DeepSeek Open-Sources TileLang Software for Huawei Ascend Chips

DeepSeek open-sourced TileLang programming tools for Huawei Ascend chips, challenging Nvidia's CUDA ecosystem.

DeepSeek announced on its official WeChat channel that it has made six software modules available as open source for Huawei's Ascend AI chip platform, including libraries handling compute and chip-to-chip communication. The centerpiece is TileLang, DeepSeek's high-level, open-source programming language for AI chips, now adapted to run on Huawei's Ascend 950 accelerators. According to the South China Morning Post via Yahoo Tech, the Ascend release brings native code generation, automatic scheduling, and synchronization. Huawei, DeepSeek said, fully supported the work, and the two companies also jointly optimized a "supernode" cluster built from 128 Ascend 950 chips.

TileLang began as a research project at Peking University, and The Decoder reports that DeepSeek has been using it internally for about a year, first testing it on older Nvidia chips; The New York Times reports it is now DeepSeek's main tool for its AGI work. DeepSeek frames it as a rival to Nvidia's CUDA, arguing it offers "a simpler programming model" while still reaching the hardware's full performance. In the company's words, building "a new generation of independent, self-controlled GPU software ecosystems" starts with a universal, easy-to-program high-level language.

The surprising part for developers is where Nvidia's dominance actually lives: not in silicon, but in an estimated four million developers who build with CUDA — an ecosystem AMD has not crossed even when its hardware looked strong on paper. The Decoder reports that research firm SemiAnalysis has called the CUDA moat "potentially dead" after testing OpenAI's Jalapeño inference chip, which beat Nvidia's Blackwell on performance per watt in most tested scenarios — though the analysts only tested relatively easy scenarios with about 8,000 input tokens and 1,000 output tokens, and have not yet run the agent-focused AgentX benchmark.

There is concrete sign Huawei's software story is maturing. The Decoder notes that in an earlier analysis of DeepSeek V4, SemiAnalysis found Huawei's CANN software stack was the only one besides CUDA to support the model on day one — and DeepSeek itself previewed V4 running on Ascend processors, a departure from its prior Nvidia reliance. The supernode work on 128 Ascend 950 chips, optimizing both computation and inter-chip communication, shows the two companies tuning whole systems, not just kernels, per Reuters reporting carried by Devdiscourse and Yahoo Tech.

Huawei has a demand problem as much as a software problem: the company admits it cannot keep up with demand at home and plans to sell fewer chips abroad, according to The Decoder. Rotating chairman Eric Xu said Huawei cannot accept a future that hinges on whether others are willing to sell chips to China — a pointed reference to US export controls that block Chinese firms from buying Nvidia's most advanced processors. Per Yahoo Tech, Huawei claims its Ascend chips hold a larger share of China's AI chip market than Nvidia's, though it offered no supporting data.

The release lands two weeks after Huawei unveiled its next-generation AI processors and supernode systems, which it expects to see wide use for model training next year, per Reuters reporting in the pack. The hardware roadmap is accelerating too: Yahoo Tech reports Huawei plans to ship its next-generation Ascend 960DT chip in the first quarter of 2027, pulled forward by three quarters, with a chip roadmap laid out through 2029. For working programmers, the open question is whether TileLang's simpler model can attract the developer community that CUDA's four-million-strong ecosystem took years to build.