Chinese AI developer Deepseek has partnered with Huawei to release open-source programming tools for Huawei's Ascend chips, aiming to bolster the domestic software ecosystem. The announcement, made via Deepseek's official WeChat channel and reported by Reuters, centers on making computational libraries and data-transfer tools freely available to developers.
What Happened
The core of the release is TileLang, an open-source programming language for AI chips originally developed by researchers at Peking University. Deepseek has reportedly used TileLang for about a year, first testing it on older Nvidia chips before adapting it for Huawei's hardware. The company argues that building an independent software ecosystem requires a universal language that is easy to program while still extracting full performance from the hardware. According to Deepseek, TileLang offers a simpler programming model than Nvidia's CUDA.
In addition to the software tools, the two companies collaborated on optimizing a "supernode," a cluster consisting of 128 Ascend 950 chips. Deepseek stated that Huawei "fully supported" the work. The New York Times reports that TileLang has become Deepseek's main tool for its work on artificial general intelligence (AGI).
Why It Matters
This partnership addresses a critical bottleneck in China's AI industry: the need for software that can maximize the potential of domestic chips. Nvidia's market dominance is not solely due to chip design but also rests on an estimated four million developers worldwide who build with CUDA. This ecosystem has historically been a barrier that rivals like AMD have struggled to cross, even when their hardware appeared competitive on paper.
The move comes as Chinese model makers, such as Z.ai and Moonshot AI, have reportedly moved faster than the country's chipmakers. Huawei aims to close this gap; two weeks prior to Deepseek's announcement, the company unveiled new AI processors and supernode systems, stating they would be widely used for model training next year. Huawei's rotating chairman, Eric Xu, noted that the company plans to sell fewer chips abroad due to high domestic demand and emphasized that Huawei cannot accept a future dependent on whether other nations are willing to sell chips to China.
The broader context involves the weakening of Nvidia's software moat. Research firm SemiAnalysis recently described the CUDA moat as "potentially dead" after testing OpenAI's inference chip, Jalapeño, which outperformed Nvidia's Blackwell on performance per watt in most tested scenarios. However, SemiAnalysis added caveats, noting their tests focused on relatively easy-to-optimize scenarios and that they had not yet run AgentX, a benchmark for multistep AI agent tasks. In August, the firm found that Nvidia remained well ahead in agent workloads, citing advantages in the software that links multiple chips into a single system rather than the silicon itself.
The Bottom Line
Deepseek's open-source release for Huawei's Ascend chips marks a strategic effort to build a viable alternative to Nvidia's CUDA ecosystem in China. While Nvidia maintains a lead in complex agent workloads, Huawei's CANN software stack has already demonstrated capability by supporting DeepSeek V4 on day one, signaling growing competitiveness in the domestic AI infrastructure landscape.