DeepSeek and Huawei Open Up a Software Path Around Nvidia's CUDA

Chinese AI firm DeepSeek announced on Wednesday, September 30, 2026, via its official WeChat account a partnership with Huawei Technologies to build programming tools for Huawei's Ascend AI chips, releasing six open-source software modules aimed at creating what the company called an "independent and controllable" GPU software ecosystem, per Reuters via Quartz. The move is the latest sign that Chinese technology firms are working to build alternatives to the Nvidia software stack as U.S. export controls restrict access to American chips. Among the releases is an Ascend-compatible version of TileLang, DeepSeek's high-level open-source chip programming language, positioned as a rival to Nvidia's CUDA, which DeepSeek says offers "a simpler programming model" and brings "native code generation, automatic scheduling, and synchronization" to Ascend, according to GitHub and the South China Morning Post via The Daily Upside. The two companies also co-developed a "supernode" solution based on 128 Ascend 950 chips, optimized for both computation and communication.

What happened

The announcement landed on DeepSeek's official WeChat account on September 30, 2026, and laid out a coordinated software and hardware push. The six open-source modules include libraries handling compute and communication workloads for Huawei's Ascend platform, the infrastructure layer that developers need to run AI models on a given chip family. TileLang, the flagship of the release, is DeepSeek's own high-level programming language for chips, and the Ascend-compatible version is the most direct challenge yet to CUDA, the Nvidia software platform that has anchored the company's dominance since its launch in 2006.

The timing tracks closely with Huawei's own hardware cadence. DeepSeek's release came about two weeks after Huawei unveiled its next-generation AI processors and supernode computing systems, which the company expects to deploy for model training next year. Last week, Huawei said it would release its next-generation Ascend 960DT chip in Q1 2027, a date pulled forward by three quarters, alongside a chip roadmap that runs through 2029. DeepSeek had earlier this year previewed its V4 model adapted to run on Huawei Ascend processors, moving away from its prior reliance on Nvidia chips, and Huawei said it coordinated with DeepSeek to ensure V4 compatibility across the Ascend lineup.

The executive commentary around the release was pointed. Huawei Rotating Chairman Eric Xu claimed the company's Ascend chips hold a larger share of China's AI chip market than Nvidia, though he provided no data to support that claim, as Quartz noted. Nvidia CEO Jensen Huang said in May that Nvidia had "largely conceded China's AI-chip market to Huawei."

Why it matters

The significance of the announcement sits less in the hardware than in the software layer. CUDA, launched by Nvidia in 2006, remains the industry standard for AI development, and as The Daily Upside's Jamie Wilde wrote on October 1, 2026, "In some ways, CUDA is to Nvidia what iOS is to Apple." Prior attempts to break that lock, including AMD's ROCm, Intel's oneAPI and the open OpenCL standard, failed to gain meaningful traction. Wilde added that TileLang "doesn't seem likely to turn many heads among the millions of developers that depend on CUDA," a reminder that ecosystem gravity, not raw chip specifications, has historically decided these contests.

Still, the combination matters for the shape of the Chinese AI market. As Reuters reported, U.S. export controls have constrained Chinese firms' access to Nvidia's most advanced hardware, and a domestic software stack that lowers the cost of writing for Ascend chips addresses the main reason developers stayed on CUDA. A 128-chip supernode co-developed by DeepSeek and Huawei, plus a V4 model already adapted to Ascend, suggests the two firms are building a vertically integrated alternative: Chinese chips, Chinese software, Chinese models. If the tooling matures, the question shifts from whether Chinese firms can run models on domestic silicon to whether they can attract the developer base that makes a platform self-sustaining.

How it is being covered

According to ground.news, the story is carried by 50 sources with a left/centre/right split of 0/58/42 percent, meaning centre and right-leaning outlets dominate the coverage. Factuality ratings are strong, with 17 of 19 rated sources in the high or very high bands. The blindspot sits on the left: ground.news records 0 percent left-leaning coverage, so readers relying on left-leaning outlets may not encounter the story at all.

What to watch

Three markers will show whether the software push gains traction. First, developer adoption of the Ascend-compatible TileLang and the six open-source modules, measurable in GitHub activity and third-party model ports. Second, Huawei's execution against its pulled-forward timeline, with the Ascend 960DT due in Q1 2027 and a roadmap through 2029, and with next-generation supernode systems expected to enter model training deployment next year. Third, whether DeepSeek's V4, already adapted to Ascend, ships as a widely used model trained or served on domestic hardware, which would be the clearest proof that the "independent and controllable" ecosystem works at scale. Any data Huawei produces to substantiate Eric Xu's market-share claim, which currently stands without supporting figures, would also be worth weighing.

By the numbers

Source snapshot