DeepSeek, the Hangzhou-based lab known for shipping frontier-competitive models at a fraction of Western training budgets, is moving into chip design. The company has spent roughly a year meeting hardware and silicon partners and hiring engineers for the effort, according to Ars Technica, which cited a Reuters report naming three people familiar with the matter. The target is inference chips for data centers, the hardware that runs a trained model, not the chips that build one.
That distinction is the story. DeepSeek currently runs on a mix of Nvidia chips it can still obtain despite US export controls and Huawei’s domestic alternative. Building its own inference silicon would cut dependence on both suppliers at once.
Huawei controls roughly half of China’s data center chip market, per Ars Technica’s reporting. Nvidia, the chipmaker most AI companies outside China rely on, could not replicate that reach inside China because of the US export ban. DeepSeek is not alone in trying to close the gap: Alibaba and Baidu have pursued their own silicon programs too.
US labs are making a parallel bet for different reasons. OpenAI and Broadcom announced Jalapeño, OpenAI’s first chip built for inference at scale, in June. Anthropic has explored custom chip design too, though it has not disclosed any public milestones.
For OpenAI, the motivation is partly to loosen its own dependence on Nvidia and partly a desire for the kind of vertical control over its stack that Apple has over iPhone hardware and software. For DeepSeek, the calculation is sharper. Export controls make chip access a political variable, not just a market one, and the company has few other levers to pull.
Owning silicon at the data center level could matter most when capacity itself is scarce. Every AI company chasing scale, from DeepSeek to OpenAI, is competing for the same limited compute, and controlling a piece of the stack that others must buy or rent is a hedge against getting squeezed out.
AI Insiders covered DeepSeek’s ambition to help build a 10 trillion dollar Chinese AI hardware ecosystem in May. This is the concrete follow-through on that thesis: a defined chip category, a year of relationship-building already logged, and named competitors racing toward the same target.
None of this is close to shipping. Reuters’ sources described planning and hiring, not taped-out silicon or a production date, and chip design cycles typically run two to three years from team formation to a working part. DeepSeek’s effort is also unproven against Huawei’s Ascend line, which already ships at volume domestically.
For operators tracking China’s AI supply chain, the signal is not a chip on the market but a widening field of buyers stepping away from Nvidia by necessity. Expect Huawei, Alibaba, Baidu, and now DeepSeek to keep splitting China’s inference capacity among themselves while Nvidia’s addressable share there keeps shrinking.
Ars Technica reported this story on July 7, 2026, citing Reuters, which cited three people familiar with the matter.
