, stack yield, thermal integration, and test. The packaging layer deserves the scarcity premium it is charging today, 2027 is a long way out, and CoWoS is sold out regardless. None of that changes the fact that the most valuable customers in the world are now engineering their way off that particular tax roll, and toll roads get repriced on the first credible detour, not the last car through.

On Qualcomm itself: the re-rating comp is MediaTek, which came into the year under 20x NTM earnings and re-rated hard once TPU exposure pulled it out of the handset bucket. The credibility gap is real, since this is the company that shipped Centriq in 2017 and walked away, but the validation stack is different now: Alphawave for SerDes, two hyperscaler custom-silicon programs contributing in FY27, a multi-generation Meta CPU agreement with production in 2H28, and the pending Modular acquisition for the software layer. The data-center line is more than $15 billion of that $40 billion, not additive to it: Qualcomm’s own bridge is $10 billion automotive, more than $14 billion IoT, and more than $15 billion data center. TD Cowen’s read, which Qualcomm has not confirmed, is that the target counts only existing customers. Named customers and measured system performance are the evidence; the guidance is the option.

Where This Breaks

The graveyard wins again. Every number in this piece is Qualcomm’s own, with zero third-party benchmarks and first silicon in 2027, from a management team whose last data-center entry ended in a walkaway. The specific failure mode is high internal bandwidth that never becomes end-to-end inference, with operator coverage, partitioning, fallback execution, and thermal throttling eating the local gain. In that case the AI250 launch is a slide deck with a good chart in it, and the packaging premium keeps collecting uninterrupted.

The stack outruns the dodge. Our own correction from the parent piece cuts against the urgency here: HBM is improving per stack, per watt, and per delivered TB/s every generation, and it is sold out through the exact window in which HBC must prove itself. The competition is not only HBM4E and hybrid bonding. Larger caches, lower-precision weights, speculative decoding, and KV-cache optimization each shrink the value of near-memory execution before AI250 reaches production. The dodge only pays if the tax keeps rising.

HBC works and the incumbents still collect. High-density DRAM from the same memory oligopoly, potential overlap with the same constrained stacking and test capacity, and a bill of materials that reroutes dollars rather than removing them. Bonding, stack yield, test, cooling, and customer validation could absorb everything saved by deleting the interposer. In that outcome the architecture survives, our asymmetry holds, and the profit pool barely moves. A reader can accept this entire piece and still conclude the only durable longs are the memory vendors and foundries they already own.

What would change our mind. Two conditions, by end-2027. Qualcomm discloses how effective bandwidth is calculated, in enough detail to separate raw DRAM bandwidth from locally consumed traffic from bandwidth visible to the accelerator. And independent testing confirms at least half the claimed decode perf/watt advantage against contemporary HBM systems under matched model quality, latency, and thermal conditions, with at least one named hyperscaler or frontier lab in production. Both, and the packaging-premium leg activates. Either one missing, and we write the follow-up saying the tax held.

Stop Asking Whether Qualcomm Catches NVIDIA

That is not what the Investor Day was about, and Qualcomm does not need to catch NVIDIA for HBC to matter. What matters is that the cost of moving data is now determining the shape of the machine, and three separate accelerator teams have concluded that the cheapest way to pay the bandwidth tax is to build something that owes less of it.

The Bandwidth Tax said the money made around each HBM stack is growing faster than the money made inside it. HBC tests the corollary: when the tax gets high enough, someone builds a machine that doesn’t pay it. Whether this particular machine works is still unproven, and the graveyard puts the burden of proof on Qualcomm. The direction is set regardless.

The tax may move. Qualcomm still has to prove the detour is cheaper, and until it does, stay long the constraint.

Like The Bandwidth Tax before it, this piece is free and citable on purpose. The thesis is the marketing. The trade expression, the names and entries across the collection-point migration, goes to paid subscribers in the follow-up.

The Next Detour

While this piece was in final review, the tier below started making the same argument. SanDisk calls it High Bandwidth Flash: NAND stacked and interfaced like HBM, pitched as memory rather than storage, and it ran through the company’s investor day last week. The pitch writes itself, which is exactly why it gets the same treatment Qualcomm got here.

Rick and I are already working it, and the boundaries preview the piece. HBF is moving from a concept into a standards and sampling story, and NAND endurance is not solved, so the honest question is not whether flash can deliver more bandwidth. It is which AI bytes actually belong on NAND. Model weights are read-mostly and plausible. KV cache could be the far larger demand pool, but it is a write stream, and the write stream carries endurance, placement, and reload costs the slide decks do not price. HBF ends up one test case inside the bigger story, which is AI pulling NAND into memory tiers it has never had to serve. If the tax moves down the hierarchy, we will map who collects there too.

That will be the third installment. The first two lived on the DRAM side of Rick’s research. The next one crosses to flash, which is the other half of it.

A housekeeping note: I will be at Hot Chips next week, where the memory track argues several of this piece’s open questions in public. If you will be there, say hello. Field notes to follow.

What to Watch

• AI200 rack deployments this year: named customers, not press releases. HBC is not in this product.

• Any HBC silicon disclosure with sustained rather than aggregate bandwidth, and the first third-party benchmark.

• DRAM supplier disclosures for HBC, and data-center qualification signals from Micron, Samsung, and SK Hynix: the mix-shift tell.

• Meta C1000 production ramp in 2H28, and whether either hyperscaler custom program is publicly named in FY27.

• Whether NVIDIA and AMD answer with LPDDR tiers or near-memory configurations of their own.

• CoWoS pricing at the margin in 2027, and hybrid-bonding tool orders at Besi, EVG and SUSS as the offset.

Sources

• Qualcomm Investor Day materials, June 24, 2026 (FY29 guidance, HBC architecture, AI roadmap)

• Qualcomm Investor Day press release, June 24, 2026 — FY29 segment bridge: $10B automotive, >$14B IoT, >$15B data center within the $40B non-handset target

• Qualcomm, “HBC near-memory computing” (OnQ, July 2026)

• Qualcomm AI accelerators, product/expertise page

• Citrini Research, “All Along the AI Watchtower” (July 19, 2026) — source of the “toll booth” framing

• BEP Research, “The Bandwidth Tax” (July 9, 2026) — HBM4/HBM4E vendor disclosures, JEDEC JESD238B.01/JESD270-4, TrendForce HBM4 premium

• Samsung, in-memory processing (HBM-PIM)

• SK Hynix, AI Infra Summit 2025 (AiMX with H100 / vLLM)

• SK Hynix, back-end process series, wafer-level packaging / TSV stacking

• UPMEM technology overview

• Micron HBM2E technical marketing brief (known-good stacked die)

• Eckert et al., die-stacked PIM thermal analysis (85°C threshold, refresh doubling, 8.5 W passive / 55 W active)

• NVIDIA CUDA Toolkit

• AMD ROCm documentation

• TD Cowen coverage of Qualcomm Investor Day (as reported)

Related BEP Research

• The Bandwidth Tax

• The Fourth Piece Ships: NVIDIA’s Groq

• The Hierarchy Rewrites

• NVIDIA Vera: When CPU Latency Becomes GPU Economics

• The Packaging Paradox: Why CoWoS, Not 2nm, Is the Real AI Bottleneck

• The Shoreline Problem: A Bear Case on HBM

About the Authors

Rui “Rick” Xie is a researcher in computer systems, memory architecture, and AI infrastructure (rickxie.cn). He holds a PhD from Rensselaer Polytechnic Institute, where his research focused on memory architecture and computer systems, with a particular interest in DRAM and SSDs. For this piece he worked from Qualcomm’s public HBC disclosures and the prior near-memory literature directly, established the different measurement boundaries sitting behind the headline numbers, refused a sustained-bandwidth reconstruction the public architecture does not support, and set the conditions under which the architecture would be proven. His views and analysis are his own and do not represent his employer.

Ben Pouladian is the publisher of BEP Research and CEO of BEP Holdings. He led the investment framing and the margin-capture analysis. Framings are credited to their originators throughout. More at www.bepresearch.com

Disclosure: The author holds positions in NVDA, NOW, LITE, CRDO, TSEM, ALAB, WOLF, SMCI, BE, NBIS, and ORCL (2027 LEAPS). No position in QCOM. This is investment research, not investment advice. Do your own work.

On August 5, Jeff Dean left Google after twenty-seven years. He didn’t leave to build another model. He left to co-found Discovery Loop, a public-benefit corporation whose stated purpose is to automate complex, multi-step science experiments end to end, with Oriol Vinyals, Quoc Le, and Sanjay Ghemawat as co-founders and Google itself as founding investor, cloud partner, and supplier of the first year of compute.

The same day, Demis Hassabis handed off day-to-day control of Google DeepMind, took its chair and Alphabet’s chief-scientist role, and said he would lean further into Isomorphic Labs, the drug company he spun out of DeepMind. Alphabet fell about five percent. And a week before either of those, DeepMind had disbanded the team behind AlphaFold, the work that won the 2024 Nobel Prize in Chemistry, and dispersed it into Gemini and into Isomorphic. John Jumper, who shared that Nobel, had already left for Anthropic in June.

Three moves, one week, all pointing the same direction: away from building the model, toward running the experiment.

I’ve been sitting on this note for a while. I held it through July on purpose. I didn’t want to put a thesis measured in years in front of people in the middle of a high-beta drawdown, when the tape is punishing anything that asks for patience and nobody wants to hear about a decade. After this week, waiting stopped making sense.

It’s still a strange thing for me to be writing about.

If you subscribe to me, you subscribe for semiconductors. Memory, optics, packaging, power, the physical layers the AI buildout has to pass through. So this note owes you an explanation, because I’m about to spend it on drug discovery, and I want you to see that it isn’t a departure. It’s the same trade, one layer of the economy over.

I didn’t come to this as an investor. I came to it after I lost both of my parents to cancer. My mother’s doctor had found a Phase 2 drug for her. The company wouldn’t manufacture it for compassionate use. That’s the part I’ve never been able to put down: the molecule existed, it had a name, and the reason she couldn’t have it was that nobody would make it. I’ve spent enough time in cancer clinics to not wish them on anyone, and to notice that most of what gets handed out there buys time rather than a cure. Standing in that gap turned private grief into a question: what’s actually broken inside biotech, and is any of it about to change? I went looking with the same lens I bring to chips, and I found the same structure underneath.

The reason this stopped feeling abstract came out of the world I actually cover. In June, three days after Anthropic launched its most capable model, the U.S. government invoked export-control authority and ordered access to Mythos suspended for every foreign national overnight. It was walked back weeks later, but the lesson wasn’t: companies that had built their products around a model they only rented watched a decision made in another building switch it off.

Alex Karp put the sharp version on CNBC on July 1, enterprises renting frontier models get “no value” while the labs “get my IP,” and what customers actually want is “control over their compute, their models, their data stack, and their alpha.” Weeks earlier Anthropic had blindsided its own partner Figma by shipping a near-identical tool of its own. The model you rent, fed by the data you hand over, becomes the product that replaces you.

I made the enterprise version of this case in December, in NVIDIA Nemotron 3: “own the data, customize the model, and eliminate vendor lock-in.” Your data should work for you, not train someone else’s model and put you out of business.

Which brings me to what almost nobody is pricing yet. The next place that rule bites hardest is science, where owning your model is inseparable from owning the instrument that generates its data, the lab bench itself. And this has stopped being only a corporate question. China has made biotechnology a national project, with a declared goal of leading the global bioeconomy by 2035 and biomanufacturing named a strategic priority in its latest five-year plan. Whoever owns the measurement layer, the instruments and the proprietary data they throw off, owns the raw material of biological AI. That isn’t a race the United States can afford to rent its way through, even though the same instrument makers book real revenue inside China today, which means a hard decoupling would cut both ways. This isn’t hypothetical, and it’s worth knowing which instruments are already in scope. In January 2025 the U.S. put export controls on high-parameter flow cytometers and certain mass-spectrometry equipment, explicitly because that hardware generates the high-quality biological data that trains AI models. That isn’t a distant category. It’s Bruker’s timsTOF, Thermo’s Orbitrap, Beckman inside Danaher, the exact franchises this note is about. China then blocked Illumina’s sequencers on its side. I’d not be surprised to see controls eventually reach further into the instrument stack, and it’s a two-way risk on the very names I’m telling you to own.

The Bottleneck Always Migrates

Here is the thing I keep coming back to in my own coverage. The bottleneck isn’t the GPUs. It’s powered land, and then getting the thing into the datacenter, the build-out and the people actually doing the building. It’s a physical constraint, not a compute one. Once you have enough compute the whole thing settles down, but right now the problem is getting enough electricians and contractors onto sites without the neighbors complaining and stalling the project. That’s what the constraint actually looks like when you go stand next to it.

The market fixates on the glamorous layer and misprices that. In datacenters the constraint was never really the GPU; it migrated to memory, to CoWoS packaging, to power, and the durable way to own that buildout was the layer every buyer had to pass through no matter which model or hyperscaler won. I made that argument most directly in The Watt Tax: “This is the trade. Not the valuation argument. The bottleneck.” Find the layer that gates everything, own it, and you stop having to guess the winner. That rule now points at a lab bench.

Why Biology Is the Hardest Business in the World

There’s just a lot of people, and time, and experimentation, to get to one molecule. And a lot of the real cost isn’t even there, it’s in the clinical trials. But if you can model that with AI, or model the human body with enough compute, maybe some of it speeds up and we get to cures faster. That’s the whole bet, and I want to be careful not to oversell it.

Start with the economics, because they would horrify any semiconductor investor. It takes ten to fifteen years and well over a billion dollars to bring one drug to market. Roughly nine of every ten that reach human trials fail, most in Phase 2, after most of the money is already spent. And drugs approved per research dollar has fallen for decades, an unsolved bottleneck with a name: Eroom’s Law, Moore’s Law spelled backwards, the exact inverse of the cost curve that made everything we cover cheap. Unsolved bottlenecks, when a new tool finally cracks them, are where outsized returns have historically hidden.

The Lab Is a Loop

The simplest way I think about it’s inventing a better key for a lock. A computer can imagine thousands of keys. Somebody still has to cut the metal and test whether it turns. The slow, painful, expensive part is cutting the metal and finding out.

The mental model that organizes the sector is a loop. A modern discovery lab is a closed loop with four steps: Design (an AI proposes a molecule), Make (it gets synthesized), Test (instruments measure what it actually does), Learn (the result trains the next design). Every AI-drug-discovery company is a race to spin that loop faster.

And here is the part the market has backwards. Venture money, headlines, and the Nobel all flow to Design, the model. But Design is the least constrained step: models diffuse in months, every lab rents the same GPUs, the algorithms are nearly a public good. The binding constraint is Test, physically finding out which molecules actually bind, fold, permeate, and don’t kill a cell, on instruments only a handful of companies can build. The scarce input to a great biology model isn’t compute but measured data at scale. The moat is the measurement.

The Stack, One Layer at a Time

To see where the value sits, you have to see what actually happens at each station of the loop. What follows is the whole machine in plain language, nine layers, following a molecule from idea to drug. The full primer and the interactive map take each one apart; this is one level down from the headline.

Before the walk, where my own attention actually goes. Design, make, and the hands are the most interesting to me. The top half of the stack is the meat. Make especially, because of companies like Twist and IDT and GenScript: you’re about to have so many things you could build and test that the world is your oyster. You start making more and more of them and hoping one of them hits. That’s really the point for me. When AI makes ideas cheap, somebody still has to absorb all of that physical throughput. That’s where the make layer gets interesting, and it’s why I moved weight into it this week.

1. Design, where the computer dreams up the molecule. A drug is a small object that has to fit against a much larger one, usually a protein, the folded machine inside a cell that a disease has broken. For fifty years the hard part was seeing the protein’s exact 3D shape. In 2020, an AI called AlphaFold largely cracked the folding problem, predicting a protein’s shape from its sequence alone, work that won the 2024 Nobel in chemistry. Now software can invent candidate molecules to fit that shape by the million, for the cost of electricity. This is the glamorous layer everyone funds. It’s also the least defensible, because the models leak out into the world within months. Underneath all of it sits NVIDIA, whose chips and software (BioNeMo) run everyone’s models. The durable position here is the toll road underneath, not the car riding on it.

2. Make, where the design becomes real matter. A molecule on a screen is worthless until someone physically builds it. That means writing custom DNA base by base, or synthesizing a compound step by step. It’s slower and harder than it sounds, and turning a digital idea into a physical sample can still take weeks. The companies that endure here, like Twist Bioscience, are the ones that sell a consumable you reorder forever, not a clever machine you buy once. The graveyard of this layer (Amyris, Zymergen) is full of impressive platforms that had no repeat-purchase razor underneath them.

3. Test, where you find out if it actually works. This is the one that matters. A model can guess that a molecule will work. Knowing requires physically measuring the real thing against a real target: does it bind, does it fold, does it poison a cell? That measurement happens on instruments that cost as much as houses, a mass spectrometer that weighs molecules, an NMR machine that reads their structure with a superconducting magnet, a cryo-electron microscope that photographs a frozen protein atom by atom. Only a handful of companies on earth can build these, and the science is decades deep. Bruker has a near-monopoly on research NMR; Thermo Fisher owns the high-end microscopes and mass specs. This is the layer that produces the measured data every AI model is starving for, and it’s the hardest in the stack to copy. It’s the moat.

4. Readout, where you read biology directly. Related to Test, but pointed at reading the code of life itself, sequencing DNA, counting the proteins in a cell. Reading DNA is now cheap and industrialized, an Illumina near-monopoly under attack from cheaper rivals. Reading proteins, which are what drugs actually act on, is still unsolved, and that frontier is where the newer public names are fighting.

5. Hands, the robots that run the experiments. Someone has to move the liquids, load the plates, and run the machines thousands of times, and increasingly that someone is a robot. But the robot arm is becoming a commodity. The real value, and the real lock-in, is the scheduling software that tells a whole room of instruments what to do and when. Most of the best names here are still private.

6. Reagents, the sugar and cups of science. Every experiment burns physical supplies: the chemicals, the antibodies, the cell food, the plastic tips and dishes. They get used up and reordered, forever. This is the beautiful part for an investor: as AI lets scientists run far more experiments, they burn far more supplies, whether or not any given experiment succeeds. The model doesn’t have to be right for the reagents to sell. It’s a toll on scientific activity itself, and it’s concentrated in a few giants like Thermo Fisher and Danaher.

7. Learn, the memory of the whole loop. All that measured data is worthless if it’s scattered across incompatible machines and formats. The software that captures, cleans, and organizes a lab’s experiments into something an AI can actually learn from is one of the most valuable layers, because whoever becomes the system of record for that data sits at the center of what everyone’s models train on. It’s scarce enough that Siemens paid $5.1 billion for one of these companies (Dotmatics), and another (Benchling) is privately valued near $6 billion.

8. Cloud Lab, renting the whole loop. A handful of companies will now run the entire physical loop for you: you write an experiment as code, their robots run it, and the data comes back, no lab of your own required. The truly “self-driving” version, where the AI itself decides the next experiment, is real but still early, and mostly confined to a few labs. Public exposure here is mostly the big contract research organizations that already run trials for pharma.

9. The Discoverers, the drug companies themselves. Finally, the tenants that assemble the whole stack to actually find medicines, Isomorphic, Recursion, Terray, and dozens more. This is where the headlines and the dreams live. It’s also where the risk lives, because a brilliant platform can still watch its one drug fail in a trial and go to nearly zero. This is the lottery ticket. It isn’t the durable position, which is the whole point of everything above it.

What Terray Showed Me

My conviction didn’t come from a spreadsheet. It came from standing in a room. I’m an angel investor in Terray Therapeutics, and Jacob Berlin walked me through their Los Angeles lab.

The first thing that struck me was that it was quiet. I had expected something industrial. What I got were trays of custom-built, ultra-dense microarrays, each one a grid of microscopic wells, synthesizing and reading chemistry at a scale I couldn’t really hold in my head, and feeding all of it back into the models. On the company’s own figures that runs to roughly a billion unique measurements a quarter, more than three times the entirety of public chemistry data every three months. It felt like an agentic molecule factory, working through combinations that had never existed before.

Humans still set the trays up today. That’s worth saying plainly, because the phrase “self-driving lab” gets thrown around loosely and we aren’t there yet. But seeing it in person is what opened my eyes to the lab in the loop, and to what happens when that last human step goes away. If you have robots and machinery running experiments around the clock, while you’re sleeping, through the weekend, millions of iterations, you generate an amount of data no human-paced lab can approach. Then you comb through it for the needle in the haystack. The bottleneck I had assumed was permanent started looking like an engineering problem.

Why Now, and Not Five Years Ago

The tell that an inflection is real is who starts spending to own it. This summer, the answer was everyone. NVIDIA shipped its BioNeMo Agent Toolkit at the BIO convention, OpenAI released LifeSci-Bench, and Anthropic, the maker of Claude, bought Coefficient Bio, a drug-discovery team of ex-Genentech researchers, for roughly four hundred million dollars, its largest acquisition.

Then the people moved, which is a louder signal than the money, and it’s where I started this note. Dean to Discovery Loop, whose roadmap runs from machine-learning research into hardware design, drug discovery and clean energy. Hassabis stepping back from running DeepMind to lean into a drug company. The AlphaFold team dispersed.

