DeepSeek ties up with Huawei on chip tools, sidestepping Nvidia

Key Points

1. DeepSeek is open-sourcing programming infrastructure built for Huawei's Ascend AI chip platform.
2. The two firms jointly advanced a supernode system running on 128 Ascend 950 chips.
3. The move deepens China's push toward a software stack independent of Nvidia's CUDA.

The latest:

Chinese AI firm DeepSeek said it is open-sourcing programming infrastructure optimised for Huawei’s Ascend chips, including compute and communication libraries, in a partnership announced Wednesday on its official WeChat account. DeepSeek said Huawei provided full support in building the toolset. The two companies jointly advanced a supernode solution based on 128 Ascend 950 chips, tuned for both computation and communication.

Details:

• The announcement: DeepSeek disclosed the partnership in a post on its official WeChat account on Wednesday, according to Reuters. The release covers programming infrastructure for Huawei’s Ascend platform, spanning compute and communication libraries, and is being made open-source rather than kept proprietary to the two companies.
• The hardware: DeepSeek said the two firms jointly advanced a supernode solution built on 128 Ascend 950 chips, with optimisation applied to both computation and communication layers. Huawei provided full support in developing the programming infrastructure, according to DeepSeek’s statement.
• The software layer: DeepSeek pointed to TileLang, a high-level open-source programming language for AI chips, as central to the work. The company said it improves development efficiency and simplifies code logic, and described it as offering a simpler programming model than Nvidia’s CUDA.
• The stated goal: DeepSeek framed the effort as building a new generation of independent, self-controlled GPU software ecosystems. It said the first priority is establishing a high-level language that is universal, easy to program, and still able to reach the hardware’s full performance potential — a need it said “TileLang was created precisely to meet.”
• The timing: The announcement landed two weeks after Huawei unveiled its next generation of AI processors and supernode computing systems. Huawei said at the time it expected its AI systems to be widely used for model training next year, according to Reuters.
• The Nvidia angle: Reuters cast the partnership as the latest sign of Chinese technology firms deepening ties as they seek alternatives to the Nvidia ecosystem. Nvidia’s dominance rests not only on its chips but on CUDA, the programming layer most AI developers have built against for years.
• What was not specified: DeepSeek did not set out a timeline for wider deployment of the toolset, name other companies or partners involved, or disclose commercial terms of its arrangement with Huawei. No performance benchmarks comparing the 128-chip supernode with rival systems were published alongside the announcement.

Background:

Huawei has been under US export restrictions for years, pushing it to develop the Ascend line as a domestic alternative to Nvidia accelerators. DeepSeek drew global attention for training competitive models at low reported cost, making its choice of hardware platform closely watched across the AI industry.

Between the lines:

The harder barrier to leaving Nvidia has never been silicon but software: CUDA is where developer habits live. By open-sourcing the Ascend toolchain and promoting TileLang as a simpler model than CUDA, DeepSeek is targeting that switching cost directly rather than competing on chip specifications. Pairing the release with Huawei’s expectation of wide training use next year points to an attempt to assemble a full domestic stack.

What’s next

Watch whether other Chinese AI developers adopt TileLang and the Ascend libraries, whether Huawei publishes training benchmarks for the Ascend 950 supernode, and whether its stated goal of wide use in model training next year holds.