Deepseek ships open-source TileLang tools for Huawei chips

Deepseek has been using TileLang for roughly a year and first tested it on older Nvidia chips, so the language is not new to the company. What is new is putting it behind Huawei's Ascend hardware as an open release, assembled with Huawei's support and aimed at a cluster of 128 Ascend 950 chips. The stated logic is that an independent software ecosystem for AI chips needs a universal language that is easy to program but still extracts full hardware performance; Deepseek argues TileLang fits that role better than CUDA.

The context the article supplies is that software, not silicon, has been the gap. Nvidia's position rests partly on an estimated four million CUDA developers worldwide, an ecosystem rivals like AMD have not crossed even when their hardware matched on paper. Chinese model makers such as Z.ai and Moonshot AI have moved faster than the country's chipmakers, and Huawei wants to close that gap: two weeks before Deepseek's announcement it unveiled new AI processors and supernode systems, promising wide use for model training next year. Huawei also says it cannot meet domestic demand and plans to sell fewer chips abroad, with rotating chairman Eric Xu framing the company's stance around US export controls.

Whether this release changes much depends on developers outside the two companies choosing TileLang, and on how quickly Huawei's hardware is actually used for training. The rival reading in the article is that Nvidia's advantage may be shifting rather than ending: SemiAnalysis judged the CUDA moat "potentially dead" after Jalapeño, yet still found Nvidia ahead on the harder agent workloads, and noted that Huawei's CANN stack was the only one besides CUDA to support DeepSeek V4 on day one.