Chinese AI startup DeepSeek is working closely with Huawei to develop software tools and model adaptations for Huawei's Ascend chips, a move aimed at reducing its dependence on Nvidia's graphics processors for training and running large artificial intelligence systems.
The partnership marks a significant step in China's effort to build a more self sufficient AI ecosystem amid US export controls on advanced chips.
DeepSeek's latest models have been optimised for Huawei hardware and the company is splanning large-scale deployments of Ascend processors in new data centers.
DeepSeek gained global attention last year for its low cost AI models and has increasingly aligned its development strategy with Huawei's domestic chip architecture. In April, the company released its V4 model adapted for Huawei's Ascend chips and Huawei said its processors were used in part of the training process.
Huawei and DeepSeek coordinated closely to maximise performance on the Ascend 950 and Ascend 910c chips, something different to that with prior DeepSeek model releases. The collaboration included joint work on programming tools,model optimisation and software frameworks that allow DeepSeek's systems to run more efficiently on Huawei hardware.
A Huawei led team used about 1,000 Ascend 910c chips to fully post-train DeepSeek's 1.6-trillion parameter V4-Pro model. DeepSeek is also building a large Huawei based cluster for inference work in Inner Mongolia, with plans to deploy at least 160,000 Ascend chips at a data centre in Ulanqab.
The shift does not mean DeepSeek has completely replaced Nvidia. The company still relies on Nvidia's advanced chips for some frontier pre-training work and it continues to use Nvidia processors for new model training where the highest performance is required.
However, DeepSeek's strategy shows a clear direction to use Huawei chips for inference and increasingly for training, while maintaining access to Nvidia hardware where necessary. The company has also discussed developing its own AI chip, but recent reports indicate it has focused instead on deepening its partnership with Huawei.
Huawei benefits from the collaboration by gaining real world feedback from a leading AI developer. This allows the company to customise its hardware and software to better support large models, improve programming tools and strengthen its position as China's main alternative to Nvidia.
DeepSeek's founder, Liang Wenfeng, has estimated that training a model comparable to the largest systems developed by OpenAI would require abbout 50,000 Nvidia GB300 chips or around 200,000 Huawei Ascend 950 chips. The difference highlights both the scale of Huawei's planned deployments and the performance gap that still exists.
The partnership comes as the Chinese government pushes domestic AI firms to reduce reliance on foreign technology. US export controls have limited Chinese companies' access to Nvidia's most advanced processors, creating strong incentives to develop local alternatives.
Huawei is expected to supply additional training chips to DeepSeek in the fourth quarter of this year or the first quarter of next year. DeepSeek plans to expand its Huawei based data centre capacity and continue optimising its models for Ascend architectures.
Full frontier pre training is still expected to rely on Nvidia for the near term, but dedicated Huawei training chips are due by early 2027. The pace of replacement will depend on hardware performance, software maturity and the availability of advanced manufacturing capacity.
DeepSeek’s partnership with Huawei does not end its reliance on Nvidia overnight, but it marks a clear strategic shift toward domestic chips and software. The success of this effort will influence not only DeepSeek’s future but also the broader balance of power in the global AI hardware market.