I’ll not pretend to be neutral about this. It felt like validation of where my head has been for months. Dean named the company Discovery Loop. That’s the same four-step diagram I drew a few sections ago, and he left the best AI lab in the world to go automate it rather than to build another model. Hassabis, meanwhile, won a Nobel for predicting protein structure and has now chosen the drug pipeline over the lab that won it. I’d not read the five percent too hard; the market is rational and irrational by turns, and a few departures don’t define a company that size. What matters is that the two people with the best information about what AI can do next both looked at the field and walked toward the bench.

My read on the AlphaFold breakup is less dramatic than the headlines. Once a company gets big enough it becomes too many things. IBM ended up spanning transistors, consulting, cloud and quantum, and at that size office politics and competing priorities make it hard to have impact. I think these people wanted to build something focused, where they could move faster. That’s a more ordinary explanation than betrayal, and I suspect it’s the right one. But the investment consequence is the same either way. A lab won the highest honor in science for a model, published it for free, and less than two years later its own owner concluded the team was worth more dispersed than together. Whatever the motive, the durable seat was never the structure predictor. It’s the pipeline those predictions feed and the instruments that check them.

And then, on August 6, while I was finishing this, the cleanest illustration of the entire argument arrived from Stanford and the Arc Institute. A generative model called Evo designed complete viral genomes from scratch, the first time whole genomes have been written by an AI rather than copied from nature. The team took 285 of those designs, physically synthesized them, and put them in a dish.

Sixteen worked.

That ratio is the thesis of this note, run as an experiment by someone else and published in a journal. The model generated 285 plausible genomes at essentially zero marginal cost, and roughly ninety-four percent of them were dead on arrival. There was no way to know which sixteen were alive except to write all 285 into physical DNA and test them on a bench. Design was the cheap half. The truth was only available by measurement. And note what those 285 were, commercially: 285 synthesis orders and 285 assays, which is the demand curve I’ve been describing all the way through this note, showing up in a single experiment.

Be careful with the headlines on this one, because the scary reading is close to the opposite of the science. The work, led by Brian Hie at Stanford and published in Science, used the Evo models trained on the genomes of roughly two million bacteriophages. Bacteriophages infect bacteria, not people. The genetic code of viruses that infect humans, animals, or plants was deliberately excluded from the training data, so the model can’t design one. What the AI-designed phage cocktail actually did was clear E. coli that natural bacteriophages could no longer kill. That isn’t a weapon. That’s a candidate therapy for antibiotic-resistant infection, one of the genuinely unsolved problems in medicine, and the first time a medicine of that class has been designed end to end by a model and then proven at a bench.

The biosecurity concern is real and I’m not going to wave it away, but read where the experts actually point. The Johns Hopkins Center for Health Security, writing alongside the paper, warned that the ability to compose viral genomes with generative AI now exists while “the governance to safely steer it does not.” Tom Ellis at Imperial called this the smallest and easiest genome there’s to make, and argued the AI-design threat is overblown next to the far easier path of modifying pathogens that already exist. And Filippa Lentzos at King’s College London made the point that matters most for this note: the highest-leverage place to intervene isn’t the model at all, it’s the moment DNA is physically manufactured, through synthesis screening. Even the people whose job is to worry about misuse conclude that the control point sits in the physical layer rather than the digital one. That’s the same conclusion this note reaches from the other direction, for money instead of safety.

And the data owners aren’t going quietly, which is the tell that the data is the asset. On the All-In podcast this month, David Friedberg, who runs the crop-genomics company Ohalo, described Anthropic approaching large life-sciences companies to pool their proprietary data into a new model in exchange for early access and an NDA. His read, and nearly everyone’s he had spoken with: they’re “basically trying to commoditize everyone’s business.” A drug company’s experiments produce the one asset it owns, so they’re saying no and building their own models instead. They have worked out where the scarce asset sits, and it isn’t the model.

I want to be fair about one thing here, because my own first read was wrong. I assumed Anthropic was moving in as a competitor to the discovery companies. Talking it through with people closer to it than me, the picture is more specific: a lot of that effort is pointed at orphan and neglected diseases, the ones the current economics simply don’t fund, using AI for a problem nobody else is paying to solve. That’s a refreshing thing to find at the end of a cynical line of reasoning, and it doesn’t change the structural argument. The data still belongs to whoever measured it. But not every model lab entering biology is trying to eat its partners, and I’d rather say that than run the tidier version.

No fully AI-designed drug has yet won FDA approval. That gap is the whole reason I’m writing this. Design raced ahead; the wet-lab validation that actually proves a molecule works hasn’t. Which is exactly where the value is.

Where I Part Ways

It’s more of a picks-and-shovels play, because you don’t know who is going to win, or even who survives if all of this works. That’s the honest starting point. You would rather own the tools and the instrumentation layer, have all of them buy into it, and then you get to win regardless of which one of them turns out to be right.

The intuitive version of this trade is to buy the drug hunters. If AI is going to transform medicine, own the companies using AI to find medicines. I think that’s the wrong expression, because it asks you to be right about a molecule, and being right about a molecule is a coin flip with a ten-year settlement date. Own the bench instead and you get paid on the attempt.

The second consensus is newer and much louder this month: the tools layer is broken. Orders are soft, budgets are tight, sell-side has been cutting, and Bruker just printed a revenue miss against a full-year organic growth guide of one to two percent. I part ways there, and this is the part I’d push hardest on. The tools layer isn’t one exposure. It’s two, running on different clocks. Instruments are capital expenditure, approved once by a committee and deferrable in a lean year, and they’re genuinely cyclical. Consumables are a toll on activity, spent per experiment and reordered next week whether or not anything worked. Those two halves reported within a day of each other this week. One beat its guidance, grew twenty-three percent and raised the year. The other missed and left the year where it was. Own the layer as a single bet and you will conclude the thesis is breaking. Own it as two and you will notice the activity half is compounding for exactly the reason I’m writing this note, which is that models are designing more molecules and somebody has to physically make them.

What this doesn’t require: that any particular drug works, that AI cures anything, that a self-driving lab arrives on schedule, or that you have a view on AGI. It requires one thing. That the number of physical experiments run per year keeps going up.

Two Ways to Go Deeper

I mapped this the way I map any supply chain. The whole thing, how a medicine gets made and how the modern lab works end to end, all nine layers with every meaningful company placed, lives as its own 50-page primer, free to read:

→ The Future of Drug Discovery Is AI-Assisted: the full-stack primer

If you would rather click through the map than read it, the whole stack is also an interactive tool, every layer and company, public and private, filterable in one view, free to explore:

→ The AI Biolab: the interactive supply-chain map

And one more, which is the one I’d actually use if I were you. The framework underneath all of this, the way I score a layer and hunt for the bottleneck, is written up as a file you can hand to an AI and work with yourself. Drop it in a project folder as CLAUDE.md and point Claude Code, or any agent that reads it, at filings, transcripts, decks, or your own notes. It encodes the method: the bottleneck migrates, own the toll not the coin flip, separate capex from consumables, what diffuses can’t be a moat. There are no positions or price targets in it, deliberately. It’s a way of thinking, not my book, and I’d rather you build your own conclusions with it than borrow mine:

→ The research framework (CLAUDE.md, free)

The Basket, and How to Watch It Live

People ask me to just name the one stock. The reason it’s a basket is the same problem as everything else here: you don’t know exactly who is going to win. A basket costs you some upside if one name absolutely dominates. What it buys you is exposure to the layer without pretending you know the single winner ahead of time. I will take that trade.

The honest way to express this is a basket, not a single pick. You can’t know which molecule wins, so you own the bench every discoverer has to buy from no matter whose drug does, weighted toward the moat and sized so the book is roughly right about the theme even when it’s wrong about any single name.

Three anchors give the shape of it, the purest picks-and-shovels seats in the book:

• Bruker (BRKR) — NMR near-monopoly plus timsTOF single-cell proteomics, and a semiconductor-metrology leg riding a second capex wave.

• Thermo Fisher (TMO) — cryo-EM, Orbitrap mass spec, and the widest consumables annuity in the industry.

• NVIDIA (NVDA) — the neutral compute toll under the in-silico layer; it collects no matter which model or drug wins.

And the thesis isn’t static. On July 20, Tempus agreed to buy Personalis for a $1.5 billion enterprise value, a data-layer company paying up to own an ultrasensitive cancer-monitoring test outright, the measurement itself. That’s this entire argument, happening in real time.

The book isn’t a static list either. It runs as a live, rebased scoreboard so you can watch the whole thesis compound (or not) against the S&P 500 and biotech (XBI) in one view:

→ The AI-Science Basket: live performance tracker

That’s the shape and the scoreboard. Behind the paywall is the actual book: all twenty names with weights and the one-line case for each, the framework that turns “own the bench” into position sizes, and the ranked ways this thesis is wrong. This is the part that doesn’t survive being given away for free.

The Framework: Own the Bench, Not the Drug

To me it means owning the tools and the instrumentation underneath the whole thing, rather than having to pick which individual drug or which AI drug company wins. If the activity explodes, all of those companies need the physical tools either way.

One thing to be clear about before the names, because it changes how you should read all of them. This is a theme I want to play, not a trade I’m timing. I’m not telling you these are cheap this week, and I’m not going to pretend I know which quarter it starts working. What I think I know is which layer the money has to move through as AI pushes into biology, and that is a question measured in years, not prints. So the way to hold it is the way you hold a theme: buy exposure to the layer, size it so a bad twelve months doesn’t force you out of a good decade, and judge it on whether the activity underneath keeps compounding. Individual names in here are expressions of that idea. None of them is the idea.

Everything above resolves to one rule. Do not bet on the company that finds the winning drug, any more than you would have bet the whole AI buildout on picking the winning model. Picking the drug is a lottery. Own the layer every one of these companies must buy from no matter whose molecule wins: the instruments that measure, the data that trains the models, the reagents every experiment burns. So the book weights toward the measurement moat and the recurring picks-and-shovels, owns NVIDIA as the neutral toll, and sizes the drug companies small, the tail where the upside and the binary risk both live.

One thing to name honestly, because the tiers hide it: the measurement core is barely diversification at all, more like a single macro bet on pharma and academic instrument capex, expressed through a few overlapping vendors. Bruker is the partial exception, with a semiconductor-metrology leg riding a different cycle. I’m making that bet on purpose, but size it knowing that, not on the illusion that four instrument names are four independent ideas.

The AI-Science Basket: Twenty Names

A proposed research expression, not a recommendation and not a final book. Every name is liquid and public. The most exciting parts of the stack (Isomorphic, Xaira, Terray, Benchling, TetraScience) are private and can’t be owned here, which is itself part of the argument for owning the tools they all must buy. Weights sum to 100%. The tier split and the three anchors are above; here is the full book, and it runs live in the tracker.

Tier 1 — Measurement & Tools, 40.5% (the moat and the ballast).

• Bruker (BRKR), 6.5% — NMR near-monopoly, timsTOF, spatial, MALDI, the purest expression of the thesis; a second, re-rating semiconductor-metrology leg lets it ride two capex waves on one set of instruments (a BEP deep dive follows). Until Tuesday it was the largest single weight in this book, and it fell about eighteen percent on August 4 after missing on revenue, even though it held its full-year organic growth guide of one to two percent and left its EPS guidance intact. I never took a position in it personally. I watched the stock. Now that it has dropped, the valuation is more compelling to me than it was a week ago, which is a more useful thing to say than pretending the print didn’t happen. The franchise is the same one it was on August 3. The growth rate isn’t, and that’s what got repriced. I’ve cut the weight here from nine to six and a half and moved it to Twist, for reasons in the bear case below.

• Thermo Fisher (TMO), 8% — cryo-EM, Orbitrap, the widest catalog, the default one-stop supplier with a consumables annuity underneath. Durability, not a moonshot.

• Danaher (DHR), 7% — SCIEX, Beckman, Leica, Cytiva, curated breadth run by the sector’s best capital allocators; the Cytiva bioprocess leg is the cyclical part.

• Illumina (ILMN), 5% — the sequencing base, sized below the giants under a three-front attack (Roche, Ultima, Element); the value/turnaround seat, not a compounder.

• Agilent (A), 5% — LC/GC chromatography and the QC plumbing every regulated assay runs through, bought whether or not any drug works.

• 10x Genomics (TXG), 5% — single-cell and spatial data generation, the highest-beta name and most levered to the models’ data hunger; the growth kicker inside the ballast.

• Tecan (TECN.SW), 4% — the only clean public automation pure-play, much of it OEM’d into other platforms as a quiet, sticky razor.

Tier 2 — Compute & Design, 16% (the toll and the design standard).

• NVIDIA (NVDA), 8% — the neutral toll under the in-silico layer (BioNeMo, Clara, Parabricks): it wins no matter which model wins. Already core; here it double-counts as the bio-compute pick.

• Schrödinger (SDGR), 5% — the licensed physics-plus-ML design standard across large pharma, with its own pipeline as optionality; the pipeline burns cash, so the design-layer call, not a sure thing.

• Certara (CERT), 3% — regulatory-grade biosimulation (Simcyp) that lets a simulated trial stand in at the FDA; moat is regulatory embedment, limiter is a smaller market.

Tier 3 — Make & Recurring, 24.5% (synthesis leaders and the razor-and-blade floor).

• Twist Bioscience (TWST), 6.5% — the volume leader in written DNA on a silicon-chip cost advantage, and after August 3 the cleanest evidence in this book for the toll-on-activity argument: revenue up 23% to $118.4M, the DNA synthesis and protein line up 39% with therapeutics inside it up 49%, 369,000 genes shipped in the quarter, full-year guidance raised, and AI-enabled drug discovery orders guided to triple-digit growth for a second straight year. Profitability was the watch item and it’s finally being answered, with adjusted EBITDA breakeven guided for next quarter, though the company is still lossmaking today. I’ve liked this one for a long time and I think it’s still under the radar. If biology becomes programmable, ones and zeros, then the company that writes the DNA isn’t a supplier to that future, it’s a toll on it. This is where the weight I took out of Bruker went.

• Repligen (RGEN), 4% — Protein A ligands and filtration in most antibody purification, a per-gram toll on biologics.

• Sartorius (SRT.DE), 4% — single-use bioprocess consumables, the purest recurring razor; the destock is the risk and, if bottomed, the setup.

• Bio-Techne (TECH), 4% — high-margin specialty proteins and reagents feeding countless assays, plus a growing instruments arm.

• Avantor (AVTR), 3% — the distribution channel on every bench; sticky but lower-margin and levered, the value seat here.

• QIAGEN (QGEN), 3% — sample-to-answer prep upstream of nearly every readout, a steady consumables-kit razor.

Tier 4 — Discoverers & Optionality, 19% (the tenants; a non-binary incumbent anchors it, the pure-plays are the true tail).

• Eli Lilly (LLY), 7% — the AI-armed incumbent whose GLP-1 cash funds the most aggressive pharma AI program in the industry; the least binary way to own the tail, without the go-to-zero risk.

• Natera (NTRA), 4% — the molecular-diagnostics data compounder in the optionality seat: cfDNA testing led by Signatera in oncology minimal-residual-disease (MRD), throwing off a proprietary clinical-genomic dataset that compounds with every test and that models can’t replicate. Closer to the moat than a binary drug bet.

• Tempus AI (TEM), 4% — the multimodal clinical-and-molecular data layer models can’t easily replicate. On July 20 it agreed to acquire Personalis (PSNL) for a $1.5B enterprise value, an all-stock deal folding the ultrasensitive NeXT Personal MRD test into its platform to own cancer-monitoring measurement outright, the data-layer thesis consolidating in real time. The book now owns both sides of the MRD race, Signatera at Natera and NeXT Personal at Tempus, so the theme pays regardless of which test wins; integration and TEM’s own multiple are the watch items.

• AbCellera (ABCL), 4% — an antibody-discovery engine with a royalty model spreading binary risk across many programs; royalties are years out, so tail exposure, not ballast.

Tier weights: Measurement 40.5, Compute and Design 16, Make and Recurring 24.5, Discoverers 19. The shape is deliberate. It’s a book built to be roughly right about the theme even if it’s wrong about any single drug, which is the only honest way to invest in a field this early.

The Bear Case

The models eat the moat. If in-silico prediction of binding, toxicity, and pharmacokinetics gets good enough to skip most wet-lab cycles, value flows back to Design and the measurement core is the wrong 40%. Note the tension: the same capital wall I cite as bullish is partly betting exactly this way. My read is the data-hunger still cuts toward measurement, because better models need more measured data, not less; watch experiments per approved candidate, and if it falls sharply the bench loses.

The buyers insource and cap the pricing. The Friedberg point cuts both ways: the same large buyers who refuse to hand over their data have the scale to bring measurement in-house, standardize on fewer vendors, and lean on price, capping the tool vendors’ pricing power in the exact layer I’m calling the moat. Watch instrument gross margins and vendor concentration in large-pharma tenders.

The real bottleneck is the clinic, not the bench. Even if AI collapses discovery from four years to one, the gauntlet that kills drugs and eats the budget is human clinical trials, and no instrument in this basket shortens those. Optimize the fast part and the slow part still sets the clock; Eroom’s Law is mostly a clinical-failure story, so faster discovery doesn’t, on its own, reverse it. Two things keep me in anyway. The tools are a toll on activity, not success, so more shots on goal means more instruments and reagents burned whether or not any molecule clears Phase 2. And the only durable lever on that 90% failure rate is putting better-validated molecules into humans in the first place, a measurement-and-simulation problem, exactly what Certara, Schrödinger, and better preclinical assays are sold to do. The bench doesn’t bypass the clinic; it decides what reaches it. Underwrite this as “AI fixes drug approvals” and you have the wrong thesis; own it as a toll on discovery volume, which rises either way.

It’s cyclical, and the cycle just bit. Publicly, and in what had been the largest position in this book. On August 4, Bruker fell roughly eighteen percent after reporting. Margins expanded and EPS came in well ahead, but revenue missed, and while management reiterated its one to two percent full-year organic growth guide and its EPS range, it guided third-quarter organic to roughly flat and pushed more of the year into a steeper fourth-quarter ramp. I’m not going to publish a thesis about owning the measurement layer in the same week its purest expression dropped eighteen percent and quietly leave that out. That print is the cyclical bear, arriving before I could publish the note. Instrument demand tracks biopharma and academic R&D budgets, those budgets breathe with the funding cycle, and this book takes that on the chin first.

What the print doesn’t do is change who owns the measurement. Nobody else made a high-field NMR magnet or a cryo-EM column last quarter either, and the reason instrument franchises are worth owning is that the physics takes decades to replicate, not that any single quarter of orders is strong. Those are different claims on different clocks, and the structural one is intact. But owning a moat doesn’t exempt you from a cycle, and a quality compounder can still trade like a disappointment for longer than you expect. So let me put the discipline in writing rather than imply it. I’m not going to hand you a multiple here, because I don’t think that’s the honest way to underwrite an inflection. This is a call about which layer compounds over a decade, not a call that Bruker is cheap at some number this quarter. The thing I’d actually judge it on is instrument book-to-bill and organic growth, and that’s precisely what just came in soft. Scale in, don’t chase, and treat a soft print as an argument about entry and pace rather than about whether to own the layer at all.

But the same week gave the other half of the picture, and it’s the more interesting half. On August 3, Twist Bioscience, which sells the synthetic DNA that physical experiments consume, reported revenue up twenty-three percent to $118.4 million, beating its own guidance of $114 to $115 million, in its fourteenth consecutive quarter of sequential growth. Its DNA synthesis and protein line grew thirty-nine percent, with the therapeutics piece inside it up forty-nine percent. It raised full-year guidance to roughly $456 million, guided to adjusted EBITDA breakeven next quarter, and said orders tied to AI-enabled drug discovery are on track for triple-digit growth for a second consecutive year. The stock rose about eight and a half percent. In the same week Bruker fell eighteen.

The tools layer isn’t one thing, and I had the split slightly wrong. Instruments are capital expenditure, approved by a committee on a budget cycle and deferrable in a lean year. Consumables get burned per experiment and reordered next week whether or not the molecule worked. I’ve argued from the start that these companies collect on the attempt rather than the outcome, and this is the first time I’ve watched both halves of that get tested inside seven days. The activity line held. The capex line broke. So I changed the book rather than just writing it down: Bruker comes down from nine percent to six and a half, Twist goes up from four to six and a half, which takes the measurement tier to 40.5% and the recurring tier to 24.5%. That isn’t a repudiation of the instrument core. It still carries the thesis, and the physics behind those franchises didn’t change on Tuesday. I’d just rather own more of the half that gets reordered. The full ranked bear case is in the primer.

I want to be straight about where that leaves me. Drug development is hard. Clinical trials are brutal, readouts are difficult, and it remains a rich man’s game. AI isn’t going to change that. What it can plausibly do is cut down the number of blind shots and get better-validated targets into humans sooner.

And I know how far that’s from enough. The drug my mother needed had already been discovered. What stopped her was that nobody would make it. So when I say the answer is the physical layer, I mean it more plainly than the framework makes it sound: the instruments that measure, the reagents that get burned, and eventually the ability to manufacture a molecule for the patient in front of you. If AI-driven manufacturing and precision medicine ever get to where a custom drug can be produced efficiently for one individual, that’s the part that would have mattered in my family. We aren’t there.

