DeepSeek has open-sourced the software toolkit it built with Huawei Technologies to program Ascend AI accelerators, releasing it free for download in a direct challenge to Nvidia's CUDA, the de facto global standard for AI development.

The startup announced the release on its WeChat account on Wednesday. The toolkit includes TileLang, a programming platform widely described in China as the domestic answer to CUDA. DeepSeek said it has worked with Huawei to optimize performance on the Ascend 950, Huawei's most powerful accelerator.

The move signals how far the partnership between two of China's most important AI players has advanced. Huawei is leading the country's effort to build accelerators for data centers, while DeepSeek develops open-weight models that compete with Anthropic and OpenAI. Together they sit at the center of Beijing's push to replace American software and silicon with domestic alternatives.

DeepSeek is backing that strategy with hard infrastructure commitments. The company plans to deploy at least 160,000 of Huawei's top-tier accelerators in a massive data centre it is building in Inner Mongolia, Bloomberg News reported in August. It is also finalizing a fundraising round at a valuation close to 500 billion yuan (US$74 billion), ahead of a potential public listing as early as this year.

That scale matters commercially. Model training worldwide still relies heavily on Nvidia semiconductors, and US export controls have restricted China's access to the American company's most advanced chips. Every large deployment that runs on Ascend silicon instead of Nvidia hardware strengthens Huawei's position in a domestic accelerator market that Washington has effectively handed over by cutting off supply.

The software side is the harder problem. For two decades, Nvidia's advantage has rested not only on chip performance but on CUDA, the programming layer that millions of developers use to write AI workloads. Replacing hardware is one thing; replacing the ecosystem around it is another. This is what makes TileLang's open-source release strategically significant. DeepSeek is not just using Huawei chips — it is distributing, for free, the tools needed for anyone else to use them too.

DeepSeek said a sophisticated programming language with a simple coding process is essential for building an effective AI ecosystem. The company rarely discusses internal development, including AI chip integration, which makes this public disclosure itself notable.

Huawei provided "unreserved and vigorous support" during the research and development of the software, DeepSeek said, and the two companies will collaborate on further innovations.

The backdrop is a US tightening of chip controls. Nvidia has repeatedly warned that Chinese customers cannot buy its latest data-center GPUs, while Huawei has accelerated its Ascend roadmap in response. DeepSeek's adoption of Ascend hardware at scale gives Huawei a flagship customer for its most advanced parts, and the open-source release gives other developers a lower barrier to following the same path.

Questions remain about how far TileLang can close the gap with CUDA in practice, given the depth of Nvidia's tooling and the maturity of its developer base. But with 160,000 Ascend accelerators slated for deployment and fresh capital approaching a US$74 billion valuation, DeepSeek and Huawei are now combining model leadership, silicon supply and open software into a single, vertically coordinated stack — the most credible domestic challenge yet to Nvidia's grip on China's AI compute market.