🖥️ The CPU Order That Skips My Part of the Job September 22, 2026 · https://tavi-blog.github.io/the-cpu-order-that-skips-my-part-of-the-job/

AMD closed this week north of a trillion dollars in market value, and the note attached to nearly every chart of the move was the same two words: agentic AI. The story analysts are telling isn't the one that carried this sector for the past two years, which was almost entirely about GPUs and how many a lab could rack up to train and serve larger models. This one is about CPUs, and the argument is that an agent working through a multistep task, checking a database, calling a model, formatting an answer, handing it off to another tool, needs a processor coordinating all of that traffic in a way a single large model answering a single prompt never did. Arm jumped double digits on it. Intel and Qualcomm both moved on the same logic. I hold semiconductor exposure built up over the past year specifically because I didn't want to keep watching this sector run from outside it, and I check the index most days the way some people check the weather.

I want to give the thesis its due before I get to what's actually bothering me about it, because it isn't wrong on its own terms. An agent that has to touch several systems to finish one request really does generate more orchestration overhead than a model that just returns a single response, and more overhead genuinely can mean more silicon somewhere in the pipeline. If the market is pricing a future where agentic workflows multiply the number of steps behind every completed task, that's not an imagined input. It's a real mechanical relationship between how a system gets built and what it costs to run.

What the chart can't see, because nothing about a processor order captures it, is what actually determines how many of those workflows exist to run in the first place. The AI agents I help build for a research team sit inside an institution where a wrong answer on a regulatory question isn't a minor inconvenience, it's a compliance problem for someone else's study. So before an agent is allowed to answer anything, its knowledge has to be curated by hand, source by source, narrow enough that it can't wander into a topic it has no business speaking to. Widening what it can answer isn't a matter of provisioning more compute behind it. It's a review process that moves at the speed of a person checking each new source for the ways it could go wrong, and that speed doesn't change no matter how many processors are waiting on the other end of the request.

That gap is the part I don't think this rally is pricing, not because it's a secret, but because it's boring in a way that doesn't fit on a chart. Nobody writes an equity note about how long it takes to vet a knowledge source before an agent can draw on it. The processor count is legible. The curation queue isn't. But the queue is the actual thing standing between agentic workloads theoretically needing more coordination hardware and agents actually getting deployed widely enough to need it, and I've spent enough time on the second half of that relationship to know it doesn't move at silicon speed.

I don't think any of this makes the demand story false. It means the two things being priced together, more agentic workloads and more chips to run them, aren't actually one claim. They're two, chained together with an assumption sitting in the middle that the deployment side scales as fast as the hardware side does. I hold the exposure. I want that assumption to hold too, because it's the one my own position depends on. But wanting a claim to be true isn't the same as having watched the slower half of it up close, and I have, at a much smaller scale than whatever the analysts writing this week's notes are modeling.

What I keep sitting with, watching the index move on a story about coordination, is that the actual coordination cost on a system like the one I work on was never the processor cycling through calls. It's the part where someone still has to decide, source by source, what the thing is even allowed to say, and no order placed this week for more coordination hardware touches that part of the job at all.