One more thing, because I don’t want to leave you with more certainty than I have. Nothing has actually happened yet. People are saying this could happen, and I am one of them, but there has not been a single AI-designed drug approved. That is what we are all still waiting for.

What I do think is that this is the trajectory. AI went from chatbots, to making images, to actually doing agentic, loop-based work at scale, to find things that would have taken years, maybe now in weeks or months. That is the unlock. And if it happens, it happens on a bench.

The drugs may or may not come on schedule. The experiments will run either way, more of them every year, on instruments a handful of companies build and reagents a handful of companies sell. Own the bench, not the drug.

On method: I used AI, mostly Claude, to help assemble this piece and the primer underneath it: organizing research, checking reasoning, and arguing with me about the weak parts. The judgments and any errors are mine, and the numbers were verified against primary sources. Worth saying plainly in a note about AI-assisted work.

Disclosure: I’m an angel/private investor in Terray Therapeutics and VistaPath Bio, and serve on the Leadership Board of the Terasaki Institute for Biomedical Innovation. In public markets I’m long NVDA, LITE, CRDO, TSEM, LSCC, ALAB, WOLF, NOW, SMCI, ORCL (2027 LEAPS), and BE. The AI-Science Basket is a proposed research expression and may not reflect current positions. This is research and commentary, not investment or medical advice.

Four Physics Bets, One Constraint

Two weeks ago we sat in front of four founders at Rosewood Sand Hill who had almost nothing in common except the thing they were all trying to get around. The Rosewood is a beautiful hotel, and the room was mostly private-side money and investment bankers. Bernstein and Société Générale had put the founders on a panel called Power & Performance at eleven in the morning, with Stacy Rasgon, who covers the public side of this fight, moderating. Nick Harris of Lightmatter, betting on a photonic interposer. Sid Sheth of d-Matrix, betting on digital compute sitting inside SRAM. Naveen Verma of EnCharge, betting on capacitor-based analog. Taner Ozcelik of Mythic, betting on flash analog after a company reset that could have ended the company instead.

Four incompatible physical bets against the same constraint. We had a question written down for the floor, and time ran out with our hand still up:

“Each of you is a different physics bet against the same constraint, HBM. Reasoning models just made inference capacity-bound, not just bandwidth-bound. Who up here benefits from that shift and who gets hurt?”

It never got asked. They would have split on it, and the direction of that split is the entire investment question.

The session ran under the Chatham House Rule, so we can tell you what the room concluded and not who concluded it. What we took away was that training was conceded to NVIDIA and inference was the battleground. Notable, in a room full of companies that exist to take silicon share from NVIDIA, and it narrows this piece, because a fight nobody is having is not where the money gets made or lost.

Photonics, in-memory digital, analog charge domain and flash analog do not agree about anything at the device level. They agree completely about how you win. Isolate a narrower workload than the incumbent has to support, strip out the hardware that workload does not need, optimize around its actual bottleneck, post a number the incumbent cannot match on that workload, and use the number to get designed in somewhere. It is the classic specialist attack, it is well funded this time, and it is the strategy that lost the last time somebody ran it against this particular company.

3dfx Won Benchmarks. NVIDIA Changed the Product.

The graphics wars are not a story about NVIDIA beating better-funded competitors on frames per second. 3dfx won the benchmarks for about three years, from the first Voodoo in 1996 through Voodoo2, and then stopped winning them: by Voodoo3 it was behind on 32-bit color and large textures. But it did not die of a bad benchmark. It bought STB and turned its board partners into competitors overnight, it missed Rampage, and it kept Glide proprietary while Direct3D and OpenGL became what developers actually wrote to. Every one of those is a decision about the system around the chip. NVIDIA agreed to buy the assets in December 2000. Some of us were gamer kids while that war was running, playing everything on NVIDIA cards, which is part of why the lesson stuck.

What separated NVIDIA was a product family spanning entry through professional, a release cadence nobody else sustained, drivers that worked across a messy installed base, and engineers embedded with game developers before volume availability. Developer support was the moat, and NVIDIA funded it as a moat while most of the field funded it as marketing. Then the definition of the product moved: GeForce 256 pulled transform and lighting onto one processor in 1999, and CUDA turned a graphics component into a general-purpose computer in 2006. A competitor could still build an excellent graphics chip. It was now competing against an installed programming model, a developer base and a roadmap.

Tokens Per Second Is Not an Economic Outcome

There is no honest leaderboard called inference performance, and the reason is that the number has too many free parameters. A published throughput figure moves on model size, weight precision, prompt length, output length, batch size, concurrency, speculative decoding, context window, cache strategy, utilization, whether the measurement is one user or the whole system, and how much model-specific optimization went into the software build. Change any one and the ranking changes with it.

People are playing benchmark games with tokens per second and tokens per watt, and tokens are everywhere now anyway; the open-source wave made them nearly free to try. But run agents making multiple tool calls, watch some of them fail, and the number a CFO opens the spreadsheet wanting is cost per task.

So we are not reprinting anybody’s headline token rate. The challenger field publishes real numbers on real silicon, but they are company-reported, on configurations the company chose, and in several cases they are projections rather than measured deployments. Cerebras publishes token generation speeds on supported models, and those are the company’s own figures. d-Matrix publishes total-cost-of-ownership advantages that are projections or preliminary results. Nobody is lying. But you cannot put two of those numbers on the same axis and have the comparison mean anything.

Frames per second had the same gap and nobody minded, because the number was never the whole purchase. Tokens per second does not tell you whether the model finished the customer’s work correctly.

The Denominator Already Moved, and NVIDIA Moved With It

In July we went through what happened when five model launches in three days all competed on the buyer’s cost per completed task. Databricks had published its own benchmark against a multi-million-line codebase and concluded that token costs are often a poor indicator of overall task costs. Sonnet 5 was roughly 1.7 times cheaper per token than Opus 4.8 and still more expensive per task, because it burned 1.9 times more tokens getting there. The buyers stopped paying for tokens some time ago, and the two numbers do not rank the same.

Then you divide by the pass rate and the spread widens again. As we wrote in Jensen Huang: “Most Companies Will Be Built on Harnesses”: “A cheap task that fails is the most expensive task there is, because a human gets paid to clean it up.” That piece worked the argument at the model layer. Four founders on one stage in July is the same argument arriving at the silicon layer, and it arrives carrying a problem for them, because none of the costs that separate a cheap token from a cheap task are measured at the chip.

We turned that arithmetic into a tool, and it went live today. Octane runs 20 models through the same 32 tasks, grades every attempt the same way, and prices each model per finished task rather than per token. The six models at the top of the board are doing work we cannot tell apart, at a 73x spread in price, and the line on its masthead is this whole section in ten words: price is what you pay for tokens, cost is what you pay for mistakes.

Dynamo separates prompt processing from token generation, routes requests to whichever GPU already holds the relevant context, pins high-value cache, and manages a memory hierarchy that runs from on-chip SRAM out to cloud storage. That is not a faster chip. It is an argument that the unit of the product is the fleet, and the argument has a number attached to it: MLPerf v6.0 in April showed the same GB300 NVL72 rack going from 2,907 to 8,064 tokens per second per GPU on DeepSeek-R1 in six months. Same silicon, same power envelope, software only. An installed base that improves 2.7x without new hardware is a very hard thing to win a socket away from.

The same move showed up in the silicon a year earlier. As we wrote in The Fourth Piece Ships: “The LPU is NVIDIA’s answer to the purpose-built inference chip, and it comes with an ecosystem moat that no startup can replicate.” Whether that ecosystem claim survives contact with a specialist who is materially better is the open question of the next two years. The shape of the response is not in doubt. When a challenger demonstrates an advantage in one layer, NVIDIA expands the stack and makes that layer one optimization inside a system it controls.

The startups need inference to become a separable commodity layer where specialized silicon can win on its own merits. NVIDIA needs inference to be a system problem where hardware, networking, scheduling, cache management and fleet utilization have to be optimized together.

Today Ran the Experiment at Both Ends

Today that fork stopped being hypothetical, twice in one afternoon. SpaceX played its Q2 call live on x.com, and we listened as Musk went all-in on NVIDIA in public. Verbatim: “Going forward, we’ve decided to build exclusively on Nvidia, because we think the Vera Rubin architecture is the best architecture. We think it’s the best AI computer.” SpaceX expects to end this year with over two gigawatts of compute and, in Musk’s framing, to be “closer to 10 GW of compute than 5 GW” by the end of next year, with a stated target of up to 20 gigawatts of power and cooling online ahead of the GPUs to fill it. The word he reached for was computer, and no throughput number appears anywhere in the sentence. A buyer planning gigawatts chose on the architecture. That is the layer, and it is this argument arriving in the customer’s own words. And consider the mouth it arrived from. Elon is regarded as the best hardware operator alive, building the hardest hardware in the world, and he is not trying to make his own AI chip and he is not looking at AMD. When that man says NVIDIA has the best architecture, it is a big, big endorsement, and it landed the same afternoon AMD printed its record quarter. He also told analysts that “our understanding with NVIDIA is that we will receive a very significant percentage of their GPUs next year.” The platform and its largest customer are allocating each other.

Then he took the platform off the planet. Starmind, in his words, is “essentially an optimized Vera Rubin NVL72 computer,” launching next year: a solar-powered rack in sun-synchronous orbit, continuous sunlight for power, vacuum for cooling, results lasered down through Starlink. If the input cost of energy is the sun, it is zero, and energy is most of the bill in this business. It also looks real rather than aspirational; his phrase was “not some sort of far future distant thing.” One more thing we cannot stop noticing: AMD named its rack Helios, the sun, probably with this customer in mind. The customer went and built the sun-powered computer instead. And the same company’s own Starmind page carries the counterweight. It says “We are AI chip vendor agnostic. Our system architecture supports compute modules from any provider,” and it describes a planned in-house AI chip fab with Tesla. Exclusive is a decision that can be unmade, and NVIDIA’s most enthusiastic customer is architected to unmake it.

Below the divider, for paid subscribers: what the AMD print actually repriced and why the earnings multiple is the tell, the two numbers Su was asked to bless and the one she re-anchored, the yield concession and the HBM admission, the ROCm.AI move nobody wrote up, and what SpaceX’s 10% number makes of the other 90%, which is where our most speculative call in months lives. Plus the ranked bear case, led by the one that would make this whole reframe commercially inert.

Read more

You got the headline numbers in the chat during the call. Here is the rest, shorter and blunter than the public post. And my read up front: I know the stock has sold off recently with the whole momentum drawdown, but the story stays intact, and so does the confidence of the CEO on this call. He has been waiting two decades to get to this moment, and they have achieved product-market fit. Or as I put it to Jack Farley three weeks ago: “20 years, no product market fit. And all of a sudden became like the killer app.”

The print in three numbers

Revenue $1,065.4M, up 165.5%, the first quarter over $1B. Operating cash flow $226.4M against $213.1M of cash used a year ago. And the one that matters most for the model: the EPS guide midpoint went from $2.05, the number I flagged for you in April in What Wall Street Missed in Bloom’s Q1 Call, to $2.70. A 32% raise on a guide they had already raised once. They were supposed to be defending $3.4B to $3.8B tonight; they raised to $3.9B to $4.2B instead. The shares jumped after hours.

What the call added

The line of the night: “We think a four-year backlog is not a trophy. It is a confession of constrained supply.” Aimed straight at the turbine incumbents selling 2029 delivery slots as strength. Management also repeated, twice, that the backlog is growing faster than revenue. Treat that one with care: it reads as an implicit answer to the short report, but they never named it, and a faster growing backlog says nothing about whether the backlog is auditable. Claim, not reconciliation.

One thing people need to really understand is that time to power is the whole argument. You can’t wait three or four years for a turbine from GE Vernova; these projects have to get online. And the call tied it to the reason communities fight data centers in the first place: air quality and noise. Bloom answers both. No combustion, negligible air pollution next to turbines and engines, negligible water use, and a server that runs quieter than air conditioning, so communities welcome it and air permits come faster, with management’s line that “every month saved on permits is a month closer to power availability.” I was making exactly this argument on Monetary Matters with Jack Farley three weeks before the call made it: Bloom is “quiet energy for the data center and it ships fast... Time to power is everything,” while the turbine alternative is “noisy and weird. That’s not the data center of the future.”

The most important new fact of the night: customers who had ordered combustion turbines and reciprocating engines cancelled those orders and chose Bloom. Nebius (NBIS) did it this quarter, one of the dozen-plus neoclouds on that roster.

It cuts the other way too. Even the reported New Mexico cancellation: the products get diverted somewhere else, because the products are fungible. Management’s word for it in Q&A was that “every project in the portfolio is fungible.” Mine is that he has built a whole Lego system. Not a big deal.

Track the roster against this letter’s own record. In February the stat that stopped me cold was “six hyperscale and neocloud end customers. A year ago, it was one”. In May, after the Q1 print, I wrote Bloom was “no longer a thesis. It is the emerging industry standard”. Tonight the CEO used the word standard himself, for all the major US hyperscalers plus over a dozen neoclouds. One to six to all of them in twelve months is the demand curve. But the Q&A reading I would give you over dinner: management would not split that roster into operating units versus in-construction versus signed paper, and that refusal is the honest tell about where this actually stands. The rest of the color: “It took us ten years” to become the standard in commercial and industrial, “I never thought within nine months we would become the standard” in AI, “This is not a faster horse. This is a car.” And scandium got exactly three sentences (enough economically recoverable supply, visibility for 25 GW of deployments, not dependent on China) before everything else was declared proprietary. A refusal to disclose is a position. My offtake gate stays open.

The numbers behind the numbers

The full cash bridge and H2 arithmetic are in the public post; the short versions. Cash: the gap between operating income ($182.2M) and operating cash flow ($226.4M) is only $44.2M, so cash tracked earnings rather than working capital games, with the standing caveats (customer deposits, part-owned JVs, and operating cash flow is not free cash flow). Guide shape: the raise implies H2 quarters roughly flat to Q2 at the midpoint and below Q2 at the low end, after a 41.8% sequential Q2. That is the shape of the guide and you should know it before the sell side finds it. Margin: 34.3% non-GAAP, slightly above the top of the guided 32 to 34% non-GAAP band; 33.4% GAAP against 26.7% a year ago. That is the April disclosure, “Ten times the output, same headcount on the floor”, delivering roughly the 200 bps I scoped to it then, not the whole expansion.

The cleanest way I am wrong: the cash swing is not independent evidence. Cash collected in advance against the same disputed backlog is the same claim arriving earlier, and if the 10-Q shows the operating cash flow print leaning on deposits and related-party collections, the wedge I have been leaning on weakens with it.

Targets

In April I told you: “Holding targets pending Q2 backlog composition, any second MSA-style announcement, and the actual cadence of Oracle deployment hitting revenue. Current targets from February: 1-Year (YE 2026) $225+, 5-Year Upside $600+, Downside Risk $80-100. Likely revision direction is up on the one-year given the EPS guide, neutral to up on the five-year given the inference TAM expansion.”

(Targets and promise from What Wall Street Missed in Bloom’s Q1 Call, April 28; the February targets from Bloom Energy Q4 Update: The Backlog Just Doubled, February 6.)

Q2 has now printed.

My targets are unchanged tonight: 1-Year (YE 2026) $225+, 5-Year Upside $600+, Downside Risk $80-100. In April I said the likely revision direction was up on the one-year, and after this print that is still the direction. But I set the condition then and I am keeping it now: the revision waits for the backlog composition in the 10-Q, the RPO reconciliation, and the related-party mix. The print cleared the profitability bar. The filing clears the target revision. What changed tonight is my position, not my targets: I am adding at these levels and holding the numbers until the 10-Q.

For the multiple: at the pre-print $184.89 close, the new $2.70 midpoint puts BE near 68x forward, down from the north of 110x I flagged in April, and higher again after the after-hours move. Cheaper against its own guide, not cheap.

The 10-Q checklist

What actually resolves the open half:

• Related-party share of Q2 revenue

• The customer-deposit line inside operating cash flow

• Capex against that cash flow

• An updated RPO figure against the $20B backlog

• And carried over from April’s watch list: whether a second MSA-scale customer ever gets named. The template is Oracle’s up-to-2.8 GW MSA from April, where the discipline was: “Anyone telling you Bloom just booked $5 billion of revenue is reading the press release wrong.” Same discipline applies to whatever gets named next. Tonight’s call gestured at one without naming it

Gates and position

Two weeks ago, after my call with Hunterbrook’s publisher, I told you “I’m still long, unhedged, and the jury is still out”. The gates I named then are unchanged and still what I am watching: the Jupiter air permit decision, the FERC ruling on the gas lateral, a named non-China scandium offtake, and the backlog to RPO reconciliation. What changed tonight is my posture against them. The discipline then was not adding until a gate cleared. Tonight’s print changes that for me. They are really executing well, they have so many doubters, and this drawdown is a buying opportunity. I am adding to my position. The gates stay open as watch items and the checklist above still decides the backlog question; what changed is my willingness to add before they clear.

One last thing from the call worth keeping. Asked whether AI investment keeps growing at this pace, management said “I do not know for sure and I will not insult you by pretending to,” and reached for Henry Ford, who could not control whether America wanted to drive, only the cost and availability of the Model T. Good color, and I would rather you leave with my version of the same honesty. In December, with the stock around $105, the line was “This isn’t a concept anymore. It’s a deployment”. Eight months later, deployment is exactly the right word for the order book, and still the open question at the AI sites themselves. Tonight moved the flipped default from claim to print, in procurement. The installed base has not caught up yet, and everything I am watching between now and the 10-Q lives in that gap.

Overall, I am excited about the company and the trajectory they are headed, and all the hyperscalers are engaged. The pitch writes itself: rather than keeping your GPUs in storage and facing obsolescence, you can get powered and going.

Disclosure: Long BE, the subject of this piece, as well as NVDA, NOW, LITE, CRDO, TSEM, LSCC, ALAB, WOLF, SMCI, and ORCL (2027 LEAPS). [BEN: confirm current positions before publish; the post announces you are adding to BE] This is investment research, not investment advice. Do your own work.

The Lead That Narrows

Lisa Su put up a chart at Moscone on Wednesday titled “AMD Helios Delivers Industry Leading Rack Scale Performance.” Throughput on the y-axis, interactivity on the x. The advantage over NVIDIA’s Vera Rubin runs +15% at low interactivity, +12% at medium, +10% at high, on Kimi K2 Thinking, a reasoning model AMD picked. I photographed it from my seat and did not think about it again until I was back in the pictures. Nobody around me reacted either. It went by like every other bar chart.

I have held NVIDIA since 2016 and I do not own AMD. Every figure below comes off one of my photographs.

Su had spent the opening arguing that agents are the shift, that 2026 is the first year the world runs more compute on inference than training, and that inference is roughly 60% of AI compute. High interactivity is where that work sits, and where Helios is least separated from NVIDIA’s next rack.

I went looking for the trick in that 15 and found the opposite. Divide Helios by its 72 GPUs and each one is doing about 40 petaflops of FP4. NVIDIA’s own Rubin documentation puts a Rubin package at 35 petaflops of dense NVFP4, and 72 of those is 2.52 exaflops, exactly what AMD’s +15% credits Vera Rubin with. Packages against packages, dense against dense. Moor Insights notes AMD counts MXFP4 where NVIDIA counts NVFP4 and the block scaling differs, but on the question that would have moved the levels, AMD took the conservative NVIDIA number and did the arithmetic straight.

NVIDIA does not market 35. It markets 50 petaflops per package, 3.6 exaflops per rack, and that delta is not silicon. It is an adaptive compression engine in the Transformer Engine that strips zeros out of the data stream in flight, replacing the 2:4 structured sparsity NVIDIA used to double marketing numbers with. Against that figure, Helios at 2.9 exaflops is 19% behind. I will not treat the 50 as realized performance, because nobody outside NVIDIA has benchmarked what that engine delivers on a real workload. Both of these are vendor numbers on unshipped parts. AMD’s +15% is a specification comparison rather than an independent benchmark, and I am holding it to the same standard I just held NVIDIA’s 50 to. Either way, AMD’s lead is a lead in dense silicon, and NVIDIA’s answer to it is not more silicon.

Which is why the curve decays. Both endpoints sit on AMD’s own slide: compute +15%, HBM bandwidth +6%. Low interactivity is batch-heavy work running close to peak throughput, so the lead sits near the flops ratio. As interactivity rises, the workload becomes increasingly sensitive to decode latency and memory movement rather than to peak batch throughput. Weights and KV cache have to be moved for every token generated, and the number that governs that is bandwidth. The line is walking from AMD’s compute number toward AMD’s bandwidth number, and the bandwidth number is six.

Scale-up bandwidth on the same slide reads Same, at 260 TB/s. Nobody said that out loud.

One discrepancy I owe you: my recording has Su saying “50% more compute” at this point, while the slide and Moor Insights both read 15, so I am treating the 50 as an artifact of my own transcription.

The keynote started at 9:30am with open seats around me. In March I stood outside the SAP Center in a line that had been forming for hours. Jensen Huang is the Taylor Swift of AI. That says nothing about the silicon and something about the demand pull AMD is still building. They did copy the DGX Spark, with the Gorgon Halo I am holding above.

The Number I Am Not Swallowing

The $1.4 trillion accelerator TAM does all the work in the bull case, mine included. Last year at this same event Su put it at about $500 billion by 2028. This year it is $1.4 trillion by 2030, so the horizon moved with the number. The server CPU guide went from $60 billion to over $200 billion, against a $25 billion base.

