DeepSeek, the Chinese AI company behind large language models that compete with systems from OpenAI and Anthropic, is preparing to build its own chips, according to Reuters.
Reuters, citing three people familiar with the work, reported that DeepSeek has been developing the plan for about a year. The company has met with possible hardware and silicon partners and has been hiring engineers for the project, Reuters said.
The reported target is narrow and practical: data center chips for inference, rather than chips meant to train frontier models. Training is the expensive process of building a model from large datasets. Inference is what happens afterward, when the model answers prompts, generates text, or handles other user requests. For a company running large-scale AI services, inference becomes a permanent operating cost, not a one-time lab bill.
DeepSeek’s apparent aim is to reduce its dependence on both Nvidia and Huawei, according to the report. That is a familiar pressure point in AI, but the Chinese market makes it sharper. Nvidia supplies much of the accelerator hardware used by AI companies in North America and Europe. In China, US export controls have blocked Nvidia from building the same kind of position.
Huawei has filled much of that space. It controls about half of China’s data center chip market, according to the reporting. DeepSeek is not alone in looking for a way around that bottleneck: Alibaba and Baidu have also been moving into AI silicon.
Export controls are forcing stack decisions
US chip restrictions are a direct reason DeepSeek’s reported project has become urgent. Those controls limit Chinese companies’ access to the most capable Nvidia accelerators, which are the default hardware for much of the global AI industry. If Reuters’ reporting is accurate, DeepSeek is trying to claw back some control over the hardware layer rather than waiting for Washington, Nvidia, or Huawei to solve its supply problem.
That does not mean DeepSeek is suddenly a chip company in the Nvidia sense. Designing a useful AI accelerator is hard enough. Producing it at scale, integrating it into data centers, and making software run well on it are separate fights. Reuters’ account points to planning, hiring, and partner talks, not a finished chip or a deployment schedule.
US AI companies are making similar bets for different reasons. OpenAI and Broadcom recently announced Jalapeño, described as OpenAI’s first chip for large-scale inference. Anthropic has also been exploring custom chip design, though there have been no public milestones reported.
For OpenAI, reducing reliance on Nvidia is only part of the pitch. Control over more of the stack, from model software down through silicon and data center capacity, can matter when compute remains scarce and expensive. The same logic applies in China, with the added complication that export policy can decide which chips are even available.
DeepSeek’s reported move shows how AI competition is pushing model developers toward hardware. Once inference demand is large enough, renting someone else’s chips becomes a strategic dependency. Companies with enough scale start asking whether they can own more of the machine. DeepSeek now appears to be asking that question under unusually tight constraints.
This story draws on original reporting from Ars Technica.
