DeepSeek is working with Huawei Technologies to design AI chips built around its models, a move that could let China's leading AI startup train advanced systems without relying on Nvidia, which remains central to training the country's most advanced models despite U.S. export restrictions.
The collaboration goes beyond a supplier relationship. According to The Information, Huawei engineers are stationed at DeepSeek's offices and fold the lab's feedback into chip designs, including how models perform calculations and move data between processors. The report cites people familiar with the partnership. The project supports Beijing's push to build an AI industry independent of American technology, a goal the government has pursued for years and which DeepSeek, with its 160,000-chip Ascend cluster in Inner Mongolia, is now turning into concrete infrastructure. (The 160,000-chip Huawei cluster in Inner Mongolia was until now aimed at inference, that is, running already-trained models.)
SSBOfficial/DeepSeek-V4-Pro · Hugging Face
Timing and chips involved
DeepSeek plans to begin using Huawei's Ascend 950DT chips for model training as soon as the end of 2026. Huawei, for its part, plans to prioritize DeepSeek for most of its next-generation Ascend 960DT chips, expected in the first quarter of 2027. The timetable is tight, and the real test will be whether manufacturing can keep up. Huawei is also trying to work around advanced lithography limits with new packaging, and production of AI chips remains constrained by SMIC's capacity.
The technical gap is still large. According to the research institute Epoch AI, Nvidia's B300 chips deliver roughly seven times the raw arithmetic performance of Huawei's Ascend 950 processors. Training a frontier model could require more than 10,000 Huawei chips operating together for weeks, which makes manufacturing capacity and system reliability critical. It is the same bottleneck that surfaced when Huawei acknowledged it still trails Nvidia on single-chip power, even while selling more AI accelerators than Nvidia in China.
TileLang and moving away from CUDA
The software side matters too. DeepSeek uses TileLang, an open-source programming language that helps developers adapt AI software across different chip architectures, reducing reliance on Nvidia's CUDA ecosystem. Last month it released a version compatible with Huawei's Ascend 950 chips. This matters because Nvidia's advantage is not only hardware: much of the AI code in the world is written for CUDA, and moving training to another platform means rewriting or adapting large parts of the software.
Deepseek Ai DeepSeek V4 Pro - a Hugging Face Space by zhudi1886
DeepSeek is not abandoning Nvidia, however, and continues to seek access to its most advanced processors, according to The Information. The shift is gradual and, for now, partial. Earlier this year DeepSeek had already returned to Nvidia GPUs for training after failures on Huawei hardware, which makes this partnership a more cautious bet.
What it means for China and the market
If the collaboration let DeepSeek train a frontier model entirely on Chinese chips, the project could become a blueprint for other AI developers in the country and a major step toward reducing China's dependence on U.S. technology. The context is one of growing tension: while Washington has barred DeepSeek products from federal agencies, China is trying to build its own chain from silicon to model.
For now, though, the project is a set of intentions more than a result. The date to watch is the end of 2026, when it will become clear whether DeepSeek can actually start training on Huawei chips.
No comments yet. Be the first!