Everyone argues about which GPU buys the frontier. Meanwhile, the biggest cloud deal of the year is for the boring chip that runs your agent's tool calls, browsing, and code execution.
Agents do not live inside the model forward pass. They live in the harness around it, and that harness is CPU work. That is why a lab would lock up seven years of general-purpose compute while everyone else fights over HBM.
We covered this before it was news: the agent workload already moved to the CPU.
Top comments (0)