She flagged it herself, in the sentence before the chart went up: “it’s really hard to put this market data up because I feel like every few months we’re changing the perspective based on what our customers are saying.”

Capex has to land as megawatts where the grid will not interconnect new load for years, and Deloitte has 11% of organizations running agents in production against 38% piloting.

The Ramp

AMD earned a win here. In February a SemiAnalysis report said manufacturing problems would push the Helios ramp to Q2 2027, and AMD denied it flatly. Moor Insights put the timing question to Su directly and reports her answer as “Declaring production means we are ready to ship,” with the ramp running through Q4 into the first half of 2027. Whether AMD can build a rack-scale system is settled.

My call is that meaningful volume lands in the second half of 2027, two quarters past AMD’s own window. Helios puts 72 MI455X GPUs and 18 Venice CPUs in one rack, tied together with UALink over Ethernet on Broadcom silicon. Helios is AMD’s first Blackwell-style rack-scale system, but with more external cabling and integration complexity in the part of the stack NVIDIA has spent years learning to manage. Jensen described the Vera Rubin tray in January as having no cables, no hoses and no fans. Six months later Su held up the Helios tray and told the room it weighs more than 160 pounds.

Nick Dorsey, posting publicly on X, reported a conversation with the ODM Wiwynn:

“He said mi455 has lot of challenges right now, that they’re not sure about how it’s ramping or going to ramp. It’s their first rack scale solution and it will have Broadcom scale up networking so AMD is dealing with a lot of complexity […] He said we will see how they come together 2nd half next year”

The same post carries the other half. Customers are optimistic, and “if AMD can make the rack work well then they are going to buy it.” That channel comment supports my second-half 2027 base case, although it is one datapoint rather than confirmation. What would prove me wrong is material Helios revenue in AMD’s Q1 or Q2 2027 results, since a clean Q4 is already guided and would tell me nothing.

The evidence against me is something I stood in front of. Supermicro, which I own, had liquid-cooled AMD racks on its stand with a large team working them, and their product manager, twenty years there, was optimistic. Build quality looked good. Those were CPU racks and Helios is a GPU rack, so what I saw is a channel building and cooling AMD silicon at volume. It counts against me anyway.

The Validators Are Incentivized. The Demand Is Still Real.

OpenAI and Meta each hold a performance-based warrant for up to 160 million AMD shares, roughly 10% of the company apiece, vesting on deployment milestones against 6 GW commitments. Together that is close to a fifth of AMD. AMD’s investment of up to $5 billion in Anthropic was reported July 22, and Tom Brown committed 2 GW the next day.

Be precise about what that does and does not mean. Nobody was paid to read a script. The warrants vest on deployment, so the incentive is to build, and Su says all three are co-developing Helios, which is a deeper commitment than an endorsement. What it does mean is that AMD’s loudest public validation came from parties holding unusually large equity claims on the outcome they were validating. You weigh that differently than a buyer with nothing at stake.

Which is why the most useful thing I heard all day happened in a hallway. A neocloud operator stopped me. His customers had told him to hold on until Helios shipped, and now he wants to scale into the hundreds of megawatts. Demand queued behind availability, from someone with no equity in the outcome.

The Part of AMD I Would Not Bet Against

Two days earlier I was on the other side of this argument. NVIDIA published its Vera whitepaper naming AMD’s flagship, EPYC Turin 9755, beside its own SPECrate score of 925 against 898, vendor-run and estimated. As I wrote in NVIDIA Vera: When CPU Latency Becomes GPU Economics the morning the embargo lifted: “AMD gets the first word almost immediately: its AI event runs July 22 and 23, and NVIDIA just handed every analyst in that room the question to ask.”

Forty-eight hours later Su answered from a stage. Venice runs at 1.8x Turin on 512 threads, and she claimed 20% higher per-core performance than Vera using NVIDIA’s own published data. Each company benchmarked the other’s unshipped part, in public, by name, in the same week.

She also put Venice in a tier she called sandboxes, dedicated CPU capacity for the agent work that is not inference. Every agent step that is not a forward pass is orchestration and tool calls, so the logic holds. But she described the tier without showing what runs there or what it costs, and the $200 billion server CPU number depends on whether anything does. The evidence I could stand in front of was plumbing: nobody liquid-cools a CPU rack, as Supermicro had, unless the thermal load justifies it. AT&T said it runs more than 45 billion tokens a day on AMD across 122 production pipelines. The CPU franchise is the part of AMD I would not bet against.

Where This Read Breaks

AMD is squeezed from two directions, and it can gain merchant GPU share while both jaws close on the economics.

Above it is custom silicon, and people read that threat as hyperscalers walking away from merchant chips. In The Third Announcement Most People Missed at Google Cloud Next I argued the opposite, that “Google is not trying to beat NVIDIA. Google is making Google Cloud the neutral venue where TPU and NVIDIA both monetize enterprise agents.” Two fleets run side by side there, and AMD is in neither. AMD loses the moment the merchant slot is contested rather than expanded.

Below it is a funded field of inference specialists, roughly $58 billion of private paper by Deedy Das’s July 25 tally, with Cerebras public at about $48 billion. AMD’s fast-inference answer is a Cerebras partnership, and Cerebras answers to its own shareholders. NVIDIA licensed Groq’s technology outright and hired Jonathan Ross with it. What would tell me the squeeze is biting is merchant GPU share falling while accelerator spend keeps rising.

AMD is also missing a layer. The stack runs silicon, interconnect, software, and models, and AMD launched the first three this week. NVIDIA ships Nemotron as a full development stack with no revenue or user thresholds, a distribution channel into every enterprise building its own AI.

The strongest case against me is software. In Top 10 Tech Predictions for 2026 I wrote that “AMD’s ROCm remains 3-5 years behind in tooling maturity.” What AMD showed was an attempt to make the coding agents developers already use fluent in ROCm, with demos through Cursor, Claude and Codex. Nobody has to learn AMD’s tooling if the agent already knows it. If that lands, the software gap closes faster than the hardware gap and my read gets worse.

So What?

The market wants a credible second source to NVIDIA, and AMD is finally offering one. The question is no longer whether AMD can ship. It is what you pay for the shipping.

At 10% of its own $1.4 trillion 2030 accelerator TAM, AMD is doing $140 billion a year, against data center revenue of $5.8 billion in the March quarter. That 10% assumes the squeeze does not bite, and it rests on Su’s claim that GPUs stay the vast majority of the market.

Wall Street ran that arithmetic the next day and got there from the other direction. Baird took its target to a street-high $1,250 from $625, modeling $147 billion of AMD AI GPU revenue by 2030 on an assumed 15% share. That is within $7 billion of the figure above, except Baird calls it 15% and I call it 10%. Fifteen percent of $1.4 trillion is $210 billion, so Baird’s share is being applied to something nearer $980 billion. The most bullish note on the street is implicitly applying AMD’s share to an addressable pool roughly 30% below AMD’s headline TAM, most likely because it does not count custom silicon, or categories it does not consider available to merchant GPUs, as reachable. The top jaw is already embedded inside the bull case.

Jeff Pu’s $640 target is explicitly 45 times 2027 earnings. NVIDIA trades in the high teens to low twenties on forward earnings, depending on whose estimate you use. Same demand, same week, and the market pays more than twice the multiple for the smaller and later-arriving one.

The warrants sit outside that multiple entirely. Those 320 million shares strike at a penny, which makes them a customer acquisition cost that lands in the share count where a P/E cannot see it. A forward estimate will not carry them either, because under ASC 260 contingently issuable shares stay out of diluted EPS until the conditions are met, and these vest on milestones nobody has hit.

Run it as illustration rather than as a target. Against 1.65 billion shares, full vesting takes roughly 16% off per-share earnings, which puts that same 45 times nearer $535 than $640. That is arithmetic on somebody else’s published number, not my price target on AMD, and it only happens if AMD ships six gigawatts, which describes a company that has earned the dilution.

What would move me is an independent inference benchmark that tests NVIDIA’s compression claim against AMD’s dense flops.

AMD no longer needs to prove it can compete. Helios is credible, the customers are real, and the market is large enough for AMD to grow dramatically. But credible competition is not the same thing as attractive relative value. AMD is being valued as though the execution, the market share and the TAM all arrive together. NVIDIA is already delivering into the same demand, without promising customers warrants equal to nearly a fifth of its current share count. AMD may win. I still think NVIDIA is the cheaper way to own the outcome.

Disclosure: Long NVDA, NOW, LITE, CRDO, TSEM, LSCC, ALAB, WOLF, SMCI, BE, and ORCL (2027 LEAPS). This is not investment advice.

Coming Up

• The Bernstein Societe Generale conference, and what the inference startups are claiming. The other jaw of the squeeze, first-hand.

• A dinner in San Francisco with AI founders, and what they say about buying compute off-stage.

• Nobody has benchmarked Helios. Every number on that comparison slide is a spec or an estimate, the microbenchmark footnotes cite one to four GPUs, and the only thing that ran live on stage was a small model on a single GPU. I am going through AMD’s endnotes line by line for subscribers.

• Venice is the same question wearing a different hat. AMD called it in production, then footnoted the comparison as preliminary estimates based on engineering projections. The SPEC26 subscores that would settle it have not been published.

• AMD reports Q2 on August 4 after the close, which tests the ramp language. NVIDIA’s FQ2 lands August 26.

Related BEP Research

• The Agentic CPU: Arm’s First Silicon Shot at x86

• NVIDIA Nemotron 3, on the model layer AMD did not launch

• The Memory Wall: Why Groq and Jamba Matter

• The Slide in NVIDIA’s RTX Spark Keynote That Sizes a New Market

• Jensen Huang: Most Companies Will Be Built on Harnesses

“Chiplet architectures are really becoming the legacy CPU architecture... chiplets levy a heavy tax on memory bandwidth and data movement.”

That was the closing statement at the Vera analyst briefing last Thursday, July 16, where I attended live and asked a question. “Chiplet tax” was NVIDIA’s phrase on the closing slide, aimed at the design philosophy under every EPYC server CPU AMD has shipped since 2017 and a growing share of Intel’s Xeon line — filed under legacy, on the record. This morning NVIDIA published the whitepaper that prices the claim: the first NVIDIA CPU document I can find that prints a current competitor’s flagship part number beside NVIDIA’s own score and publishes the full comparison configuration. It lands the day before AMD’s AI event, July 22 and 23, where I will be in the room both days.

NVIDIA sized an entire CPU around how a model behaves between GPU calls.

You Cannot Throw Cores at a Sequential Loop

Cloud economics bought 9x more cores since 2014 and 2x per-core speed, because chiplets made cores cheap. The agent loop only spends the 2x.

The slide that carried the argument: a trace of ten coding agents deployed inside NVIDIA. Green bars mark time with the model on the GPU; gray bars mark time acting on the CPU — tool calls, SQL queries, API calls, scripting. An agent session runs that loop 100 to 300 times, each pass through the gray sequential, the GPU waiting before the next reasoning step. NVIDIA drew the design consequence: “You cannot simply throw more cores at that to make it go faster. You actually need a more performant core.”

One caveat before the argument runs away: the trace does not prove that every gray second is CPU-bound. API calls, storage requests, and database operations can also wait on networks or remote services, and no core shortens a round trip to somebody else’s datacenter. Vera earns its premium only where its core, cache, memory, and fabric actually compress that interval.

Where it does, the metric shifts underneath. Tokens per second prices the green bars only. An agent task is billed end to end, and the most expensive meter in the datacenter keeps running while the CPU works: on the critical path, every second Vera can remove from the gray is a second Rubin no longer spends waiting at GPU rates. The unit of account for the agentic era is cost per completed task — the Denominator Stack from last week’s harness piece: “Sellers converged on the denominator, and the buyers were already measuring in it.” Vera is the silicon floor of that stack, and NVIDIA’s application benchmarks are task metrics rather than token metrics — RL evaluations completed in a window, sandbox jobs finished, and P99 latency on a trading feed. Compress the gray and cost per task falls with tokens per second unchanged.

So NVIDIA built the core. Vera carries 88 Olympus cores, custom-designed by NVIDIA for data-center workloads rather than using the licensed Arm core found in Grace: a wider, faster core tuned for the branchy, pointer-chasing code agents generate, sharing a 164MB system-level cache on one monolithic compute die, fed by SOCAMM2 LPDDR5X at up to 1.2 TB/s per socket. Pressed on manufacturability, Ian Finder, NVIDIA’s head of datacenter CPUs, scoped the monolithic claim with a candor you rarely get from a chip vendor: “We actually are using chiplets. We use them for our memory controller die... we have an IO die.” So “monolithic” means the compute die, and the testable claim is that latency under load stays flat where a chiplet-fabric CPU saturates.

176 vs. 256: The Claim That Matters

AMD EPYC Turin 9755. Two sockets, 256 cores, 512 copies. GCC 15.2, identical optimization flags, Micron DDR5-6400.

That is the whitepaper’s test-configuration fine print. Chip vendors benchmark against unnamed “leading x86” silhouettes as a rule; NVIDIA named the part, listed the compiler, matched the flags down to the jemalloc allocator, and published the copy counts on both sides.

SPECrate 2026_int_base — a widely followed measure of aggregate integer throughput — estimated, dual socket against dual socket: Vera 925, EPYC Turin 9755 898. That is 176 cores outscoring 256 — 3% more score from 31% fewer cores. The caveats ride with the number: vendor-run, estimated on a non-SPEC-compliant reference system, and while the SPEC suite itself is fixed, NVIDIA’s exhibit selection and commentary emphasize the subtests closest to agentic work. Anyone with a Turin box can attempt the AMD side tomorrow, and Phoronix has already tested Vera independently — although NVIDIA disabled power and frequency monitoring and limited the initial workload list to Vera’s target domains. Independent, on NVIDIA’s terms. The posture I set in May in The Last x86 Island holds: “Treat the performance as measured. Treat the efficiency as asserted.”

The 925 matters more than any per-core claim because it attacks x86’s strongest remaining defense. That defense is one I would argue myself: the agent loop is sequential per session, but a datacenter runs thousands of sessions at once, and fleet-level concurrency is a throughput problem, which is what core count buys. The socket number meets it on its own terms — if 176 cores ship more aggregate work than 256, the buyer is paying for 80 extra cores and the power to feed them. NVIDIA now claims all three tiers at once: per-core speed, socket throughput, and memory bandwidth per core.

In NVIDIA’s own loaded-latency testing, Turin falls off a cliff between 350 and 400 GB/s of delivered memory bandwidth while Vera stays flat past 1 TB/s, with a claimed 40% lower peak loaded latency and 12.7 GB/s of memory bandwidth per core against Turin’s 3.1. One caveat NVIDIA does not surface: the two systems ran different memory generations — Vera on LPDDR5X-9600, Turin on DDR5-6400 — and Turin’s cliff sits roughly where a DDR5-6400 subsystem runs out of headroom. NVIDIA chose not to run a same-memory comparison, and AMD has an obvious rebuttal available on Wednesday, July 22: NVIDIA compared two different memory generations and may be attributing too much of the resulting gap to architecture — a memory gap AMD can close far faster than an architecture gap. Caveat priced, the accounting holds — the system-level ledger Rick Xie and I ran in The Bandwidth Tax, now applied to the x86 socket with the line item named.

One more flag before the divider: the most serious long-run bear is not a benchmark dispute — it is software compressing the gray bars themselves. The full argument is below.

Below the divider, the stock implications: the ratio math under NVIDIA’s $200 billion claim, the memory supplier in the fine print, the honest answers on AMD and on Intel’s CPU-demand rally, the ranked winners, five bears, and the watchlist.

Read more

Quick one today, on two positions you know from this letter. In February I told ServiceNow readers to watch Now Assist ACV; the company deleted the metric. Bloom readers got told the backlog had doubled to $20 billion — and a short report just built its case on what that backlog isn’t.

ServiceNow: Two Clocks Into July 22

I published The Orchestration Layer in February with the stock at $107. As I wrote then: “ServiceNow isn’t the victim of the agentic revolution. It’s the infrastructure.” The stock closed Friday at $107.71. Five months, a repackaged catalog, an autonomous agent shipped to GA, three more acquisitions, and the market had moved the stock seventy-one cents. It has ticked up to $111.26 as I send this; the point stands.

Read more

“With an open model like NVIDIA Nemotron, a LangChain harness, the NVIDIA OpenShell runtime, and a company’s own data, every enterprise can build custom agents that understand its business, use its tools, and turn knowledge into action.”

That’s Jensen Huang on Wednesday, launching the NemoClaw Deep Agents Blueprint with LangChain. In the fireside released alongside it, with LangChain’s Harrison Chase, he went further than the product pitch: “most companies are built on business processes. … In the future most companies will be built on harnesses.” That’s his answer for how agents actually reach the enterprise. A fast open model, Nemotron, does the specialist work. A harness runs the loop around it. And the company deploying it owns the stack.

It’s been one of the busiest weeks the AI world has put together this year, and almost none of the coverage connected the events that mattered. So here’s the week read as one story. Pour yourself a coffee; this one runs long.

Huang sells the silicon every piece of that stack runs on, so his pitch arrives with a vendor discount attached. It also happens to be the key to everything else that shipped. Between July 7 and 9, NVIDIA published its case that the CPU, of all things, is now an AI part, and the blueprint moved LangChain’s own agent eval roughly 10x on cost through configuration alone. Cursor and xAI shipped a model trained on a harness’s telemetry, OpenAI shipped a model family that leads its pitch with token efficiency per task, and Meta charged for a model for the first time in its history.

Cover them separately and you get five product stories. I think they’re one. Somewhere between Tuesday and Thursday, AI stopped selling talk and started selling finished work. The unit that matters now is not the token. It is the completed task, and every stage of the stack, from config files to a CPU, just started competing on it.

The Model Talks. The Harness Does the Work.

Strip the jargon and a large language model is a brilliant mind that can only talk. It can’t run code, clone a repository, execute a test suite, or click a button. Everything that turns talk into work comes from the machinery wrapped around the model: the system prompt, the tool definitions, the middleware passing messages, the logic that spawns sub-agents, the sandbox where the work executes. That machinery is the harness.

You’ve probably used one this week without the word for it. Claude Code, Cursor, and Codex CLI are harnesses. Each wraps a model in its own prompts, tools, retry logic, and context management, and sells the package as a coding agent. The model is the engine; the harness decides which tools it reaches for, when it retries a failed test, what context it keeps, and when it stops. Anyone who’s swapped the same model between two of those tools has felt the difference. One setup one-shots the refactor; the other burns twenty minutes calling the wrong tools on identical weights. Adel El Hallak, NVIDIA’s VP of Product for AI, posted the cleanest definition of the week on LinkedIn Thursday morning, “An agent is more than a model. It is a model plus a harness,” and named the practice of tuning the scaffolding instead of the weights. “We call that harness engineering.”

Part of what the market has been pricing as model capability is harness capability. Tokens are an input price. A completed task consumes token bursts, but also tool calls, repository clones, test runs, sandbox time, failed branches, and retries, and the harness decides how much of that is waste. Dollars per completed task prices the whole loop, retries included.

This is what I meant in The Compute Confession when I argued that a frontier price sheet is a leaked bill of materials. A token price leaks the hardware bill underneath the model; a task price leaks the model, the harness, and every wasted call between them.

Hardly anything referees this in public. There’s no fixed task set and no agreed success threshold, so every vendor scores itself.

Every Launch This Week Was Selling Tasks

Start with the incumbent. OpenAI opened GPT-5.6 Sol, Terra, and Luna on Thursday at unchanged or lower token prices; Sol launched at GPT-5.5’s exact sheet, $5 in and $30 out per million tokens. The pitch went to the new unit instead, with Sol billed as 54% more token efficient on agentic coding tasks. Altman told CNBC: “Every enterprise now is thinking about spend and the value they’re getting in exchange for AI, and this is what we really want to do.”

Meta ran the same play from the price end. Muse Spark 1.1 is the first model Meta has ever charged for, and per Bloomberg’s coverage it lands at $1.25 in and $4.25 out per million tokens with a 1M-token context window, 4x below GPT-5.5 on input. Coverage of Meta’s eval tables, vendor numbers all, shows the aim. It sits near the top on tool use (JobBench 54.7 against 48.4 for Opus 4.8; MCP Atlas 88.1) and a step back on one-shot coding (SWE-Bench Pro 61.5 against Opus’s 69.2). That price sheet is built for task volume, not benchmark trophies.

And Grok 4.5 only makes sense in the task lens. Cursor and xAI announced it as the first jointly trained model between a frontier lab and a harness company. (The lab’s own launch page styles the entity “SpaceXAI”; I’m writing xAI until the consolidation is confirmed.) The training corpus is the tell: trillions of tokens of Cursor telemetry, priced at $2 in and $6 out per launch coverage. A tuned profile bends the harness to the model. Grok 4.5 bends the model to the harness. Either direction, the thing being optimized is finished work per dollar.

The other two launches were infrastructure, and they bookend the story. LangChain and NVIDIA’s blueprint shipped a per-model harness profile as pure config, no fine-tuning, carrying a 10x cost claim on their own eval that I take apart below. In the fireside, Harrison Chase names that comparator: Claude Opus, at 87 to Nemotron's 86, ten times the cost.And NVIDIA’s Vera CPU posts landed July 7, two days before any of the model news, pitched in NVIDIA’s own press language as a shift “from cores per dollar to tokens per dollar.” A CPU launch marketed in AI-economics terms is itself a tell, and the engineering behind it is the strongest confirmation of the whole thesis.

