DeepSeek is developing its own AI chip, according to a Reuters report Tuesday citing people familiar with the matter. The chip is aimed at inference — the stage where a trained model answers a user’s prompt — rather than the far heavier job of training new models. The effort is said to have run about a year and is still early, with the Hangzhou company expanding a chip-design team and leaning on outside semiconductor partners to get there.
The reflex read is “another crack in Nvidia’s China business,” and it’s fair as far as it goes: US export controls already bar DeepSeek from buying Nvidia’s best accelerators, and it has moved much of its newer work onto Huawei’s Ascend line after starting out on Nvidia’s H800. But look at the second name on that list. The chip DeepSeek is designing would route around Huawei too — the domestic supplier Beijing has spent two years pushing its champions toward. When the country’s most celebrated AI startup decides it would rather build silicon than keep buying a fellow national champion’s, that is not only an Nvidia story.
And it is pointed at the profitable half of AI. Training a model is a one-time capital burn; inference is the meter that runs every time someone hits enter, and at DeepSeek’s user scale that compute bill is a permanent line item, not a launch cost. Owning the inference chip means owning the margin. The report landed on an already-nervous tape: a gauge of semiconductor stocks fell about 6% Tuesday, with Micron off roughly 5% and equipment names like KLA, Lam Research and Applied Materials down more than 4%, as Samsung’s results and the DeepSeek headline together revived the question of whether the AI-hardware trade has sprinted ahead of the fundamentals.
Our take: The market filed this under Nvidia. It reads better as vertical integration — and it rhymes with the top of the stack, where Anthropic just signed a 20-year, $19 billion lease to control its own power and compute. Every serious lab is trying to own the thing that gates it. DeepSeek’s gate is chips it’s either barred from (Nvidia) or renting from a rival (Huawei); building its own is the logical end of that squeeze. The number that matters isn’t Tuesday’s 6% — it’s whether a first DeepSeek chip has actually taped out and yields at volume, which the reporting does not claim. Design teams are cheap. Working silicon at scale is the hard part, and it’s where every ambition on this list still has to prove itself.
What to watch
- The fab question. A homegrown inference chip still has to be manufactured somewhere, and at the leading edge yield — not design — is the wall. No manufacturing partner has been confirmed; the first named foundry will tell you how realistic the timeline is.
- Huawei’s bind. DeepSeek is one of Ascend’s marquee domestic customers. If China’s AI darling defects to its own silicon, Huawei’s pitch to the rest of the country’s developers gets harder — the same week Samsung’s record memory profit couldn’t hold its stock up.
- The inference-first pattern. DeepSeek isn’t chasing Nvidia’s training crown; it’s aiming at the cheaper, higher-volume half of the market — the lane Huawei, Alibaba and Baidu are already crowding, and the one the storage and memory suppliers are betting on too.