Hi Ben, this is a wonderful piece, and I thoroughly enjoyed reading it. I do have a question about the investment thesis around agent orchestration—a theme that’s been coming up more frequently, especially in the recent earnings calls from Intel and AMD regarding strong demand for server CPUs. While I’ve seen some commentary suggesting that CPU demand won’t mirror that of memory, since CPUs don’t directly improve model performance, I’m curious to hear your take on the role of CPUs in the future of agentic workflows. Thanks!
Thanks! Great question. This is something Jensen touched on recently that deserves more attention.
The key insight is that agents won’t replace CPU workloads, they’ll multiply them. GPUs power the agents (the reasoning, the inference), but those agents then execute traditional tasks on CPUs. When an agent writes code, queries a database, sends emails, or manipulates spreadsheets, all of that runs on conventional compute.
Think of it like having unlimited minions. Sure, the minions are “AI-powered,” but they’re still hammering your Jira board, spamming Slack, and running Excel macros like the rest of us. More minions = more traditional work = more CPU demand.
So we’re not looking at GPUs eating CPUs. We’re looking at an explosion of digital minions doing digital busywork. The memory thesis still holds, but CPU TAM expansion is the underappreciated B-side of the agentic story.
An interesting take. If you think about AI agents as the equivalent of employees on a network, we’re basically going to increase the population by 10x without much training. Sort of like ICE. One area (I hope and believe) will serve as a template is industrial. Controlling, managing, and extracting value from machine data is an art. I wrote case studies on a wastewater utility that could predict sewer blowouts 12 hours in advance. Then there were people who couldn’t figure out why trucks were blowing up for years. It’s a test environment for the rest of us.
Michael - the ICE parallel is sharp. Industrial is exactly right as a proving ground: bounded failure domains, existing sensor infrastructure, and actual downtime costs that justify verification investment. Spent 14 years scaling a commercial lighting company and the gap between ‘sensor data exists’ and ‘we actually act on it’ was massive. The wastewater case is perfect—predictive windows give you time to verify before acting. Consumer AI doesn’t have that luxury. Appreciate the framing
Great question. The path is hardware-rooted attestation. TEEs like NVIDIA Confidential Computing let GPUs cryptographically prove what code they’re running. Combine with Merkle tree verification of model weights and you can verify “Agent X ran Model Y” without auditing every inference.
For scale, probabilistic verification with cryptographic commitments. Make cheating detectable, not impossible.
The real catalyst will be enterprise demand forcing the silicon stack to respond.
You’ve nailed the core tension. The latency tax on formal verification is real, but the alternative is hoping your LLM-as-Judge catches the 0.1% of cascading failures that actually matter. Working on a piece exploring whether the answer is tiered verification lightweight sampling for routine paths, formal methods only at trust boundaries. Curious what you’re seeing in multi-agent coordination that might work
Hi Ben, this is a wonderful piece, and I thoroughly enjoyed reading it. I do have a question about the investment thesis around agent orchestration—a theme that’s been coming up more frequently, especially in the recent earnings calls from Intel and AMD regarding strong demand for server CPUs. While I’ve seen some commentary suggesting that CPU demand won’t mirror that of memory, since CPUs don’t directly improve model performance, I’m curious to hear your take on the role of CPUs in the future of agentic workflows. Thanks!
Thanks! Great question. This is something Jensen touched on recently that deserves more attention.
The key insight is that agents won’t replace CPU workloads, they’ll multiply them. GPUs power the agents (the reasoning, the inference), but those agents then execute traditional tasks on CPUs. When an agent writes code, queries a database, sends emails, or manipulates spreadsheets, all of that runs on conventional compute.
Think of it like having unlimited minions. Sure, the minions are “AI-powered,” but they’re still hammering your Jira board, spamming Slack, and running Excel macros like the rest of us. More minions = more traditional work = more CPU demand.
So we’re not looking at GPUs eating CPUs. We’re looking at an explosion of digital minions doing digital busywork. The memory thesis still holds, but CPU TAM expansion is the underappreciated B-side of the agentic story.
An interesting take. If you think about AI agents as the equivalent of employees on a network, we’re basically going to increase the population by 10x without much training. Sort of like ICE. One area (I hope and believe) will serve as a template is industrial. Controlling, managing, and extracting value from machine data is an art. I wrote case studies on a wastewater utility that could predict sewer blowouts 12 hours in advance. Then there were people who couldn’t figure out why trucks were blowing up for years. It’s a test environment for the rest of us.
Michael - the ICE parallel is sharp. Industrial is exactly right as a proving ground: bounded failure domains, existing sensor infrastructure, and actual downtime costs that justify verification investment. Spent 14 years scaling a commercial lighting company and the gap between ‘sensor data exists’ and ‘we actually act on it’ was massive. The wastewater case is perfect—predictive windows give you time to verify before acting. Consumer AI doesn’t have that luxury. Appreciate the framing
Great question. The path is hardware-rooted attestation. TEEs like NVIDIA Confidential Computing let GPUs cryptographically prove what code they’re running. Combine with Merkle tree verification of model weights and you can verify “Agent X ran Model Y” without auditing every inference.
For scale, probabilistic verification with cryptographic commitments. Make cheating detectable, not impossible.
The real catalyst will be enterprise demand forcing the silicon stack to respond.
You’ve nailed the core tension. The latency tax on formal verification is real, but the alternative is hoping your LLM-as-Judge catches the 0.1% of cascading failures that actually matter. Working on a piece exploring whether the answer is tiered verification lightweight sampling for routine paths, formal methods only at trust boundaries. Curious what you’re seeing in multi-agent coordination that might work