That’s the free half: the vocabulary and the news. Behind the paywall is the part that makes it investable. The mechanical proof of the Denominator Stack, including the exact figures on the config file that moved an agent eval an order of magnitude with the weights untouched, and the enterprise that measured the same lever on its own multi-million-line codebase. The only third-party board pricing coding agents in dollars per task, what its same-day rerun of GPT-5.6 did to the open-versus-closed narrative, and why the loser of that reading isn’t who you’d guess. The purpose-built-silicon read on NVIDIA’s Vera CPU. The bounded NVDA claim, and the one company that may end up selling the loop to everyone who can’t build it. Four ranked risks, including the one aimed at my own framing, and the watch items for the next two weeks.

The Referee Ratified the New Metric Same-Day

Every model launch markets efficiency, and two things separate this one. The token price stayed put while the whole pitch moved to relative task cost, and the pitch got checked within hours by a board OpenAI doesn’t run. DeepSWE, a third-party benchmark from Datacurve, is built on 113 long-horizon coding tasks written from scratch across 91 real repositories, with hand-written verifiers, every model run on the same frozen harness so the models are comparable. Late Thursday it re-ran its leaderboard with the new family.

Sol tops the board at 73% pass@1 for $8.39 a task, against Claude Fable 5’s 70% at $21.63. Terra matches Fable 5’s pass rate at $4.95, and Luna matches GPT-5.5’s 67% at $3.03 versus $7.23. OpenAI’s launch copy claimed a sixteenth of Fable 5’s cost for its cheap tiers; the board measured roughly a quarter at matching accuracy, deflated but directionally intact. The rows also cut against the week’s easiest hot take, because the closed incumbent now tops the buyer’s metric among the cheapest per success. Whatever repriced this week, it wasn’t open models eating the closed frontier.

The buyers are measuring too. On July 8, Databricks published an internal benchmark built from real merged PRs against its multi-million-line codebase and concluded that “token costs are often a poor indicator of overall task costs.” The cleanest cell is Sonnet 5, about 1.7x cheaper per token than Opus 4.8 and still more expensive per task, $2.09 against $1.94 at a lower completion rate, because it consumed 1.9x more tokens getting there. Cheap tokens bought expensive work.

The Harness Went on Sale, and Open Source Got Hired

The quietest launch of the week carries the biggest repricing. On July 8, LangChain and NVIDIA shipped the NemoClaw Deep Agents Blueprint, and inside it a harness profile for Nemotron 3 Ultra, a config layer that adapts the scaffolding to one model’s quirks. No fine-tuning, no serving changes. On LangChain’s own agent eval suite, the tuned profile posts a 0.86 aggregate score at $4.48 per run; the next closest performing model cost $43.48, “roughly 10x lower inference cost” in LangChain’s phrase. Vendor eval, vendor suite, three parties with product to sell. Discount all of it and the mechanism stands. Config alone moved the task number an order of magnitude, weights untouched, which means part of the closed-model premium was scaffolding all along, and scaffolding now ships as a file.

Databricks corroborated the lever from neutral ground the same day, running the same model at the same thinking effort through two harnesses, Claude Code and Codex versus the lighter Pi. Cost per task differed by “more than 2x in some cases” while “quality remained the same,” mostly because Pi sent about 3x less context per turn. Harness choice rather than a tuned profile, but the same lever, measured by a buyer. DeepSWE can’t see any of this, since it freezes the harness by design; the harness has no public referee yet, though buyers have started grading it privately; Databricks published its methodology the same week.

Who runs which model inside the loop is settling too. In production, the tiers compose, a frontier model orchestrating with open models executing underneath. NVIDIA says its own research agent won Deep Research Bench with Opus planning over Nemotron sub-agents, Coinbase reported cutting AI spend nearly in half by routing execution to open-weight defaults, and Databricks’ board landed GLM 5.2 statistically tied with Opus 4.8 on quality at $1.28 a task against $1.94. That composition is the argument of The Open Source War Is Over, published Tuesday on X. Neither side won; both got hired into the same loop, and the seam where the hiring decision gets made is the harness router. El Hallak said it on the BEP Research podcast back in March: “All agents will be a system of models.”

The nation-scale version arrived the same week. Palantir is taking NVIDIA’s open models to frontier grade for the US government, agencies keeping the weights, and Karp put the buyer’s frame plainly on CNBC: “I wanna own the GPUs. I wanna own my data. I wanna own the model.” The All-In panel’s read of the same deal was that most foreign nations will run open source on NVIDIA hardware. Sovereign work is picking the open executor tier, and it mints on NVIDIA silicon either way.

Vera Is the Part NVIDIA Built for the Loop

Now look at where the loop physically runs. The doing in an agent task, the tool calls, repo clones, test suites, and sandbox spin-ups, is CPU work; the GPU fires in bursts between those steps. When the industry priced itself in dollars per million tokens, the GPU carried the whole story. Price by completed task and a large share of the work never touches the GPU at all.

This is the mechanical constraint I laid out in The Agentic CPU in March: “Agentic AI doesn’t reduce CPU demand. It explodes it.” Vera is NVIDIA acting on that arithmetic in silicon, three and a half months later.

The engineering reads like a spec sheet for loop stitching. NVIDIA’s claims run to 88 in-house Olympus cores on a monolithic compute die, 1.8x sustained per-core performance under full socket load, and more than 3x per-core memory bandwidth at less than half the power of a traditional x86 datacenter CPU; single-threaded speed and sandbox concurrency are exactly what the stitching starves for. Perplexity, the one named third party in NVIDIA’s own post, cloned a repository and ran its test suite in sandboxes about 1.5x faster than x86, and started concurrent sandboxes up to 1.9x faster. And NVIDIA’s own press release re-denominates the business in as many words, “AI factory economics are shifting from cores per dollar to tokens per dollar.”

The dates say NVIDIA saw this coming. Nemotron, the free open executor model, shipped in December. In March a product VP sat on my microphone describing agents as systems of models, and the CPU the loop runs on arrived in July with agents on the label. A company reacting to a metric change doesn’t have the part ready the same week; NVIDIA had already built for it.

Call it the Denominator Stack: five sellers at five stages, config to silicon, all competing on the buyer’s task cost inside three days, each with its own eval and its own units. The hard cells across the board are the price sheets and the DeepSWE rows; the rest is vendors grading their own homework.

The equity read is narrower than the excitement. NVIDIA doesn’t bill Grok 4.5, GPT-5.6, or Muse Spark, and all three can be served off non-NVIDIA silicon. What the week supports is the plumbing claim: NVIDIA sells a lever at more stages of the agent xloop than anyone else, the GPU for the token bursts, Vera for the stitching, the harness profile for config, the OpenShell runtime for the sandbox. Who bills the models running through the loop is a different question, and mostly the answer is someone else. I’ve held NVDA since 2016 and I’m not adjusting anything on this week’s news; what changed is the shape of the claim I’m willing to make for it.

The forward question is who sells the loop to everyone who can’t build it. DoorDash, Coinbase, and Databricks are the sophisticated tail; they staffed their own routers and built their own benchmarks. Most enterprises never will. If the harness decides task cost and the composition decides the bill, the other 95 percent buy the loop as a product, governance included, and ServiceNow is positioned as exactly that layer, “the layer that every model depends on but no model can replace” (The Orchestration Layer). Maybe the cleanest way to play the loop is the company selling it, not the companies inside it. That’s a question I’m taking into the paid follow-up, not a call I’m making today, and the same bound applies: ServiceNow bills none of the models either.

Where This Breaks

The stages cannibalize each other. A 54% token cut, a 4x model price cut, and a 10x config gain don’t multiply into a compound saving. They compete for the same savings line in the same enterprise budget, and a buyer who captures one may stop shopping for the rest.

The denominator has no general referee, and the two most dramatic multiples of the week, the config gain and the silicon gain, are publicly unauditable. A metric this loose invites gaming, and vendors game loose metrics every time.

Value capture is the soft spot. The assumption that the unlocked loop volume lands on NVIDIA iron is the load-bearing one, and it’s contestable. Muse Spark runs on Meta’s own buildout, and TPU, Trainium, and Groq serve agent loops today. The sovereign builds cut both ways, open volume weakening closed-lab capture while minting on NVIDIA silicon, which refines the risk without answering it. If the volume pools on someone else’s silicon, the loop thesis survives but the ticker translation doesn’t.

There’s a fourth risk, and it’s mine: curation. Token-efficiency claims, price cuts, and CPU launches happen constantly, and the line through these five is my drawing. Two of the five price sheets are still written in tokens; the translation into the buyer’s unit is this piece’s work, not the sellers’ announcement.

Underneath all four sits the capability floor, because the denominator is completed tasks. Divide each DeepSWE cost by its pass rate: Sol’s $8.39 becomes $11.49 per successful task, up 37%; Terra’s $4.95 becomes $7.07, up 43%; Opus 4.8’s $13.22 becomes $22.41, up 70%; Sonnet 5’s $26.40 becomes $48.89, up 85%. Failure inflates every bill, and it inflates the low pass rates most. A cheap task that fails is the most expensive task there is, because a human gets paid to clean it up.

The first two risks change the slope of the story. Value capture can change its direction. What flips me: loop volume visibly pooling on non-NVIDIA silicon, or the harness lever failing to reproduce. One profile is a demo; three at the same quality is a production line, and LangChain has shipped one.

So What?

Five takeaways, and one chain to keep. The harness drives the agent, the CPU carries the harness’s work, and the completed task is what the buyer pays for. Every link in that chain got a launch this week, and NVIDIA sells a lever at more of those links than anyone else.

1. AI started selling finished work. Five launches in three days all competed on the buyer’s cost per completed task, a third-party board repriced the frontier in that unit within hours, and an enterprise published its own task-cost benchmark the same week. Sellers converged on the denominator, and the buyers were already measuring in it. The general referee seat is still open, and whoever fills it inherits pricing power.

2. The harness is the lever, and it just became a product. LangChain turned harness capability into a config file worth 10x on its own eval; Databricks measured 2x from harness choice on its own codebase. The moat question in agents moves from who has the best weights to who owns the loop around them.

3. The open source war ended in a hiring decision. Production composes the tiers, frontier orchestrating, open executing, and the sovereign version of the same pattern is forming at nation scale. The fight that matters now is over the router that makes the hiring decision.

4. Vera is what it looks like when the silicon vendor believes the thesis. The doing in an agent task is CPU work, NVIDIA built a CPU for it, and its own press language now reads “cores per dollar to tokens per dollar.” NVDA lands inside the plumbing claim only; I hold the position and I am not extending the claim past the plumbing. Of the four risks above, value capture is the one I’d underwrite first.

5. The forward question is who sells the loop. The sophisticated tail builds its own routers and benchmarks; the other 95 percent will buy the loop as a product, and ServiceNow is positioned as that layer. Watch three things: the next two LangChain profiles (does the config lever reproduce), a Grok or Nemotron row on DeepSWE (the neutral-ground test), and the neocloud disclosures (where the loop volume actually pools).

I don’t know which stage wins the savings claim in an actual enterprise negotiation, and I’m skeptical of anyone who claims to this week. The sellers have already picked their unit.

Price the loop, not the model.

Coming Up

• Ill be at AMD Advancing AI 2026 later this month in San Francisco

Related BEP Research

• The Reasoning Tax: Why GPT-5.4 Just Validated the Memory Wars

• AI for the Rest of the World: Why Muse Spark Changes the Meta Thesis

• The Orchestration Layer Is the New Platform War: NVIDIA’s AI Agent Strategy (GTC 2026 podcast)

• NVIDIA Nemotron 3: Why This Changes Enterprise AI Economics

• The Last x86 Island: The One Socket NVIDIA Didn’t Own

Resources

• LangChain: NemoClaw Deep Agents Blueprint launch post (Jul 8, 2026; figures verified against this post)

• Jensen Huang fireside with Harrison Chase, LangChain (Jul 8, 2026; “built on harnesses” quote verified against the video)

• NVIDIA Developer: Vera CPU boosts AI factory throughput for agentic workloads (Jul 7, 2026)

• NVIDIA: Vera, the max single-threaded CPU at scale (Ian Buck, Jul 7, 2026; Perplexity figures)

• NVIDIA Newsroom: Vera, the CPU for agents (”cores per dollar to tokens per dollar”)

• DeepSWE leaderboard v1.1 by Datacurve (113 tasks, 91 repos; updated Jul 9, 2026, read same day)

• Databricks: Benchmarking Coding Agents on Databricks’ Multi-Million Line Codebase (Jul 8, 2026; read same day)

• CNBC: Altman on GPT-5.6 Sol and enterprise spend (Jul 9, 2026)

• OpenAI: GPT-5.6 launch page

• Bloomberg: Meta starts charging for AI with Muse Spark 1.1 (Jul 9, 2026; pricing source)

• xAI: Grok 4.5 launch post (page styles the entity SpaceXAI; pricing per launch coverage)

• Cursor engineering blog: Grok 4.5

• Ben Pouladian: The Open Source War Is Over (X article, Jul 7, 2026)

Disclosure: Long NVDA, NOW, LITE, CRDO, TSEM, LSCC, ALAB, WOLF, SMCI, BE, and ORCL (2027 LEAPS). This is not investment advice.

Co-authored with Rui “Rick” Xie, researcher in computer systems, memory architecture, and AI infrastructure (rickxie.cn). Rick led the engineering and standards analysis; Ben Pouladian led the investment framing.

published “HBM: High-Bandwidth Mistake” on May 31, and it has been making the rounds ever since. The spicy takes, verbatim:

“I believe HBM is a mistake and will be phased out (90% drop in volume from peak) within the next 7-10 years.”

“DRAM stocks prob have another double or triple left in them.”

“At some point in the next 3-10 years, DRAM stocks will drawdown at least 70% from peak.”

The same week those lines were making the rounds, SK Hynix was booking the other side of the argument: its roughly $28 billion Nasdaq ADR offering, the second-largest US listing by a foreign company since Alibaba, came more than seven times oversubscribed ahead of Friday’s debut under SKHY. The loudest memory bear note of the summer and the most heavily bid memory listing in history landed in the same tape.

The volume-collapse call is the loud half of that note, and it is the half we think least matters to a portfolio. The quiet half is the one that reprices the memory complex: long before anyone phases HBM out, the money being made around each HBM stack starts growing faster than the money made inside it. We went looking for that line in the standards documents and the vendor disclosures. This piece is what we found, including the place where the data corrected our own first draft.

Everyone Is Long Memory on Scarcity. That Call Looks Right, for Now.

Start with what the bear note gets right, because it gets a lot right. Every major HBM vendor is effectively sold out. NVIDIA’s HBM4 bandwidth demand makes qualified stacks the binding constraint into 2027, and greenfield DRAM supply does not arrive in size before 2027-2028. Scarcity keeps pricing power with Micron, SK Hynix, and Samsung. Even the bear gives the equities “another double or triple” before the reckoning. We are not shorting memory into a sold-out tape, and neither is the author of the bear note.

But scarcity and value capture are different questions. One asks who is sold out today. The other asks who collects the incremental dollar as the system gets re-architected around the constraint. Those answers are already diverging, and the divergence is visible in public documents that most of the people sharing the bear note have not read.

Rick read them.

What the Standards Actually Say

Our working draft of this piece assumed what most bandwidth-tax arguments assume: that each HBM generation delivers bandwidth at worsening cost and power per bit, and that the curve inside the stack is bending the wrong way. Rick went to the JEDEC standards and the vendor disclosures to anchor that chart, and came back with the correction that reshaped the piece: the public data do not show a worsening energy or cost curve per unit of delivered bandwidth. The stack is improving.

The numbers, from the vendors’ own disclosures:

• Micron’s HBM3E delivers more than 1.2 TB/s per stack at pin speeds above 9.2 Gbps.

• Micron’s HBM4 delivers more than 2.8 TB/s at above 11 Gbps: 2.3x the bandwidth of its HBM3E with over 20% better energy efficiency.

• Samsung’s HBM4 reaches up to 3.3 TB/s at up to 13 Gbps, with reported 40% better power efficiency, 10% lower thermal resistance, and 30% better heat dissipation versus its HBM3E.

• Samsung’s HBM4E samples run a stable 14 Gbps, scalable to 16, at up to 3.6 TB/s, with 16% better energy efficiency than HBM4.

A note on measurement boundaries: published pJ/bit figures often measure different parts of the memory path. Samsung’s 2023 HBM3 interface work reported 0.29 pJ/bit, while Intel’s 2025 HBM3 host-PHY paper reported 0.5 pJ/bit. Neither figure represents complete stack-access energy, and the two should not be combined into a cross-generation product series. The publicly available peer-reviewed data do not support a clean, apples-to-apples HBM3E-to-HBM4 stack-level pJ/bit chart. We wanted to build that chart. We couldn’t, so we didn’t.

So where is the tax? It moved. The bandwidth tax is levied at the system level: everything the platform must pay around the stack to turn a faster cube into delivered bandwidth. HBM4 doubles the channel count from 16 to 32 and doubles the interface width. Feeding that interface pulls the base of the stack onto advanced logic: Samsung builds its HBM4 base die on a 4nm logic node, and TSMC offers N12 and N3 processes for HBM4 base dies. JEDEC raised the nominal stack height from 720 to 775 microns and formally blessed 16-high configurations, which is more vertical room and also more layers, more parasitics, and more heat to manage within the defined HBM package envelope. TrendForce expects the manufacturing complexity to support a price premium above 30% for HBM4, driven by the doubled I/O count, larger die area, and the logic base die. Package area, power delivery, cooling, yield, qualification, and constrained packaging capacity all scale up alongside. That bill lands on the system, and it grows every generation even while the stack itself gets more efficient.

Hybrid bonding fits the same pattern, and this is Rick’s framing worth keeping: hybrid bonding is not a cost cure, it is an extension whose value rises as more layers, more channels, denser interconnects, and tighter thermal budgets have to fit inside that 775-micron envelope. Samsung reports hybrid copper bonding supports 16 or more layers while cutting thermal resistance by more than 20% versus thermal compression bonding. The instinct is familiar to BEP readers. As I wrote in The Memory Wars: “NVIDIA Feynman (2028) won’t fight SRAM physics. It routes around them: 3D-stacked SRAM using AMD X3D-style hybrid bonding.” Memory vendors are now doing to HBM what NVIDIA is doing to SRAM: routing around a wall rather than pushing through it, and paying for the route in packaging complexity.

The Migration Map

The claim we will defend is sharper than the one we started with. The consensus bull case says HBM stays scarce and the oligopoly keeps the margin. The circulating bear case says HBM gets phased out and the DRAM equities eventually round-trip. Rick’s read of the data supports a third path: the profit pool broadens before HBM volume or pricing ever peaks. HBM can remain scarce and highly profitable while foundries, base-die logic providers, packaging suppliers, and system integrators capture a rising share of each incremental bandwidth dollar. “Peak HBM” is in here somewhere as a pricing-power moment, but the broadening comes first.

There is already a first signature of it in Micron’s disclosed numbers. As I wrote in Micron Just Proved the Memory Thesis: “Sanjay confirmed that non-HBM margins are currently higher than HBM margins. That inverts the popular narrative that HBM is where all the pricing power lives.” One quarter of relative-margin commentary is a data point, not proof of a structural relocation. But when a top-three HBM vendor tells you its richest margins sit elsewhere in its stack, that is the tape’s first reading in the direction of the broadening.

Not every candidate destination is real. Rick placed each layer; the monetization read is mine.

Three rows deserve words the table cannot hold. The packaging row is the argument from The Packaging Paradox: “the real bottleneck isn’t transistor density anymore. It’s advanced packaging.” The 30%-plus HBM4 manufacturing premium is that bottleneck showing up in a memory price. And the CXL row got materially stronger on June 29, when Meta disclosed Vistara, a custom CXL memory-expansion ASIC that reuses DDR4 from decommissioned servers as a slower capacity tier. Meta reports deployment across millions of production servers and up to a 25% reduction in server count for disaggregated ML inference. That is a hyperscaler saving real money with tiered memory in production, and it stays inside CXL’s actual lane: none of that DDR4 is doing HBM’s job.

The third row is the newest. SanDisk and SK Hynix began standardizing HBF this February: NAND stacked in an HBM4-matching footprint and power envelope, roughly 1.6 TB/s of read bandwidth per stack at 8 to 16 times the capacity of comparable HBM, with samples targeted for this half and the first inference devices in 2027. It is a capacity tier, not an HBM substitute, and demand forecasts cluster around 2030. The reason it earns a row now is who is building it: a flash vendor buying into the bandwidth profit pool is the broadening thesis in its purest form.

Rick’s one-line summary of the map is the line to keep: the investable question is where the next unit of value accrues while qualified DRAM dies and stacks remain scarce.

The Base Die Is the Cleanest Signal

Both JEDEC standards allow an optional logic die at the bottom of the stack for signal redistribution and vendor-defined functions. For most of HBM’s life that die was the memory vendor’s own, on a DRAM-adjacent process, and nobody outside the packaging team thought about it. That is over. Samsung’s HBM4 base die is built on its 4nm logic node. TSMC offers N12 and N3 base-die processes to the industry. Micron plans both standard and customized HBM4E base logic dies fabbed at TSMC, and has told investors it expects the customized products to carry higher gross margins.

