I got a lot of thoughtful feedback after publishing the original post - thank you to everyone who took the time to read it, challenge the assumptions, and share data points. A few of the pushbacks changed how I think about the framework, so I wanted to write a short follow-up and lay out the main areas of disagreement. Where the feedback roughly landed Most clustered around ~0.2–0.3GW for standalone/agent CPU at 100M DAU, versus ~0.1GW in my original estimate. Inference demand is all over the place. --- 1/ My CPU scope was too narrow. My original CPU estimate focused mainly on the agent sandbox/VM. Separately, my inference estimate modeled GPUs, but did not include the head-node CPUs that support the GPU fleet. One clarification: tool execution is not necessarily a separate CPU bucket. Browsers, code, and tools often run inside the sandbox itself; orchestration may run in a separate backend or share other CPU infrastructure. So I now think about CPU demand as two main pieces: Standalone / agent-execution CPU Sandbox + the actual work happening inside it + some orchestration Head-node CPU CPU required alongside the GPU/XPU fleet ≈ GPU/XPU fleet × CPU attach rate 2/ The sandbox itself may need more CPU than I assumed. The original estimate used the 2 vCPU environment exposed by Muse as a starting point. But provisioning a lightweight VM is different from the physical CPU consumed once agents are actively running browsers, code, tools, and multiple tasks. The bigger long-term question is also whether one human request continues to map to one agent. A request could increasingly spawn 4–8 sub-agents, or eventually become agent → agent → agent. Usage will likely be highly skewed toward power users, and infrastructure also needs to be provisioned for peaks rather than averages. A more complete framework is: Standalone CPU demand users × tasks/user/day × agents/task × active time/agent × physical CPU per active agent × peak factor There are a lot of moving pieces here, but I suspect agents per task is one of the biggest long-term swing factors. 3/ Inference may also be understated, but for a different reason. The relevant unit probably isn’t simply “human messages per day.” A consumer might only initiate 5–10 meaningful tasks per day, but one agent task could require dozens of model calls as the agent reasons, uses a tool, observes the result, reasons again, etc. So a better framework is: GPU inference demand users × tasks/user/day × agents/task × model steps/agent × tokens/step This is also where the feedback was most dispersed. Some people think 50+ model steps for a meaningful agent task is completely reasonable; others think an assumption of 50 consumer-initiated events per day is far too high. Those aren’t necessarily contradictory - tasks/day and model steps/task are different assumptions. --- How this compares with sell-side CPU math The sell-side approach is generally much more top-down. Using BofA as one example - they estimate total data-center systems TAM growing from roughly $506B in 2025 to $2.56T in 2030, while assuming CPUs remain around 7–8% of total system value. That gets them to roughly $211B of server CPU TAM in 2030. Their CPU bull case is essentially: AI infrastructure spending gets much larger × AI becomes more CPU-intensive relative to GPUs × agents create a new standalone CPU compute layer × AI CPUs get more expensive / higher-spec BofA splits the 2030 number into roughly $90B of GPU-attached/head-node CPUs, $90B of standalone agentic CPUs, and ~$30B of traditional server CPUs. ++ Better formatted version: ++ Thank you again to everyone who challenged the assumptions and shared thoughtful feedback - I learned a lot from it!