A report that Huawei engineers are working inside China’s best-known AI start-up adds pressure on Santa Clara’s Nvidia, and on a Washington chip policy that has loosened this year, a month before President Trump and Xi Jinping meet in Shenzhen.
By Staff Reporter
China’s most prominent artificial intelligence start-up, DeepSeek, is working with Huawei Technologies to design AI chips around its own models, a collaboration that could eventually let the company train advanced systems without processors from Nvidia, the Santa Clara chipmaker, according to a report published Monday by The Information.
Citing people familiar with the arrangement, the report said Huawei engineers have been stationed at DeepSeek’s offices and are incorporating the start-up’s feedback into chip designs, including how AI models perform calculations and move data between processors, according to a summary by Investing.com. DeepSeek plans to begin training models on Huawei’s Ascend 950DT chips as soon as the end of this year, and Huawei intends to give DeepSeek priority for most of its next-generation Ascend 960DT chips, which the report said are expected in the first quarter of 2027.
A Chinese-language account attributed to Tencent Technology and relayed by PANews on Tuesday described the Huawei engineers’ presence as long-term and said a complete pre-training run could require more than 10,000 chips running continuously for several weeks, depending on testing. The Inquirer could not find a public statement from either company addressing the reports.
What the report says, and what it does not
The account is not a declaration of independence from American silicon. DeepSeek is not abandoning Nvidia and continues to seek access to its advanced processors, The Information reported, and Nvidia hardware remains central to training China’s most advanced models despite U.S. export restrictions.
The performance gap is also large. Citing the research group Epoch AI, the report said Nvidia’s B300 chips deliver roughly seven times the raw arithmetic performance of Huawei’s Ascend 950 processors, which is why a frontier-scale training run on Chinese hardware would depend on linking thousands of chips together reliably for weeks at a time.
What is new is the depth of the partnership. Huawei laid out its chip plans publicly at its Huawei Connect conference in September 2025, when the rotating chairman, Eric Xu, presented a roadmap that listed the Ascend 950DT, “optimized for both the decode stage of inference and for model training,” for the fourth quarter of 2026, and the Ascend 960 for the fourth quarter of 2027. If the timeline in Monday’s report holds, a version of the 960 would arrive several months earlier than that public schedule, with much of the initial supply reserved for a single customer.
The software side has been moving, too. On Sept. 30, DeepSeek released six open-source software modules built for Huawei’s Ascend chips, the South China Morning Post reported, including a version of its TileLang programming language that now officially supports the Ascend 950. Tools like TileLang matter because much of the world’s AI software is written for Nvidia’s CUDA platform, and that software ecosystem, as much as the chips themselves, has kept developers tied to the company.
Why it matters in Santa Clara
For Nvidia, China has become a market defined more by politics than by demand. In December 2025, President Trump said the government would allow sales of Nvidia’s H200 chip to approved Chinese customers in exchange for a 25 percent cut, and on Jan. 13 the Commerce Department’s Bureau of Industry and Security issued a rule moving license applications for the H200, AMD’s MI325X and similar chips to case-by-case review. Applicants must show that the exports will not reduce production capacity available to American customers, that the Chinese buyer screens its customers, and that the chips have passed independent testing in the United States. AMD, like Nvidia, is based in Santa Clara.
The opening has produced little so far. “The bottom line is very few shipments against licenses for H200s and equivalents have taken place. It’s a very small quantity of chips,” Jeffrey Kessler, the under secretary of commerce for industry and security, told a congressional hearing in July, CNBC reported. Nvidia has excluded potential Chinese AI chip revenue from its forecasts, and its chief executive, Jensen Huang, said in May that he had told investors to “expect nothing” from Chinese sales, according to the same report.
That leaves a tension at the center of U.S. policy. Supporters of tight controls argue that denying China leading-edge chips slows its military and AI ambitions; skeptics argue that restrictions push Chinese buyers toward domestic suppliers and speed the rise of a rival ecosystem. A DeepSeek model trained largely on Huawei hardware would likely be cited by both camps.
The diplomatic calendar
The report lands as Washington and Beijing negotiate ahead of a meeting between Mr. Trump and Mr. Xi on the sidelines of the Asia-Pacific Economic Cooperation summit in Shenzhen in November, their second meeting in two months, according to the South China Morning Post, which reported on Tuesday that negotiators were racing to finalize investment and trade agreements. An essay in The Diplomat this month noted that the leaders are expected to meet on Nov. 18 and 19 and that both sides have scheduled further talks on artificial intelligence before the end of November.
For California, the stakes run beyond one company’s share price. The Bay Area’s chip designers, cloud providers and AI laboratories have built their businesses on American hardware and software setting the global standard. A credible Chinese alternative, even one that lags in raw performance, would give developers in China and elsewhere a second option.
Whether DeepSeek can meet the timeline is the open question. The earliest evidence would likely come from DeepSeek itself, which has published technical papers alongside its major models.
