DeepSeek made its name on models that matched Western labs at a fraction of the cost. Now the Hangzhou company is going after the layer beneath them: it is designing its own AI chip, three people familiar with the matter told Reuters, aiming to cut its dependence on both Nvidia and Huawei.
The chip
The reported chip is meant for inference — the stage where a trained model answers user queries — rather than for training. It would be fabricated by SMIC, China's largest foundry, on a 7-nanometre process. That focus is deliberate: inference is now expected to drive roughly 70% of AI compute demand, and it is where a purpose-built chip pays off and where China's export-limited silicon is already closest to competitive.
The export-control squeeze
SMIC has been cut off from the most advanced chipmaking tools by US and Dutch export controls, and is widely reported to be stuck on a 7nm node several generations behind Taiwan's TSMC. DeepSeek, meanwhile, faces limits on buying Nvidia's top accelerators and is wary of leaning on Huawei's Ascend line. A homegrown inference chip is a hedge against both constraints at once.
A quiet buildout
The effort began about a year ago, according to one source. DeepSeek has quietly increased hiring of chip-design engineers in recent months without public job postings, and is holding discussions with external foundry and memory partners. The company has not confirmed the project, which remains in early stages.
The stakes
If it works, DeepSeek would become the first Chinese frontier-model lab to own its inference silicon — a threat to Nvidia's remaining China business and to Huawei's pitch as the default domestic alternative. It also extends a broader pattern: after Google's TPUs, Amazon's Trainium and OpenAI's custom-silicon push, model makers increasingly want to control the chips their models run on. The limits are real, too — a 7nm process caps performance-per-watt, and yields on a first design are rarely kind.