Training was the story. Inference became the business.

Where the money actually went

Between 22 January and 18 August 2026, twelve disclosed AI chip funding rounds raised a combined 5.37 billion dollars. Inference-focused companies accounted for eight of those twelve deals and roughly 67 percent of the capital. The largest single round was SambaNova's 1 billion dollar late-stage raise led by General Atlantic, at an 11 billion dollar valuation, for custom silicon and cloud services aimed squarely at inference.

That distribution is a reversal. For most of the last decade the money and the mythology both sat with training, because training was the visible act of creation and the thing that consumed entire data centres for months. Serving the finished model was treated as a downstream detail, the boring half of the pipeline.

Training is a project. Inference is a utility.

The economics separate cleanly. A training run is bursty and finite, closer to a capital project: you build the thing, then you stop. Inference is a per-request cost that never stops, scales with adoption rather than ambition, and gets charged to somebody every time a user presses enter. One is a construction cost; the other is an operating cost that runs forever.

Industries reorganise around the operating cost eventually, because that is where the compounding is. The telephone business spent its early decades on the capital problem of building exchanges and stringing copper, and the rest of its history on the metering problem of the calls that ran across them. The exchange was the achievement. The calls were the business.

The foundries and the laptops agree

The manufacturing side reflects the same demand. TSMC, the world's largest contract chipmaker, reported August revenue up more than 53 percent to a record. Foundry revenue is a lagging but honest indicator: it counts wafers that somebody has already committed to buying rather than intentions announced on stage.

The client side is moving in step. Intel's Panther Lake, shipping as the Core Ultra Series 3, is a laptop part built to run models locally. Inference on the device is the same shift viewed from the other end, where the request never reaches a data centre at all and the marginal cost falls to the battery.

Compute keeps moving toward the work

This is the oldest rhythm in the industry. Computing began centralised because the machine was scarce and expensive, and every time the cost fell far enough, the compute migrated toward wherever the work actually happened. Mainframe to minicomputer, minicomputer to workstation, workstation to personal computer. Then the pendulum swung back to the cloud when the network got cheap and the scale advantages returned.

Inference silicon is the next pass of that same cycle. The expensive central act, training, stays centralised because it genuinely needs the scale. The frequent, latency-sensitive act, serving, spreads out toward the user, which is why the capital is following it. Nothing about this is unprecedented. It is just the part of the story that happens after the interesting invention.