Inference Chips Are the New Gold Rush, and Most of Us Missed It
Everyone spent September arguing about which model was smartest. October's funding told a different story: the money went to the chips that run those models. Positron AI's $875 million raise for its Asimov inference chip and Tenstorrent's $1.5 billion Series E both landed in the back half of the month, and neither is about training a new frontier model. They're about serving the ones we have, faster and cheaper.
Here's why that matters to anyone who just uses the tools. The price you pay per query is mostly an inference cost. When better silicon shows up, the per-token price drops, and the cheap tier gets good enough to do real work. We already saw it this month: Claude Haiku 5.5 came in about 90% cheaper than its predecessor, and GPT-6 Sol dropped roughly 50%, with three labs converging on the same $2 in and $10 out pricing within nine days.
The risk is concentration. If two or three chipmakers own the efficient path to serving models, they own the toll road every app drives on. That's why the open-weight crowd matters here too: a model you can run on your own hardware is a hedge against someone else's silicon tax.
For builders, the practical move is routing: send the high-volume, low-stakes 80% to a cheap fast model on cheap silicon, and save the premium compute for the judgment calls. The dispatcher, not the model, is the real cost lever now.
I'm not buying stock. I'm just noticing that the interesting money stopped chasing brains and started chasing the plumbing, and the plumbing is what actually decides who can afford to ship.
Updated October 28, 2026: the raises below were announced October 19-20.
Frequently asked questions:
Q: Which inference chip startups raised money?
A: Positron AI closed $875 million at a $5 billion valuation for its Asimov inference chip, and Tenstorrent raised a $1.5 billion Series E - both about serving models, not training them.
Q: Why does inference silicon matter to users?
A: The price you pay per query is mostly inference cost. Better silicon drops per-token prices - Haiku 5.5 arrived about 90 percent cheaper than its predecessor.
Q: What's the risk of concentration?
A: If two or three chipmakers own the efficient path to serving models, they own the toll road. Open weights are the hedge - a model you can run on your own hardware.