The Information 深度报道：DeepSeek 与华为深度合作，将用昇腾 950DT 预训练、并拿下 960DT 大部分产能，进一步松动英伟达 CUDA 锁定

The extensive collaboration, which hasn’t previously been reported, brings two of China’s most important tech companies into a unified effort on an issue central to Beijing’s technology strategy. China’s government wants the country’s AI built on homegrown chips and has limited how many Nvidia processors China’s largest tech companies can buy. Huawei already leads the country’s AI chip market, and Nvidia CEO Jensen Huang told CNBC in May that his company had “largely conceded” the market to domestic rivals.

In particular, the cooperation could help China crack one of the biggest impediments to fully homegrown AI: the powerful chips needed to train new models quickly. Despite attempts by both the Chinese and U.S. governments to limit Chinese companies’ access to Nvidia chips, [China’s top AI developers still train their best models on chips from the U.S. company](https://www.theinformation.com/articles/chinese-ai-startup-moonshot-seeks-nvidia-blackwell-chips-next-model), sometimes obtained through smuggling. If the collaboration yields a Huawei chip that DeepSeek can use to fully train a top-tier model, other Chinese AI labs could follow suit.

Last month, DeepSeek founder Liang Wenfeng told investors at a meeting that using more Chinese chips for training was a major priority, The Information [reported](https://www.theinformation.com/articles/deepseek-bets-big-huawei-chips-bypass-u-s-export-controls?rc=jnr9wn). DeepSeek is planning to start using Huawei’s current flagship AI chip, the recently released Ascend 950DT, for pre-training—the stage where a model learns its basic abilities from vast amounts of data over weeks or months—as soon as the end of this year, according to one of the two people and a separate person with direct knowledge of the effort. Tests are still underway and the timetable could change, they said.

Huawei, meanwhile, is planning to give DeepSeek most of the supply of the Ascend 960DT, its next training chip, when it comes out next year, rather than spreading it across China’s AI developers. DeepSeek plans to use it both to train models and to run them, one of the two people said.

Huawei has said it will release a new generation of Ascend chips annually, including the 960DT in the first quarter of 2027, the 970 in 2028 and the 980 in 2029.

**Hurdles to China’s Chip Efforts**

The chip effort still faces significant obstacles, including constraints on how quickly Huawei’s manufacturing partner can ramp up production to levels needed for the large chip clusters used to train models. U.S. export controls have made that harder, with Washington blocking Huawei from using the world’s top chip manufacturer, Taiwan Semiconductor Manufacturing Co., in 2020 and limiting China’s access since 2024 to the advanced high-bandwidth memory AI chips require.

And even with DeepSeek’s help, it will likely be a while before Huawei’s chips rival the power of Nvidia’s. The U.S. company’s B300 chips deliver roughly seven times the raw arithmetic performance of Huawei’s Ascend 950, according to research institute [Epoch AI](https://epoch.ai/publications/huaweis-roadmap-to-2031), which expects Huawei to stay three to four years behind Nvidia on single-chip performance through 2030.

While Huawei’s weaker chips make models slower to build, they come from a supplier that U.S. export rules can’t control. DeepSeek is choosing to “go a little slower but go a lot longer,” said Qingyuan Lin, a senior analyst at Bernstein who covers China’s chip industry.

Only a few companies have trained frontier models on non-Nvidia chips, including Google on its own tensor processing units and Anthropic on Google TPUs and Amazon chips. Previously, DeepSeek used Huawei chips for running its models and for post-training—the stage of refining a pre-trained model for reasoning and other tasks.

DeepSeek isn’t walking away from Nvidia. It is trying to access more of Nvidia’s advanced Blackwell chips, along with older Nvidia processors, one of the two people said. Blackwell chips are barred from export to China, but DeepSeek has obtained them anyway and used them to train its V4 models, [The Information reported](https://www.theinformation.com/articles/deepseek-using-banned-nvidia-chips-race-build-next-model). Intermediaries bought Blackwell servers in countries where sales are allowed and set them up in data centers that passed inspection. The servers were then taken apart and the chips shipped to China.

**Nvidia’s Lock-In Loosens**

A key component of DeepSeek’s effort with Huawei chips is TileLang, an open-source programming language built by a team led by Peking University researchers and released in January 2025. It lets engineers write small programs known as kernel code that carry out an AI model’s calculations with less code than older methods require. TileLang is an alternative to writing code directly in CUDA, Nvidia’s widely used programming system for its graphics processing units, and can run on different chips.

DeepSeek first used TileLang with Nvidia GPUs, publishing TileLang code alongside versions written in CUDA when the Chinese company released an experimental V3.2 model in September 2025. TileLang now handles most of those programs in model development for DeepSeek’s V4 models.

Last month, DeepSeek released a version of TileLang for Huawei’s Ascend 950 chips, along with five sets of code covering the calculations and data transfers needed to run its models. It said every TileLang program it uses in training now has an Ascend version. DeepSeek said Huawei’s engineers gave “unreserved” help in building them.

TileLang could further loosen Nvidia’s grip. For years, CUDA has been one of the U.S. company’s most important tools for keeping customers using its chips, because moving to another type of chip requires a long, laborious rewrite. With TileLang, code still has to be adjusted, but DeepSeek doesn’t have to start from zero when it moves work from Nvidia to Huawei chips.

Huawei is taking other steps to make up for the weaker performance of its chips. The company is linking many processors into one system big enough to train models, and recently began shipping these supernodes with Ascend 950DT at high volume, the people said.

Huawei Rotating Chairman Eric Xu said last month that 950DT training tests had gone well, that large-scale supply would start around the turn of the year, and that he expects many developers to train on them in 2027. He also said Huawei couldn’t meet domestic demand and was limiting overseas sales.

DeepSeek’s newer tests use more than 1,000 processors, the people said. A full pre-training run could need well over 10,000 Ascend 950DTs running nonstop for weeks. One failed chip or a delay in moving data can leave the rest waiting, and repeated disruptions can derail the run. Some of the software needed to manage systems that large is still unfinished, the people said.