Chinese AI Industry Unites as DeepSeek Releases Open-Source Software for Huawei Ascend Chips

Meta description: Deepseek and Huawei use TileLang and a 128 Ascend 950 supernode to build AI chip software amid CUDA ecosystem pressure and agent workload tests.
Tags: AI chip software, Deepseek, Huawei, TileLang, CUDA
Featured image: an abstract technology-style illustration with no real people, no company logos, and no text.

Deepseek and Huawei are trying to turn AI chip software ecosystem work into a practical path for Chinese accelerators. According to Deepseek, Huawei "fully supported" the work, and the two companies optimized a so-called supernode, a cluster of 128 Ascend 950 chips. The effort centers on TileLang, an open-source programming language for AI chips that Deepseek says can serve as a universal programming layer for hardware that must be programmed simply while still delivering full performance.

Background: CUDA's Developer Base and China's Hardware Gap

The problem is not only silicon. Nvidia's dominance has long rested on CUDA, its software platform for graphics and AI processors, and on an estimated four million developers worldwide who build with CUDA. That developer base has formed a moat that rivals such as AMD have not crossed, even when their hardware looked competitive on paper.

That dynamic is especially important for China's AI industry, where domestic chips need software that can get the most out of them. The New York Times reported that Chinese model makers like Z.ai and Moonshot AI have moved faster than the country's chipmakers. Huawei wants to close that gap. Two weeks before Deepseek's announcement, Huawei unveiled new AI processors and supernode systems and said they would be widely used for model training next year.

The partnership also reflects supply constraints and policy pressure. Huawei has admitted it cannot keep up with demand at home, so it plans to sell fewer chips abroad. Referring to US export controls, Huawei's current rotating chairman Eric Xu said the company cannot accept a future that hinges on whether others are willing to sell chips to China.

Development: TileLang, Huawei, and the 128-Chip Supernode

TileLang is at the core of the release. It was originally developed by researchers at Peking University, and Deepseek has been using it for about a year. The company first tested the language on older Nvidia chips, then moved it into its core workflow. TileLang is now Deepseek's main tool for its work on artificial general intelligence (AGI), The New York Times reported.

Deepseek's argument is that an independent software ecosystem for AI chips must begin with a language that is easy to program but still extracts full performance from hardware. In the company's view, TileLang offers a simpler programming model than CUDA. The collaboration with Huawei pushes that idea onto Ascend hardware. According to Deepseek, Huawei "fully supported" the work, and the two companies optimized a supernode made of 128 Ascend 950 chips.

Latest Progress: Competing Inference Chips and Agent Benchmarks

The CUDA moat may be weakening, but it has not disappeared. Research firm SemiAnalysis tested Jalapeño, OpenAI's inference chip, and called the CUDA moat "potentially dead" because OpenAI gets new models running on its own hardware so quickly. SemiAnalysis said Jalapeño beat Nvidia's Blackwell on performance per watt in most scenarios tested. It also reported that OpenAI models helped design the chip and that those models run on Nvidia GPUs.

The analysts added caveats. Their tests covered scenarios that are relatively easy to optimize, with about 8,000 input tokens and 1,000 output tokens. They have not yet run AgentX, a benchmark measuring how AI agents handle multistep tasks. That gap matters. In August, SemiAnalysis found Nvidia well ahead on AgentX-type workloads. With AMD's current software stack, Nvidia would still come out cheaper per token even if AMD gave its hardware away. The authors said Nvidia's lasting advantage is not the silicon itself, but the software that links many chips into one system.

Huawei's chips were not part of the AgentX comparison. In an earlier analysis of DeepSeek V4, however, SemiAnalysis pointed out that Huawei's CANN software stack was the only one besides CUDA to support the model on day one.

Current Stage: Software, Not Silicon, Is the Deciding Layer

The timeline now points to a phase where hardware announcements are only half the story. TileLang's move from older Nvidia testing to AGI work, Huawei's 128 Ascend 950 supernode, and the planned wider use of new processors for model training next year all show an attempt to build software around domestic accelerators. At the same time, SemiAnalysis's Jalapeño results and AgentX caveats show that performance per watt can improve quickly, while agent workloads and chip-linking software remain harder to replace.

What to watch is whether TileLang and CANN can attract enough developers to compete with CUDA's four-million developer base, whether Ascend supernodes can support training next year at scale, and whether future AgentX comparisons include Huawei hardware. For now, the race is being decided by software layers, developer adoption, and benchmark gaps as much as by chip design.