Read that last disclosure the way we do: a memory vendor telling you that the customized logic at the bottom of the cube is where its own margin expansion lives. The margin is migrating inside the incumbent’s own product line, from the DRAM cell it has always sold to the logic and integration it is learning to sell.

Rick’s discipline on this point is worth quoting directly, because it is the part the custom-HBM hype skips: the base die does not add supply. It does not fix the shortage of DRAM core dies, known-good-die yield, stacking capacity, or packaging throughput. What it does is raise the system value of each scarce stack: interface optimization, offloaded host-die overhead, support logic pulled into the cube. Marvell claims its custom HBM architecture can cut memory-interface power by up to 70%, free up to 25% of compute-die area, and support 33% more memory capacity. Those are vendor architecture targets, not industry-measured results, and we label them as such. But the direction of every one of those numbers points the same way: the value of a stack increasingly depends on logic and packaging work done by someone other than a DRAM fab.

The Optical Off-Ramp Is Real. It Is Also Not Next.

The bear note’s proposed endgame is memory disaggregation over co-packaged optics: “HBM is a mistake. The solution is to disaggregate memory with a real PHY that drives co-packaged optics.” And a widely shared PhotonCap piece argued that the Micron-Anthropic HBM agreement accelerates that future, right in the headline: “The More Anthropic Buys Micron HBM, the Faster Optical Memory Pooling Arrives.”

Keep the claim narrow, because the sources themselves do. ’s own body text concedes the deal “does not pull optical revenue forward so much as it strengthens the case for why optics becomes necessary,” and keeps optical memory revenue in FY2028 and beyond, with electrical CXL pooling entering revenue first. That matches Rick’s sequencing exactly, so we will state it as the joint view: CXL-based capacity pooling near term, optical scale-up interconnect next, optical memory pooling later. Optics improves reach and bandwidth density. A deployable memory system also needs end-to-end latency, coherence, data placement, reliability, yield, and customer qualification, and none of those ship on a press release. Marvell identifies all-optical scale-up interconnect as Celestial AI’s first commercial application. Optical memory pooling is also part of the platform vision, but its deployment timing is less clearly defined.

The same lesson is arriving from the latency side. A widely shared thread from the anonymous chip account Big Boss walked the interconnect math on sharded inference this week: hundreds of nanoseconds of SerDes and switch overhead, then tens of microseconds of software, wrapped around payloads measured in kilobytes. Whatever you make of his conclusions, the overhead math cuts our way: remote memory is only as useful as the fabric and placement layer serving it.

HBM scarcity strengthens the case for disaggregation. It does not, by itself, move the optical timeline. Both halves of that sentence matter, and most takes this month kept only one.

Three Ways We Are Wrong

1. Scarcity outruns the migration. If qualified HBM supply stays tight through 2028, the oligopoly’s pricing power persists and the broadening stays small relative to the cell’s profits. This is the strongest bear on our timing. Note what it does not touch: the direction. Even in this world, base-die value and the packaging premium grow; they just grow in the incumbents’ shadow for longer.

2. The destinations slip. CXL beyond the hyperscaler in-house case, SOCAMM volume, and custom base-die design-ins all carry integration friction. Meta building Vistara for itself is not the same thing as a merchant CXL market. If the “early” rows stay early, the incremental dollar has nowhere to migrate and defaults back to the cell.

3. The incumbents capture the migration internally. Micron fabbing custom base dies at TSMC and keeping the margin uplift is exactly this scenario, and Micron itself is telling you it plans to do it. If SK Hynix, Samsung, and Micron own the base die, the packaging relationships, and the customization revenue, the profit pool broadens but the tickers that own it barely change. Hold onto that nuance: a reader can accept our entire thesis and still be correctly long Micron. The difference is what you are long: the customization and integration line, not the commodity cell cycle.

What would change our mind. We are wrong on direction, not just timing, if two things happen together: the HBM base die stays overwhelmingly memory-vendor internal logic with no foundry co-design design-ins by late 2027, and Micron’s non-HBM margin leadership reverses for two or more consecutive prints. That is the clean disconfirming test, and we will run it in public.

So What?

Stop arguing about whether HBM disappears. On the published data, HBM is getting better per stack, per watt, and per delivered terabyte-per-second, and it is sold out. The question that reprices the complex is different: the system bill around each stack is growing every generation, and a growing share of that bill can accrue beyond the DRAM fabs. Foundries offering advanced-node base-die processes, including TSMC’s N12 and N3 options. HBM4 carrying an expected 30%-plus premium associated with higher manufacturing complexity. A hyperscaler cutting up to a quarter of its inference servers with a memory-tiering ASIC it designed itself.

For a decade, long memory meant long the DRAM cell. The bandwidth tax says the durable version of the trade is one level up: long the constraint, which is qualified bandwidth delivered into a rack, and long whoever gets paid to relieve it. Sometimes that is still the memory vendor; Micron’s custom base-die margins are the incumbents’ counterattack. Sometimes it is the foundry, the packager, or the CXL controller. The tells are concrete and they are all watchable this year.

And one ungated takeaway, so the free read carries a position and not just a concept: the layer already collecting the tax is advanced packaging, where TSMC and the packaging suppliers are pricing scarcity today, and the layer to underwrite next is the base die, where the foundry-plus-custom-silicon axis meets the memory vendors’ customization push. The under-covered names at both layers are the paid piece. The layers themselves you now have for free.

This reframe is free and citable on purpose; the thesis is the marketing. The paid follow-up, “The Bandwidth Tax: The Migration Call,” carries the trade: the entry framework for the pricing-power question, the under-covered base-die and OSAT names positioned for the broadening, position sizing, and the full watchlist with tickers. Rick and I will also do a joint Q&A in the Substack chat the day it drops, walking the engineering feasibility and the monetization read side by side.

Long the constraint, not the cell.

Sources

• Irrational Analysis, HBM: High-Bandwidth Mistake (May 31, 2026)

• Bloomberg on the SK Hynix US ADR offering (7x+ oversubscribed) (July 8, 2026)

• Micron HBM3E product page; Micron HBM4 volume production release

• Samsung commercial HBM4; Samsung HBM4E samples; Samsung hybrid copper bonding (GTC 2026)

• JEDEC JESD238B.01 (HBM3) and JESD270-4 (HBM4), stack height and channel-count specifications

• TSMC North America Technology Symposium release — N12 and N3 logic base die for HBM4

• TrendForce on HBM4 manufacturing complexity and premium

• K. Chae et al., ISSCC 2023 (HBM3 interface, 0.29 pJ/bit); A. H. Mehta et al., VLSID 2025 (HBM3 host PHY, 0.5 pJ/bit) — interface/PHY boundary measurements

• The Register on Meta’s Vistara CXL ASIC (June 29, 2026)

• PhotonCap, The More Anthropic Buys Micron HBM, the Faster Optical Memory Pooling Arrives

• Big Boss (@0xBADB01E), thread on interconnect overhead in sharded inference (July 5, 2026)

• SanDisk and SK Hynix on HBF global standardization (Feb 25, 2026); TrendForce on HBF demand timeline

Related BEP Research

• The Memory Wars: Why NVIDIA’s 2028 Architecture Ends the AI Chip Competition

• Micron Just Proved the Memory Thesis. Now What?

• The Packaging Paradox: Why CoWoS, Not 2nm, Is the Real AI Bottleneck

• The Hierarchy Rewrites

• The DRAM Squeeze

About the Authors

Rui “Rick” Xie is a researcher in computer systems, memory architecture, and AI infrastructure (rickxie.cn). For this piece he went to the JEDEC HBM3 and HBM4 standards and the primary vendor disclosures directly, built the feasibility triage in the migration map, and corrected the piece’s original framing when the data disagreed with it. That correction is the piece. His views and analysis are his own and do not represent his employer.

Ben Pouladian is the publisher of BEP Research and CEO of BEP Holdings. He led the investment framing and the margin-capture analysis. Framings are credited to their originators throughout. More at www.bepresearch.com

Disclosure: The author holds positions in NVDA, LITE, CRDO, TSEM, ALAB, LSCC, WOLF, ORCL (2027 LEAPS), and BE as disclosed in prior pieces. This is investment research, not investment advice. Do your own work.

Hope everyone had a great Fourth of July holiday. Amazing to see America turn 250 years young.

I spent the last few weeks between London, Paris, and New York City meeting with clients and portfolio companies while my family did some sightseeing. Grateful for the chance to build, invest, and spend time with family across three incredible cities.

Lets dive in we have Wolfspeed in here along with some other updates.

“It’s a little bit of a prediction at this point... like in EVs we had similar predictions and then as the market grew the competition increases and then there’s some pressure on it.”

That was Cengiz Balkas on a call with me last Tuesday. Balkas has been at Wolfspeed since 2006, ran the power and RF device businesses for twelve to thirteen years, and since last year runs every business unit in the company. When the CFO talks, you hear the balance sheet. When Balkas talks, you hear the fab.

Quotes from Wolfspeed management in this piece are from my July 1 call with the company.

The prediction he is hedging is the ladder Gregor van Issum gave me in May, the silicon carbide dollar content per megawatt in the AI rack that I published in The Mohawk Number. A month later, the operator who has run these products for two decades backed the ladder, hedged the dollar figure, and told me one new thing that sits in zero models I have read. This is a short position update, not a new thesis.

Behind the paywall: the operator’s sign-off and his EV caveat in his own words; the three questions he would not touch; the date the 800V story now carries and how this morning’s NVIDIA roadmap fight changes it; the 300mm silicon carbide disclosure with zero revenue and zero coverage; the beats on Gen 5, GaN, 10kV, and CHIPS; how I am handling the run-up; and what breaks the position from here.

Read more

“…unwilling to be associated with the financial statements prepared by management.”

— Ernst & Young, resigning as the company’s auditor, disclosed in an October 30, 2024 filing.

That is close to the worst sentence an auditor can write about a client. Eighteen months later, the same company priced a seven-billion-dollar raise to buy the components for a thirty-nine-billion-dollar book of AI server orders. Both facts are true at the same time, and the distance between them is the whole opportunity. I am going to make you wait for the name, because the name is exactly what the market is using as an excuse not to look.

The market spent two years pricing this company as a governance problem: a delayed 10-K, a short report, an auditor resignation, a delisting clock, federal subpoenas, and this spring a co-founder arrested for conspiring to divert Nvidia servers to China. Each headline knocked the multiple down another leg, and even after the books were refiled the company that builds the racks for the most demanded compute on earth still trades like a distressed box-shifter.

The scandal repriced the governance. It never repriced the demand.

The Discount Came From the Books, Not the Backlog

Every leg of the selloff traces to a governance event, and not one of them touched a unit of demand. Walk the timeline. Ernst & Young flagged the audit committee in July 2024. Hindenburg published its short thesis on August 27. The company delayed its annual 10-K the next day, drew a Nasdaq non-compliance notice on September 20, and confirmed a Justice Department inquiry that same month. Ernst & Young resigned on October 30 and the stock fell roughly 30% in a session. A new auditor was engaged in November, a special committee reported no evidence of fraud or misconduct in December, and the company refiled its delinquent statements by the February 25, 2025 Nasdaq deadline and regained compliance. Every one of those is a real governance failure. None of them is a statement about whether hyperscalers and neoclouds want liquid-cooled rack-scale AI systems.

A company can be a governance mess and a demand monster at once, and the market hates that combination because it cannot underwrite it cleanly. So it discounts the whole enterprise to the level of its worst headline, and the name screens as broken while it ships into the tightest capacity market the data center has ever seen.

This is the inverse of the setup I described in The Quality of the Beat: Why Credo (CRDO) Beat, Raised, and Fell: “The market did not reprice the size of the beat. It repriced the quality of it.” Credo printed a clean beat and fell because the market repriced who the revenue came from and when the next leg arrived. This company is the mirror: the market repriced the integrity of the books and left the quality of the demand alone. Same machine, opposite stain.

The Asset Is the Cooling Loop, and It Just Became Mandatory

This company does not screen as an AI winner because the market still files it under “server assembler,” and that filing is two rack generations out of date. The thing it builds is no longer a box. It is a direct-liquid-cooling rack-scale system, and at the power densities Nvidia is shipping into, liquid cooling stopped being an option and became a physical requirement. This is the company that stood up the racks for xAI’s Colossus in Memphis — 100,000 Nvidia GPUs in 122 days on liquid-cooled systems — a cluster that has since grown past half a million GPUs across H100, H200, and Blackwell GB200 and GB300 parts. The cooling loop is the reason it could be built that fast.

I laid out the mechanism in The Watt Asymmetry. Rack power density is on a three-year ramp from roughly 120 kilowatts at Blackwell NVL72 to 300 kilowatts at Rubin to a target near one megawatt per rack at Feynman. You cannot cool a megawatt with air. Above that threshold liquid cooling is the only path that works, which is why the integrators who can deliver it at rack scale move from commodity to chokepoint as the density curve climbs. This company is one of the named GB200 and GB300 NVL72 direct-liquid-cooling integrators in the Nvidia ecosystem, and it markets the loop on its own numbers: roughly 40% power savings and 98% heat capture per rack against air.

The market is pricing it for the margin profile of bending sheet metal. The asset it actually owns is the ability to deliver the cooling architecture every megawatt-class rack now requires, at the speed the buildout demands.

It Sits Inside the Financing Loop, Not Outside It

The reason the demand never wavered is that the rack integrator is the step where the money becomes a machine. In The Token Dollar, I argued the AI infrastructure economy runs one loop: borrow dollars, buy GPUs, mint tokens, sell tokens for dollars, service the carry, roll the principal, with the GPU fleet as collateral. The rack integrator is the step nobody models, the point where the borrowed dollars convert into the physical asset that backs the loan. Before a GPU mints a single token, somebody has to rack it, cool it, and power it. That somebody is the DLC integrator, and the June raise is exactly that conversion priced in the open.

The June 10 pricing term sheet spells it out: 45.45 million common shares plus 75 million depositary shares of a new 7.00% Series A mandatory convertible preferred, roughly $1.25 billion of common and $3.75 billion of convert, with up to $2 billion more available through an at-the-market program. The use of proceeds is not vague either; the filing states the money funds “the purchase of components to satisfy the approximately $39 billion of orders … from more than 20 customers” the company plans to fill in future quarters. A load-bearing node in a loop measured in the hundreds of billions, with its booked demand written into an SEC filing, still priced like a vendor that might lose its listing.

And the loop is still widening. Morgan Stanley now models the five largest hyperscalers spending roughly $805 billion of capex in 2026 and approaching $1.1 trillion in 2027, up from $765 billion and $950 billion only a quarter earlier, with Oracle’s line compounding at a 116% annual rate off the lowest base. That is the demand curve in dollars, and every one of those dollars has to be racked, cooled, and powered before it computes anything. The rack-and-cool layer is the toll booth the whole number passes through.

Here is the tell. Every other major name that builds these machines has been repriced for that demand this year. Dell, HPE, and Lenovo have all re-rated hard in 2026, each of them at least doubling on the year. One name on the same demand curve has done the opposite and fallen, and the only thing separating it from the rest is the stain on its books. That divergence is the entire trade, and I will name the company below.

xAI Is the Catalyst That Forces the House Clean

The clean-up will not happen because the board found discipline. It will happen because the demand got too large to leave stranded behind a governance cloud. xAI merged with SpaceX in February 2026 into a single entity valued near $1.25 trillion, the most valuable private company in the world, and that treasury now funds the Memphis Colossus build-out where this company is a direct-to-chip liquid-cooling partner. The campus already runs at two gigawatts across three buildings, and on our read the trajectory runs toward roughly 3.5 gigawatts as the SpaceX-funded phase comes on, underwritten by the combined balance sheet rather than by xAI’s own cash flow, which is still deeply negative. When a customer of that scale leans on a vendor, the governance overhang stops being the vendor’s problem alone and becomes a shared problem the supply chain has every interest in resolving fast.

And the balance sheet just got larger. SpaceX came public this month in a reported $75 billion raise, and the dollars behind that print are not really a bet on rockets. They are a bet on power: the gigawatts, the racks, and the cooling loops the model and its agent fleet consume. This is the live case of what I mapped in The $4 Trillion Scarcity Trade — private-market scarcity does not stay private, it becomes capex, and the capex lands on whoever builds the physical machine. The integrator that racks and cools those gigawatts is one of the names that premium has to pay. Today the buyer doubled down again, striking a $60 billion all-stock deal for Cursor, the AI coding agent, buying the demand layer that consumes the very compute these racks run. The money is not slowing; it is integrating vertically, from the agent down to the watt.

You can already see it routing through the system. Google is reportedly paying SpaceX on the order of $920 million a month for Blackwell capacity tied to xAI, the same scarcity marked at the cloud layer instead of the integrator layer. Every dollar of it runs through racks that have to be built and cooled, and the integrator captures a slice of each one. This is one of the cheapest ways to stand in that flow, precisely because the accounting cloud is still keeping institutional money on the sidelines.

And the vendor is not being shy about it. As the IPO priced, the company’s own CEO went on X to congratulate Musk on “SpaceX’s greatest IPO” and said he was “proud to co-build another new Gigawatt AI datacenter for SpaceX and xAI within a year,” calling it his fastest build yet. Read that as the supply side raising its hand in public. The integrator the market is still discounting for an accounting scandal is the one announcing a brand-new gigawatt cluster for the most aggressive compute buyer alive, four days before I wrote this.

Below the paywall: the name and the ticker, the CEO’s on-record line that put the catalyst on a one-year clock, the re-rating gap against the peer group that left this name behind and what it decomposes into, why Nvidia needs this house cleaned, the full bear case including the live China export-control indictment and the customer who holds the pricing whip, and where I stand on the name with sizing.

The company is Super Micro — ticker SMCI. If you guessed it from the auditor and the Colossus build, you were meant to. The point of holding the name was never that it is unknowable; it is that the market is using “the SMCI accounting scandal” as a reason to file the name under untouchable and skip the demand entirely. Now the work.

And the CEO put it in writing. On June 12, as SpaceX’s IPO priced, Charles Liang posted to Elon Musk: “A big congratulations to SpaceX’s greatest IPO and proud to co-build another new Gigawatt AI datacenter for SpaceX and xAI within a year. This will be our fastest TTO yet, and easily one of the biggest and most advanced datacenters in the world.” That is the bull case stated by the one person with full visibility into the order book, four days before this note. It does not erase bear one — a CEO talking his book on X is not an audited number, and this is the same chair whose last CFO sat through an auditor walking out — but it moves the xAI catalyst from “reported” to announced by the vendor, on the record, with a one-year clock attached.

Why Nvidia Needs This House Cleaned

At one megawatt per rack, the number of partners who can actually deliver direct-liquid-cooling at scale on the Rubin and Feynman timeline is small, and every integrator that strands itself on its own balance sheet is capacity Nvidia cannot ship into. A customer with a half-trillion-dollar order book does not have the luxury of letting a key cooling partner delist itself over a filing delay. The pressure on Super Micro to get its house in order is structural: it is Nvidia protecting its own deployment cadence. And it is not hypothetical. After the China smuggling indictment, Jensen Huang publicly pushed Super Micro to tighten its export-control compliance — the literal source of the get-your-house-in-order framing. Nvidia needs its direct-liquid-cooling integrator both solvent and clean, because a key cooling partner tangled in an export-control case is capacity it cannot safely ship into.

So the governance discount is priced as if SMCI is the marginal supplier the ecosystem can shed, while the density curve says it is one of the few that can do the hard physical thing the whole buildout now requires. That is the asymmetry the stain is hiding.

Left Out of the Re-Rating

The cleanest way to see the discount is to set it beside the names that build the same machines. In 2026 every other AI-server integrator got repriced for the demand, and not modestly. Dell is up roughly 250% on the year. Lenovo, the third big rack-scale integrator, is up around 140%. HPE is up about 100% after a 40% revenue quarter. Super Micro is down on the year. Same end market, same Nvidia dependency, the same direct-liquid-cooling demand curve, and it is the only one of the four the market refused to carry higher.

The multiple says what the tape says. On forward earnings Super Micro trades in the mid-teens against Dell near 32 times; on sales it is roughly 0.7 times against Dell’s 2.3. Same demand, a fraction of the multiple, on the AI-server pure-play that actually re-rated. HPE and Lenovo price below Dell as well, and Lenovo is cheaper than Super Micro on both measures, but those are slower-growth, lower-beta franchises, not the GB200-rack pure-play that Dell and SMCI both are. Against the comp that matters, Super Micro is the one carrying a governance discount the others do not, and it has not started to close.

Be honest about why, though, because not all of that gap is mispricing. The peers do not have a co-founder under federal indictment or a $2.5 billion smuggling case in the headlines, and the 33% drop on that news is part of why SMCI sits where it does. Some of the discount belongs there. The narrower, defensible claim is this: the market has priced the overhang as if the demand were impaired too, when the order book and the re-rating in every comparable name say it is not. You are paid for the governance risk; you are not being asked to pay for a demand risk that has not shown up.

