传DeepSeek与华为深度共研昇腾AI芯片，规划950T预训练、960DT主供DeepSeek以替代英伟达

DeepSeek is reportedly working closely with Huawei to influence the design of its future AI chips, a link reportedly closer than typical supplier-customer relationships. The collaboration shows the growing ties between two of China's largest technology firms to sidestep Nvidia.

According to The Information , Huawei engineers have been embedded at DeepSeek's offices over stretches of time to respond to the AI company's design suggestions. The feedback, which includes model calculation and data processing methods, will be incorporated into chips set to launch at the beginning of next year.
DeepSeek founder Liang Wenfeng said the company will aim to use more Chinese chips for training and told investors that the goal is to rely only on Huawei chips for this step in the future. The AI startup plans to use Huawei's recently released Ascend 950T for pre-training, possibly starting at the end of this year, The Information reported.
As for Huawei, it plans to sell most of the Ascend 960DTs it produces next year to DeepSeek. The AI developer intends to use them for both model training and inference.
Currently, DeepSeek relies on Huawei chips primarily for inference, while the harder training stage is still mainly done using Nvidia chips. Given US restrictions on its most advanced processors to China, these are often smuggled into the country or accessed outside it through data center services.
Nevertheless, Nvidia's market share in China has dropped by roughly a third in the first half of this year . Huawei is now closely behind it, and rotating chairman Eric Xu has even said its Ascend accelerators have surpassed Nvidia in sales in China. The Chinese government restricts imports of Nvidia chips in a bid to encourage domestic development of homegrown advanced chips.
DeepSeek's contributions to TileLang , an open-source programming language that can be used on different chips, may also help chip away at Nvidia's dominance. TileLang was developed by Peking University researchers, but DeepSeek has since become a significant user and contributor to it.
In September, the AI developer released Ascend support for the language. TileLang's versatility across chips marks a major difference from Nvidia's own programming language, CUDA, which is tied to its own chips. This has long driven sales, as shifting to another chip would require rewriting a large amount of code.
However, China is not yet self-reliant on the most advanced semiconductor technology, and Huawei still lacks the production capacity to satisfy domestic demand . It is using scale to overcome its weaker single-chip performance, forming AI supernodes of its Ascend chips. It has so far deployed over 1000 Ascend 910C supernodes, and it has designed architectures including as many as 1 million NPUs.