DeepSeek, the Chinese artificial-intelligence startup whose low-cost language models rattled Silicon Valley last year, is designing its own data-center chip for AI inference and has spent about a year lining up hardware partners and hiring engineers for the effort, according to a Reuters report cited Tuesday by Ars Technica and Semafor.
The move, if it delivers working silicon, would make DeepSeek the latest Chinese technology company to answer Beijing's push to wean domestic AI off American components, while widening a homegrown challenge to Huawei, which controls roughly half of China's data-center chip market. It also mirrors an in-house chip push under way at OpenAI and Anthropic, which have grown wary of Nvidia's grip on AI compute.
What's in the plan
Ars Technica, citing three "people familiar with the matter," said DeepSeek is focused on chips for inference — the workload of running a trained AI model — rather than training. The company has been meeting with potential silicon and hardware partners and hiring engineers for the effort for about a year, Reuters reported.
Why now
U.S. export controls have steadily eroded Nvidia's presence in China, Semafor said, leaving Huawei's chip line as the dominant homegrown option in a data-center market it now controls at about 50 percent. Alibaba and Baidu have also stepped up in-house chip designs, part of what Semafor described as a collective response to Beijing's directive to minimize dependence on American technology.
Not just China
OpenAI and Broadcom announced Jalapeño, OpenAI's first chip designed for inference at scale, "just a couple of weeks ago," Ars Technica said. Anthropic has been exploring its own silicon design, though it has not disclosed publicly visible milestones. Ars Technica said full-stack control — from data center down to chip — is becoming a competitive edge in a market where compute is likely to remain scarce and multiple companies are fighting for it.
DeepSeek has not confirmed the effort publicly, and Reuters's account did not identify a foundry partner, a tape-out timeline or a performance target against existing Nvidia or Huawei parts. Chinese officials have said nothing about the project, and none of Tuesday's reporting drew on sources outside Reuters's three "people familiar with the matter."
Reuters said DeepSeek's silicon push has been under way for about a year — a timeline that, if it holds, would put the company's first working samples on a path to surface within the next 12 months.