And the case for the cheaper builder is not only that it is cheaper. The reason xAI stood Colossus up on Super Micro racks in 122 days is the same reason it is the higher-beta way to own this demand: Super Micro is the direct-liquid-cooling specialist, not a diversified box vendor that also offers cooling. It manufactures its own cold plates, manifolds, and coolant-distribution units, and sells the rack as a building-block system engineered for time-to-online, which is why it tends to ship new Nvidia architectures first; its next-generation Vera Rubin NVL72 platform is a full liquid-cooled stack it markets at ten times the throughput per watt of its Blackwell systems. Dell is the better-capitalized, more diversified franchise with the cleaner balance sheet, and that is precisely why it is the lower-beta, fuller-priced way to own the same demand. The choice is not cheaper-versus-better in the abstract; it is the concentrated cooling pure-play with the governance hair against the diversified incumbent without it, at a large multiple discount. That edge is real today and erodes over time, which is what bear two is about. Whether the discount is a gift or a warning is the entire debate, and it is why I own it small rather than not at all.

The Bear Case

Three ranked risks. The first can change the direction of the call; the other two change the slope.

Bear one: the books are worse than the refiling shows. The whole thesis rests on the demand being clean and the accounting being a contained, resolved governance failure. The 2025 refiling and the new auditor’s clean opinion are the load-bearing assumption. If a future restatement reveals that revenue recognition or related-party dealings were materially misstated in a way the refiling missed — on a name that was already delisted once in 2018 and settled SEC accounting charges in 2020 — then the discount is not a discount, it is the correct price of books that cannot be trusted. The concrete tripwire is the next audit cycle: a clean opinion with no new material-weakness disclosure keeps the call alive; a fresh restatement breaks it.

And the governance risk is not all historical, which is the part that should give a bull pause. In March 2026 a Super Micro co-founder and senior vice president, Yih-Shyan “Wally” Liaw, was arrested on federal charges of conspiring to smuggle roughly $2.5 billion of Nvidia-powered servers to China through a shell-company transshipment network, and the stock fell about 33% on the news. Two things matter for the call, and they cut in opposite directions. The company says it is not itself accused, has opened an independent review, and is cooperating; Liaw is charged, not convicted, and is presumed innocent. That is the mitigant. The aggravant is that an active federal export-control case against a co-founder, on a name whose entire discount is a trust problem, is exactly the overhang that says the house is not yet fully in order. This is the live bear, alongside the refiling, and it is a real reason part of the discount belongs there.

Bear two: it is a thin-margin assembler, and the customer holds the whip. The call is a re-rating bet, not a free lunch, because the integrator layer earns a thin margin on a large revenue line. If Dell, Lenovo, the ODM-direct players, and the Taiwan integrators all build comparable direct-liquid-cooling capacity on the same Rubin timeline, the layer commoditizes and the per-rack value capture flattens toward sheet-metal economics. The customer structure makes that worse, not better. When your demand concentrates in Nvidia upstream and Elon’s compute empire downstream, you are a price-taker on both ends: Liang builds what Musk orders, and Musk, sitting on a fresh $75 billion and the most valuable private company on earth, holds the leverage that comes with being the buyer everyone needs. That guarantees the volume and squeezes the price. It is why the rack-assembler group as a whole, Super Micro included, trades at a fraction of the multiple the chip and cooling-component names command. The demand shows up; the slope of what SMCI keeps from it is the open question.

Bear three: the xAI catalyst is announced, not a signed floor. Liang’s June 12 post moved the Colossus build-out from reporting to an on-the-record vendor claim, which is why this sits at bear three and not higher. But an enthusiastic post is not a multi-year contract. If that spend cadence slips, re-sources to Dell under the same split-supply arrangement that built Colossus 1, or simply proves smaller than the announcement implies, the most visible near-term catalyst weakens. The structural liquid-cooling demand curve survives, but the specific forcing function that cleans the house on an aggressive timeline loses its urgency.

So What?

The lesson here is bigger than one stock. The AI capex cycle has gotten large enough, and the scarcity premium on the private side extreme enough, that the market can no longer afford to leave a key physical node stranded over a governance cloud. The supply chain leans toward self-correction because the demand underneath is too valuable to waste. A scandal that would have killed a vendor in a normal cycle becomes, in this one, a discount on an asset the ecosystem needs cleaned up — usually, not always, which is exactly what the bear case is for.

For the reader, the practical frame is this. Do not screen SMCI as a fundamentals story whose accounting you have to trust line by line. Screen it as the highest-beta, lowest-priced way to express the direct-liquid-cooling super-cycle, with the governance clean-up as the catalyst, the gap to the re-rated peers as the part of the discount that can close, and the refiling integrity as the one risk that actually breaks the call. Watch the next audit opinion, watch the export-control case for any sign it reaches past the charged individuals toward the company, watch the Colossus spend cadence and whether it stays split with Dell, watch the DLC bench for new entrants that would commoditize the margin, and watch the August 13F filings to see which institutions actually took down the $7 billion raise. The $7 billion was sold into the public market through JPMorgan, Goldman Sachs, and Citigroup, so the offering documents name the banks, not the buyers; the identity of the funds that stepped in at a discounted strike, with the scandal and the indictment in plain view, is its own read on how the smart money is scoring the overhang. I want to see exactly who paid.

The backdrop moved this week, too, and it is worth being precise about how. The clean event was geopolitical: on June 14 the United States and Iran signed a memorandum of understanding to halt the fighting and reopen the Strait of Hormuz, lifting the oil-shock overhang that had been weighing on global growth for three and a half months. The Fed, on June 17, was the muddier read: it held at 3.5 to 3.75%, but Kevin Warsh’s first meeting turned hawkish, pulling the cut bias and nodding at a hike, with inflation at a three-year high the committee pinned on the Iran war. So one overhang cleared cleanly and the other only half-cleared. The link, though, runs the right way: a reopening Strait is exactly what should pull the energy inflation back out of the tape that just made the Fed cautious. For the highest-beta way to own the AI buildout, a market that stops pricing an energy shock is the friendlier backdrop, even with the Fed leaning higher-for-longer, which is its own reason to keep the position small.

Where I stand: I am long SMCI, speculative and sized small. Two facts moved me off the sidelines. JPMorgan, Goldman Sachs, and Citigroup underwrote the $7 billion raise and the institutional market took it down with the scandal and the indictment in full public view, which is real capital funding the name at these levels, not retail hope, even at a discounted strike and a 7% coupon on the convert; and the $39 billion order book says the demand the discount is ignoring is already booked, not forecast. That is enough to own it. It is sized small, not full, because bear one is real: an active export-control case and the standing risk that a future audit cycle surprises me are exactly the kind of thing that takes the equity to the wrong number, and on this name the cost of being wrong about the books is the whole position, not a drawdown. So I am paid to own the demand the scandal never touched, sized for the chance the governance is worse than it looks. If the next audit opinion prints clean and the export-control case stays contained to the charged individuals, I add. If either breaks, I am small enough to be wrong cheaply.

The accounting is the discount. The cooling loop is the asset. Underwrite the demand the scandal never touched — small, and with your eyes open.

Coming Up

A follow-up note lands once the next audit opinion is on the record, with any change to the position behind the gate. I am also waiting on the August 13F filings to see which institutions bought the $7 billion raise; the moment the buyers are on the record, paid subscribers get the breakdown. If the Colossus spend cadence prints harder over the next quarter, paid subscribers get the updated read first.

Related BEP Research

• The Token Dollar

• The Watt Asymmetry: $693B in Capex, One Megawatt Per Rack

• The Quality of the Beat: Why Credo (CRDO) Beat, Raised, and Fell

• The $4 Trillion Scarcity Trade

• Nebius: The Tokenmaxxing Factory

Disclosure: Author is long NVDA, LITE, CRDO, TSEM, LSCC, ALAB, ORCL (2027 LEAPS), BE, and SMCI (speculative, sized small). No position in DELL or HPE at publication. This piece is research and commentary, not investment advice. Do your own work.

That is Jensen Huang in March, working the crowd at the Nebius booth at GTC, telling the company’s CRO that “Nebius will take care of you.” Nebius posted the clip with its own two-word caption underneath, “We will.”

On Tuesday, at Nebius’s first Inflection forum in San Francisco, I watched what that promise looks like as a business model. The product story was agents. The investment story was duration.

The market already understands the first neocloud species: CoreWeave, the contracted utility. Borrow heavily, lock in long contracts, satisfy lenders, scale with NVIDIA’s roadmap.

Nebius is trying to be the second species: the merchant generator. Fund with more equity, customer prepayments, backstop contracts, and converts; avoid five-year price locks; sell scarce GPU capacity short into a rising market.

Same GPU shortage. Opposite duration bet.

CoreWeave is a contracted utility. Nebius is a merchant generator.

The market color I keep replaying came out of the hallway. I posted it Tuesday evening

I heard this from a credible source I cannot name. I am not treating hallway color as a signed contract. I am treating it as a live price signal from people close enough to the shortage to know where the bids are. The important part is the structure: premium pricing, short duration, no spare capacity, and buyers still taking the capacity. That is the merchant book in one anecdote.

’s Sources heard the same market at the same event, with a number attached: Google reportedly paying SpaceX $920 million a month for Blackwell capacity, a high premium on an effectively 90-day rolling commitment.

The hallway quote is not the thesis by itself. It is the price signal.

Below the paywall, I’ll walk through the actual underwriting question: whether Nebius has built a merchant GPU book that can reprice into scarcity, or whether it is just another capital-hungry GPU landlord arriving before the next supply wave.

The Workload Without an Incumbent

Start with the keynote. One line has stuck with me since.

“Agentic AI will be a new workload for everyone. No decade of accumulated experience. Everyone starts from scratch.”

That is a claim about moats, dressed as a product pitch. The standing assumption in cloud investing is that AWS, Azure, and GCP carry an unassailable advantage, twenty years of operational scar tissue. The keynote’s claim is that the next workload class resets that clock to zero. If agents really do behave differently from humans, with parallel API calls and continuous multi-step execution, then nobody has a decade of experience serving them, Amazon included.

Nebius wants to be the AWS of true AI, and it said so on stage. It is not pitching models or apps; it wants to own the AI infrastructure stack from rack to token to agent, and the more of the stack it controls, the more it can optimize cost per token. The product slide names the layers SCALE, BUILD, RUN, with Agent Echo, the new agentic layer, sitting on top.

A hallway conversation put ground truth under the vertical-integration claim. Devang Sachdev, ex-NVIDIA, Twilio, and Snorkel AI, who led Nebius’s last two acquisitions, told me the data center heritage runs through Andrey Korolenko, the Head of Infrastructure who spent years building Yandex data centers, and that the racks are proprietary, designed in house. In the company’s telling, owning the rack design squeezes out more tokens per watt than anyone else. That is a company claim relayed by an employee and I weight it as one, but it points at the variable the whole stack is built to win. Tokens per watt is the floor under cost per token. Customers arrive on bare metal, GPUs by the hour, in verticals from robotics to financials.

The two deals Devang led each buy margin above hosting. Eigen AI is a model-optimization team that squeezes more tokens out of every NVIDIA GPU, the same cost-per-token axis as the proprietary racks; Tavily builds real-time search infrastructure for agents, the capability the agent layer was missing. Eigen was announced May 1 at roughly $643 million, Tavily on February 10 at $275 million and up to $400 million on milestones. Over $900 million of announced M&A puts real money behind the stack climb.

I introduced the NeoCloud framework in The NeoCloud Hypothesis: a cloud provider whose “primary business model is deploying NVIDIA GPUs at scale”, with no competing silicon program and no legacy estate in the way. Nebius fits, and I owe readers a correction: the original piece mapped CoreWeave and Oracle, and Nebius was not on my board. It should have been. Consider this piece the fix. What it adds is a layer neither CoreWeave nor Oracle built, a developer platform that climbs from bare metal to tokens to agents. CoreWeave’s edge is deployment speed on NVIDIA’s roadmap; Oracle’s is the database-to-GPU pipeline; Nebius is betting the agentic workload class is new enough that a platform built for it from scratch beats the retrofit.

Agent Echo is the layer I keep coming back to. The keynote described it as a stateful runtime rather than stateless model serving, with orchestration, durable execution across tasks that run minutes to hours, full trace observability, cost accounting, and permissioning with spend caps. Whoever operates the runtime owns the traces, and they become the raw material for optimizing customer agents the way inference endpoints get tuned today. If everyone starts from scratch, the operator accumulating the most agent traces stops starting from scratch first.

The hardware corroborates the workload shift. One customer reportedly runs on the order of 100,000 CPUs against roughly 3,000 GPUs, with early Vera CPU adoption and CPU attach rising; agent workloads are CPU-heavy in a way GPU-by-the-hour hosting does not capture. Whether any of it monetizes still hangs on the trust problem in bear three.

The Ramp

None of the positioning matters unless the buildout under it is real. Nebius went from 10 MW of running power to more than 200 MW in under two years, a 20x ramp, and the stated target is 0.81 GW running by the end of 2026, roughly 4x from here in six months. Contracted capacity sits at 3 GW today and is guided to 4 GW-plus by year end, with two thirds of that 4 GW on owned land and owned power shell rather than leases.

All stated company targets, not my estimates. The customer proof points were the checkable part: P99 inference latency of 400 to 500 milliseconds that one customer called phenomenal, training time cut from months to hours, and an unnamed customer’s 70 percent time reduction on tasks that took two to three weeks. Those numbers are the utilization side of the merchant model; capacity that performs stays sold out between repricings. The scale claim, hundreds of thousands of GPUs on a path the company describes in millions, would put Nebius alongside only the three hyperscalers among publicly available clouds.

The Merchant Model: Equity Buys the Right to Sell Short

The financing structure is built backwards from CoreWeave on purpose — the most consequential structural choice I heard all day. The capital staircase: $2 million of reservation capital 23 months ago, roughly 1 percent of total capex; billions raised last year as the data center build began; tens of billions reserved this year for the gigawatt-scale GPU buildout; hundreds of billions now targeted, structured through customer prepayments, backstop contracts that cheapen construction financing, and convertible notes. Those are order-of-magnitude figures as stated on stage, and announcements were teased as coming very soon; I am treating the staircase as a stated plan until they print.

Nebius is deliberately financed with more equity than debt so that it does not need long-term contracts to satisfy lenders, which frees it to sell short-term reservations into a market where GPU prices are rising monthly, capturing the upside that a long contract locks in for the customer.

The session described demand as unlimited, every new megawatt sold the moment it lands, supply the only binding constraint. Company statements at the company’s own event, and the unlimited assumption gets its own entry in the bear case. The checkable version is narrower — customers now request reservations out of fear of under-supply, not for pricing certainty.

CoreWeave’s lenders require duration. The weighted average contract length stretched to five years, the backlog sits in the tens of billions, and the debt-to-equity ratio printed near 894 percent at the Q4 report. That structure is rational for a debt-funded builder betting that locking today’s GPU price for five years is good business. Duration is the underwriting variable that splits the two books.

In The Token Dollar, I argued that the neocloud financing loop runs borrow dollars, buy GPUs, mint tokens, sell tokens for dollars, service the carry, roll the principal, with the GPU fleet as collateral. Nebius is running the same loop with one leg swapped, equity instead of carry. No carry means no lender forcing duration onto the book; the token side of the loop floats with the market price of compute.

The market validation the session reached for was a recent on-demand deal, capacity sold at what was described on stage as fairly insane prices with no long-term obligation. That is the stage version of the hallway price signal up top.

The margin philosophy was profitability from day one, not revenue multiples. When I pushed on the bare-metal-to-full-stack revenue multiplier in my notes, the question was raised on stage and not answered with a number. That non-answer is the honest gap in the story, and it is the number I want before underwriting the stack-depth claim.

Token Deflation Is the Business Model

The investor session says GPU prices are rising monthly. The customer panel says half of enterprise tasks are migrating to open models at a tenth the cost. Both are true, and for the same reason. Buyers have learned AI is expensive and useful at once, so demand keeps exploding while everyone turns price sensitive.

Roman Chernin, Nebius’s co-founder, posted the reconciliation himself three weeks before the event. From his tweet of May 22: “Start with frontier models → grow usage → collect data → tune cheaper models -> get margins.” He pointed at Anthropic’s revenue and token growth as the preview for open-source tokens. In this plan, per-token deflation is the margin engine. An investor reacting to the post called it Layer 4, where the greatest margin sits — tuned open models serving a myriad of users efficiently.

Amir Efrati of The Information, who moderated the customer panel, framed the moment as a rerun of the public cloud’s adolescence, when AWS customers woke up $20 million over budget and an entire cost-optimization industry was born inside a decade. The stance has shifted from “let’s try AI” to “let’s maximize AI efficiency, but with clear, documented ROI,” and every tool serving that stance routes inference spend toward open models on cheap infrastructure. The panel was not unanimous on the premise; Cognition CEO Scott Wu argued the GPUs are “expensive, but they’re not that expensive” next to the humans they stand in for, and that the real shift is buyers measuring outcomes rather than consumption.

Roughly 80 to 90 percent of enterprise AI tasks do not need the smartest model, open-source models handle 50 to 60 percent of tasks at roughly 10x cheaper and faster, and the hardest 10 to 20 percent still warrant frontier models. The tooling response is converging on AI gateways, one control point with full visibility and free model switching. Databricks runs all internal work through one tool, per VP Nikita Shamgunov, and the visibility surfaced bottlenecks that were never model calls at all, like stacked PRs waiting on CI/CD.

Nebius put a price on the routing table from its own stage. Anubhav Maheshwari, the company’s VP of ecosystem strategy, demoed a healthcare compliance agent that cost $637 per run on a closed frontier model; rerouted onto open-weight models, DeepSeek and then NVIDIA’s Nemotron, the same run cost $24 and finished in 15 minutes instead of an hour. Heath’s Sources has the writeup.

Seen from the supply side, the migration is the same story. The typical customer journey starts on frontier APIs, hits an economics or data-sensitivity wall, fine-tunes an open-source model, and shifts the workload. The open-source frontier lag is roughly six months. Shopify was the concrete example. Latency-sensitive inference stays on GCP near the control plane; everything else gets distributed across providers on price, with fine-tuned models doing merchant search. The cost reckoning is not shrinking AI budgets. It is redistributing them down the price curve, toward open-model inference at scale, the middle layer of Nebius’s stack. I think Nebius wins in exactly that world, where large enterprises understand the use cases but are picky about which models they use for what, and want full control over their data and, above all, their costs.

The orchestrator-worker pattern is the concrete shape of the migration. NVIDIA’s own internal practice, as I understand it, runs Opus as the orchestrator and Nemotron as the worker. The expensive frontier model plans and delegates; the cheap tuned open model executes the volume. My base case is that we end up with more models, and more specialized ones, rather than one god-model carrying every workload. Anthropic’s terms of service now restrict using Claude for frontier-LLM development work, per SemiAnalysis reporting, and per its model card the new Fable model is hardened against distillation and LLM-development use, pushing exactly that work toward open models and the clouds that tune and serve them.

Every voice in that story has an interest in it. The first disinterested signal is showing up in spend data. Citadel Securities’ “Tokenomics” note, built on Silicon Data’s LLM Expenditure Index, reads the index’s recent decline as substitution away from expensive frontier usage toward cheaper models, and cites reports of unexpectedly large token bills. That is exactly the world Nebius wants — and the scarce asset in it is still GPU capacity. When a sellside macro desk charts that shift, the migration has graduated from conference talk to a data print.

Now run the declining index against the merchant book. Tokens deflate per task, total volume grows faster than the deflation under the Jevons dynamic the panel described, and every tuned open model still runs on GPUs somebody has to rack and power. What Nebius sells short-term is capacity, not tokens, so token prices and reservation prices can fall and rise at the same time: the cheaper the work gets, the more of it gets done, and the megawatts stay sold out.

Nebius's CRO opened the day pitching the room on graduating from tokenmaxxing to valuemaxxing, making tokens count rather than counting them. The infrastructure underneath is indifferent to which slogan wins. Nebius Token Factory bills for the GPUs that mint tokens, so a falling price per token pulls more volume onto the same megawatts. That is the tokenmaxxing factory, and the floor under its book is the total number of tokens produced, whatever any single one ends up selling for.

This is also why I am skeptical of the bitcoin-miner AI pivots. Power and shells get you into hosting, not into the token factory, and hosting margins compress as the gigawatts land. If margins migrate up the stack, the miners become the control group for what happens when you own megawatts but not the platform. My read: many of those pivots do not survive the compression, and the failures will get blamed on AI demand when the real problem is stack position.

What This Means for NBIS, CRWV, and NVDA

The split is no longer hypothetical. The newest, fastest-ramping entrant chose the merchant structure, funded it with equity on purpose, and says it is selling every megawatt the moment it lands. That is a market signal. The people closest to GPU pricing are betting it stays supply-constrained through the window that matters. The open-source migration is part of the same bet; per-token deflation grows the workload base that keeps capacity scarce.

NBIS is the purest public expression of the merchant structure and the only neocloud selling a developer platform up to the agent layer. A merchant book marks to the GPU price faster than any contracted backlog, in both directions. The demand filling it skews toward production inference tied to live products with users, a higher-quality book than speculative training reservations. I am watching the financing announcements and the full-stack revenue multiplier before underwriting it properly.

CRWV should now be read through duration. The five-year backlog is the defensive asset in a falling price curve and the capped asset in a rising one.

NVDA wins either way, but the short-duration book is the better tell on pricing, because it marks the GPU price to market every month while a contracted book hides it for years. The scramble underneath is already visible; everyone at the compute layer is trying to learn how much Vera Rubin supply it can actually get, and how much its competitors are getting, a sorting process the hallways described as cutthroat. And the hyperscalers’ real exposure is whether the agentic workload class grows up on someone else’s platform.

The Bear Case and the Tripwire

Four risks, ranked. The first can change the direction of the thesis. The rest change the slope.

Bear one: the merchant book has no floor. Short-duration reservations capture a rising GPU price curve, and they reprice within a quarter if the curve rolls over. A Rubin-era supply wave or a hyperscaler custom-silicon ramp that eases scarcity would hit Nebius’s revenue line faster than any neocloud carrying a five-year backlog; so would plain macro softness. The hallway consensus at the event put dates on that risk: constrained until sometime next year, possible softening mid-2027, and some attendees pushed relief out to late 2027 or 2028. The merchant book has to earn its premium inside that window.

Bear two: the merchant structure is not proprietary. Anyone with access to equity capital can copy the balance sheet, and if the curve keeps rising, every well-funded GPU cloud can shorten its book and chase the same repricing. The harder half to copy is the stack above the financing, the half the Eigen and Tavily checks just bought, and the bitcoin-miner pivots are the live control group: power and shells with no platform, competing for the same hosting dollars. If Nebius’s margins over the next year look like hosting rather than Layer 4, the merchant premium was never defensible and the company is one well-financed landlord among many.

Bear three: Agent Echo is a product for an unsolved problem. The differentiated top layer monetizes only if enterprises deploy agents at scale, and that needs a security and verification layer nobody has shipped. Nebius’s own security session was candid: agentic security is unsolved industry-wide, the preferred approach is identity-based access control at the agent runtime, and the panel conceded the industry has worked the problem for three years without cracking it. If enterprise agent deployment stalls on trust, the agentic layer monetizes late and the stack-depth premium compresses toward bare metal.

Bear four: the staircase needs open capital markets. Hundreds of billions is a target, not a balance sheet. The whole staircase works while credit is loose and GPU collateral is bid. A spread blowout forces an ugly choice: sign the long contracts the model was built to avoid, or slow the buildout while competitors with contracted backlogs keep building. Mostly slope, but at the tail it bends direction, because the merchant edge dies the day a lender demands duration.

And the tripwire, so this is checkable later: the first quarter Nebius’s realized reservation price declines sequentially, or the first three-year-plus contract it signs. Either one says the merchant thesis is breaking.

I do not hold Nebius today. I expect that to change soon, and I will disclose when it does.

Underwrite the duration of the book, not the size of the buildout.

Related BEP Research

• The NeoCloud Hypothesis

• The Token Dollar

• The $4 Trillion Scarcity Trade

• The Verification Gap: Who Audits the Agent Swarm?

• TurboQuant and Dynamo: Why the Market Has the Memory Thesis Backwards

Elsewhere: covered the same event from the product side in Sources, including the panel debate this piece draws on and the $920M-a-month Google-SpaceX detail. If you follow the AI race, his newsletter belongs in your inbox.

Disclosure: Long NVDA, LITE, CRDO, TSEM, LSCC, ALAB. Long ORCL 2027 LEAPS. Long BE. Long WOLF speculative. No position in NBIS at publication; I expect to initiate one soon and will disclose when I do. Positions current as of publish. This is opinion and analysis, not investment advice.

“My job is not to stay independent. My job is to do whatever we can to create the maximum value.”

That was Gregor van Issum, Wolfspeed’s CFO, on a call with me last Friday afternoon. The man running the balance sheet of the only fully-built 200mm silicon carbide device fab on US soil told me, unprompted, that maximizing value is the only mandate and independence is one option among several. He then pulled up his notes and walked me through numbers I have not seen in a published model on this name.

I have spent three weeks underwriting Wolfspeed as an asset-vs-replacement-cost mispricing in The Watt Tax. The call did not soften that frame. It hardened it.

Behind the paywall, in the CFO’s own words: the exact dollar-per-megawatt SiC content for the 800V rack and the next generation up (the number I have not found in a published sell-side model), cross-checked against the BofA Global Research AI power semi model Vivek Arya published this week; why Wolfspeed is already printing data-center revenue at roughly 10 percent of the company while the eight named cohort suppliers BofA tracks sit at zero in CY25; the five markers from the call that establish this is a legitimate competitor with relationships and shipping product, not a bankrupt comeback story; the real Mohawk Valley utilization figure and why it is the bull case; the Apollo step-up clock that triggers in June and why Gregor will not let it set his refi timeline; the on-the-record signal that a competitor “valued far beyond us” is infringing Wolfspeed’s IP; the cleanest way this trade breaks and what I am watching; and how I am sized. This is the work. If you want the number, step inside.

Read more

Two quick notes. This explainer is free. BEP Research just crossed 3,000 subscribers, and I’m marking the milestone by taking it out from behind the paywall. I also built an interactive version so you can watch a single link flap stall a GPU cluster in real time. If it’s useful, a share is the best thanks.

What a “Flap” Actually Is, in Plain English

A “link flap” is a network connection that keeps dropping and reconnecting. On a laptop, it is the flaky Wi-Fi blip that makes a video call freeze for a second. Annoying, not catastrophic. The reason it matters for AI is that an AI training cluster is the opposite of a laptop: it is one hundred thousand GPUs that all have to stay in perfect lockstep.

Here is the mechanism, and it is the whole piece in one paragraph. Training a large model means the GPUs repeatedly do a synchronized step called an all-reduce, where every chip shares its partial result with every other chip and waits for everyone before moving on. It is a group operation. One slow runner stalls the entire group. So when a single optical link between two switches flaps, even for a moment, the GPUs that depend on that link fall out of sync, the collective operation stalls or restarts, and thousands of otherwise-healthy GPUs sit idle waiting. The flap is local. The damage is cluster-wide.

Now scale that to the part that makes it an investment story rather than an engineering footnote. The component most likely to flap is the laser inside a commodity optical module, also called a transceiver, the part that turns electrical signals into light and back. Lasers are the least reliable, most thermally sensitive, most failure-prone element in the link. As clusters grow toward gigawatt-class power draw and the link count per cluster explodes, the math gets ugly fast: more links, each with a small flap probability, means more flaps per hour, which means more idle GPU time. Credo CEO Bill Brennan has said publicly that one customer was losing 20 to 30 percent of its productivity to link flaps before deploying the Zero-Flap approach. At hyperscale, that is the gap between the fleet a customer paid for and the fleet it can actually use. Reliability here is not a feature. It is the limiting factor on usable compute.

This is the same point I made when Credo first preannounced its blowout quarter. As I wrote in Credo Just Proved the Optical Thesis: “One hyperscaler came to Credo because they were losing 20–30% of their uptime fighting optical link flaps.” That was framed then as a copper-versus-optics story. The Zero-Flap optical line is the same reliability philosophy, pushed out to the optical distances where copper physically cannot reach.

▶ See it in motion: I built an interactive simulation. Toggle commodity vs Zero-Flap optics and watch a single link flap stall the cluster and idle the GPUs in real time. tools.bepresearch.com/zero-flap

How Zero-Flap Works: Three Layers, One North Star

Zero-Flap attacks instability on three layers at once, which is exactly why Credo calls it a system and not a component. Start with what every optical link has to do: convert an electrical signal to light using a laser, push it down a glass fiber, and convert it back using a photodetector, with a digital signal processor (DSP, the chip that cleans up and reconstructs the signal) on each end. Every one of those steps is a place a link can degrade. Zero-Flap engineers against degradation at all three of the layers it can control.

1. Hardware: fewer lasers, fewer failure points. Through the Dust Photonics acquisition, Credo brings in silicon photonics, an approach that integrates much of the optical function onto a silicon chip and uses substantially fewer lasers per link. Fewer lasers means fewer of the single most failure-prone parts, better power and cost per link, and less exposure to the industry-wide laser supply shortage. Paired with an optical DSP tuned specifically for link stability rather than raw throughput, the hardware is built to flap less in the first place.

2. Software: catch the flap before it happens. Credo’s “PILOT” software, integrated at the switch level through the switch maker’s software development kit (SDK), continuously reads link-health telemetry, the bit error rates, signal-quality histograms, and laser-degradation signals coming off each module. Think of it as the check-engine light for the optical link: it reads the signals that precede a failure and autonomously detects and mitigates an instability condition before the link actually drops the cluster. Prevention, not cleanup.

3. Integration: own more of the stack, see more of the link. Because Credo controls both the DSP and, through the silicon-photonics intellectual property, the photonic integrated circuit (PIC, the chip where the light is generated and routed), it gets tighter DSP-to-PIC integration than a vendor selling one piece. Tighter integration means richer telemetry, deeper diagnostics, and system-level optimization that a component supplier stitching together someone else’s parts simply cannot match.

The result management claims is meaningful improvement in network reliability, faster time to a stable cluster, and longer uptime. This is the same mechanism I flagged the day the Q3 call landed. As I wrote in Credo’s Q3 Call Answered Every Question, what makes Zero-Flap a moat is the system integration delivering “autonomous detection and mitigation of link flap events before they impact the cluster.” The defensible part of a system like this is the hardware integration beneath it, far more than any single component, and that distinction is what the bear case below turns on.

Why It Matters Commercially: Three-Digit ASPs Change the Mix

The reason Zero-Flap is the highest-leverage part of the Credo story is that it sells a system, not a part, and systems capture more dollars per link. On the Q4 FY26 call, Credo’s CFO framed the economics plainly in Q&A: Zero-Flap optics carries three-digit average selling prices (ASPs), versus two-digit ASPs for the discrete building blocks, the standalone optical DSP and the photonic integrated circuit. The same physical link, sold as an engineered reliability solution instead of a bag of components, captures roughly an order of magnitude more revenue. That is the whole commercial argument in one ratio.

Stack that ASP onto the volume story and you get management’s own math: as Zero-Flap ramps, it becomes the largest revenue contributor inside Credo’s optical portfolio. By management’s framing on the Q4 call, which is worth confirming against the filing rather than the auto-generated transcript, Zero-Flap is guided above $100M in FY27 as one of three optical legs each above $100M, for a total optical business above $600M. The discrete DSP and PIC legs ramp more linearly. Zero-Flap is the hockey stick, steep and back-half FY27 and FY28 weighted, because it is gated on hyperscaler qualifications converting to volume rather than on smooth quarterly shipments.

That hockey-stick shape is the source of both the upside and the risk, and I will be precise about both. The upside: an order-of-magnitude ASP advantage on the fastest-growing optical leg, ramping into a reliability problem that gets worse, not better, as clusters scale. The risk lives in the back-half timing, which is the bear case below.

The Strategic Option: A Direct Path to Co-Packaged Optics

The Dust silicon-photonics roadmap is the part of Zero-Flap the consensus model mostly overlooks, and it is low-cost optionality on the next architecture. The same silicon-photonics intellectual property that lets Credo use fewer lasers today scales across 800G, 1.6T, and a path to 3.2T network speeds. That roadmap is also the direct on-ramp to co-packaged optics (CPO) and near-packaged optics (NPO), the architectures where the optics move onto or next to the switch chip itself, with initial revenue around FY2028.

This is where Zero-Flap becomes more than a product line and starts to look like a position in the next architectural transition. The physics is the lane-speed jump from 100G to 200G per lane, which makes the laser the binding, reliability-limited component and reshuffles who captures the margin. Zero-Flap is Credo’s answer to that physics from a different layer: instead of competing to build the most reliable laser, it engineers a link that needs fewer of them. As the industry moves toward CPO, the company that already owns the DSP-to-photonics integration and the reliability software is positioned to participate rather than be displaced.

Why It Matters Technically: Reliability Becomes the Binding Constraint

As AI clusters scale toward gigawatt-class power and denser architectures, even isolated link instabilities stop being a tail risk and become the dominant driver of cluster availability. The variables that matter at that scale, cluster breakup time, GPU utilization, overall system availability, are all downstream of link reliability. Add more links and the aggregate flap rate rises. Pack them denser and thermal stress rises, which is exactly the condition lasers hate. The reliability problem does not get diluted by scale. It compounds with it. Credo calls reliability its “North Star,” and at gigawatt scale that label describes the binding constraint rather than dressing it up.

This reframes what Credo actually is. For a year the consensus model has treated Credo as an active-electrical-cable (AEC) company, a copper interconnect supplier, defending a niche against the optical transition. Zero-Flap inverts that. Credo is now selling optical transceivers directly, as a system-level reliability product, moving up from a SerDes and DSP and AEC supplier into the optical TAM long dominated by Coherent, Lumentum, and the merchant transceiver vendors. The product the market skips in the model is the one that changes what the company is.

The Bear Case

Three risks would break this thesis, and I am ranking them by what changes the conclusion versus what only changes the magnitude.

1. The reliability claims are management-asserted, not independently benchmarked. “Zero-flap” is a marketing term. The reliability and autonomous-mitigation claims come from Credo, not from a neutral third party with published field data across multiple hyperscalers. One number deserves special care: the often-quoted 1000x reliability figure belongs to Credo’s copper AECs measured against laser-based optical modules, and there is no comparable independently published figure for the Zero-Flap optical line itself, so do not carry it across. Optical reliability products live or die on the first deployment year of real-world data, and the high-ASP optical line is barely into that window. If early field data disappoints at a major customer, the system-moat story compresses fast and the stock reverts to being priced as an AEC incumbent with an optical side project. This is the risk that would change the direction of the thesis, not just the slope. It is also the risk I flagged when I corrected my own Credo view in Credo Was the Wrong Question: the Zero-Flap ramp is real, but it is a concentrated bet on one product line through the first year of field proof.

2. The high-ASP ramp is the latest, least-proven, most supply-chain-dependent leg. The hockey-stick shape means the execution risk is concentrated in the back half of FY27 and into FY28. A hockey stick that does not bend on schedule is just a flat line that disappointed. This one depends on qualification timelines holding, additional hyperscalers signing on, and the silicon-photonics supply chain delivering into a known laser shortage. This risk changes the slope and the timing, not the direction, but in a name where the ramp is the whole story, slope is most of the value.

3. Optics is a far more competitive arena than AECs, and software-defined reliability is hard to defend. In copper AECs Credo holds a dominant share. In optics it is the entrant against Coherent, Lumentum, Marvell, Broadcom, and the merchant transceiver and DSP vendors, all entrenched. Three-digit ASPs are an invitation to competition, not a moat by themselves. And the software layer, the part Credo leans on hardest as the differentiator, is exactly the kind of telemetry-and-SDK function a large hyperscaler could choose to build in-house. If the customers who most need Zero-Flap decide reliability software is something they own rather than buy, the moat thins from the most valuable end first.

So What?

Stop reading Zero-Flap as a marketing name and start reading it as the answer to the question the whole AI buildout is converging on: how do you keep a hundred-thousand-GPU cluster usefully online when the single most failure-prone part, the laser, sits in every link and the link count is going parabolic? Network reliability has quietly become the binding constraint on usable compute, the same way memory bandwidth and power density became binding constraints before it. Zero-Flap is the only optical product engineered, in hardware and software together, to attack that constraint directly.

For Credo specifically, that is why this one product line is the highest-leverage part of the story. It carries the richest ASPs, it ramps into a problem that worsens with scale, and on management’s own math it becomes the largest leg of the optical business and cheap optionality on the co-packaged-optics transition. The catch is price: even after its post-earnings drop Credo trades around 32 times trailing sales, a multiple that already capitalizes much of this optical optionality, which is why the proof matters more than the next print. I hold Credo, and Zero-Flap is the reason I read it as a multi-product optical company now rather than the copper incumbent it was a year ago. What I am watching is the first year of independent field-reliability data from the hyperscaler deployments, because that is what turns a management claim into a moat.

The market is modeling the company Credo was.

The product it skips is the one that says what Credo is becoming.

Read the reliability line, not the AEC line.

Coming Up

A follow-up deep dive on the Dust silicon-photonics roadmap and the path to co-packaged optics is in motion, pending a conversation with Credo’s optical leadership on the qualification pipeline beyond the anchor hyperscaler and the FY2028 CPO timeline. I will also be tracking the next two optical earnings prints from Lumentum and Coherent for the laser-supply read-through that sits underneath every Zero-Flap link. Paid subscribers get the optical-leadership conversation first.

Related research: The Great Photonic Divergence (why the 100G-to-200G lane jump makes the laser the binding, reliability-limited component and reshuffles the margin).

Disclosure: The author holds positions in NVDA, BE, LITE, CRDO, TSEM, LSCC, ORCL 2027 Calls, and ALAB. This is not investment advice.

Read the Two Lines Brennan Buried, Not the Headline Number

I’ve been pretty busy the last few days from GTC Taipei to Credo Earnings. Excited to be up in SF next week for the Nebius inflection event. Give me a shout if you would like to meet up.

I read the earnings release and listened to the call. My biggest takeaway: read the 1.6T transition exchange before you read the revenue line.

When Brennan was asked about 200G-per-lane signaling, he said FY27 revenue tied to that transition will still be relatively light because the industry is not there yet. He even added, almost as a throwaway, that there is always a rumor about delays.

On the optical DSP side, he went further. Yes, Credo has several customer programs. But he also admitted Credo does not ultimately control when those transceivers actually go to market.

That matters. On his own print day, the CEO effectively put the highest-ASP transition in Credo’s roadmap on someone else’s calendar.

The number on the tape was excellent. Revenue of $437 million was up 157 percent year over year, at a non-GAAP gross margin of 68.3 percent. Non-GAAP EPS of $1.16 cleared the roughly $1.03 the Street was carrying; GAAP EPS was $0.88. Full-year FY26 revenue tripled to about $1.34 billion, and non-GAAP net income rose roughly fivefold to about $662 million. Management then guided Q1 FY27 to $465 to $475 million and framed the full year as more than 80 percent revenue growth, with non-GAAP operating expenses growing roughly half that fast and non-GAAP net margin held near 50 percent. The operating leverage is real. By every headline metric, this re-rates a stock higher.

It re-rated lower, by roughly 12 percent. Part of the reason is mechanical and unforgiving. The $437 million beat the published consensus near $432 million but missed the buy-side whisper closer to $440 million. For most stocks a few million against an unofficial number is noise. For a stock that went into the print near 52-week highs at roughly 41 times trailing sales and about 127 times GAAP earnings, it is not. That is the gap between perfection and merely excellent, and a perfection multiple does not survive merely excellent.

Three forces hit the stock at once, and they are worth separating because only one is about Credo. It had run up into the print. Capital was rotating out of data-center-capex names and into beaten-down software the same week, a positioning move that does not care what Credo reported. Underneath both sat the company-specific repricing: the quality of the beat. The first two are market structure and they pass. The third is the one that should change how you underwrite the name, because when a company beats, raises, and falls, the market is telling you the numbers were already in the stock and that something the numbers do not fix is now the binding constraint on the multiple.

The market did not reprice the size of the beat. It repriced the quality of it.

Quality has two pillars here, and both sat under the revenue ramp the whole time the line was tripling. The first is who: in fiscal 2026, two customers at roughly 39 percent and 32 percent of revenue, three hyperscalers at roughly 88 percent. The second is when: the FY27 inflection that justifies the forward multiple is back-half-and-FY28 weighted, and management openly says it does not set the clock. Concentration is the direction risk. Timing is the slope risk. Together they are the discount the market just attached to a demand story it otherwise confirmed.

Pillar One: The Beat Did Not Dilute the Customer List

A flawless quarter does not reduce concentration, and concentration is the risk that went unpriced while the revenue line tripled. In fiscal 2026, two customers were roughly 39 percent and 32 percent of revenue, and three hyperscalers were roughly 88 percent of the total. Credo went into the print near 52-week highs, at roughly 41 times trailing sales, on a multiple that needed a flawless quarter to hold. It got one. The multiple still could not hold, because a beat-and-raise cannot dilute the customer list, and the customer list is what a perfection multiple was quietly ignoring.

This is also where the April bear case and this print part company. In April, the Credo bear was architectural: that co-packaged optics would kill copper and run the AEC business over. That case was answerable from NVIDIA’s own roadmap. As I wrote in Credo Was the Wrong Question: “the entire bear case rested on a timeline assumption that does not match NVIDIA’s own disclosed plan of record.” That was true of the copper-versus-optics bear, and copper stays inside the rack through at least 2028. It is not true of the concentration bear. Customer concentration is not a timeline assumption that resolves with the next roadmap slide. It is a structural feature of the business, and the market just decided to start charging for it.

It also tests one of my own prior reads. Three months ago in Credo’s Q3 Call Answered Every Question, I flagged the same top-two shape, largest customer at 39 percent and second at 32 percent, and called the trend improving fast off the prior 86 percent single-customer base. The improvement is real and directional. But landing at roughly 88 percent across three hyperscalers is not the diversified base a perfection multiple assumes. The direction was right. The level is still the live bear. That is the gap between the demand being good and the equity being cheap.

Behind the paywall: why the FY27 optical guide is a forward promise and not booked revenue; the two-path 1.6T architecture math and what each path pays Credo per port; why the highest-ASP optical leg is also the latest and least proven; how the Weaver gearbox just intersected the BEP memory thesis on a named customer; the emerging FY28 stack the multiple is paying for early; the three ranked bears that would change the position; and where I am holding CRDO.

Read more