A new open-source toolkit for Huawei's Ascend chips offers Chinese developers an alternative to Nvidia's CUDA framework
Nvidia’s biggest advantage in AI has never been only about chips. It has also been about the software developers use to program them.
On September 30, 2026, Chinese AI startup DeepSeek and Huawei Technologies went after that software layer directly. They unveiled a suite of open-source programming tools built specifically for Huawei’s Ascend AI chips. Bloomberg reports that the pair are chipping away at Nvidia’s competitive edge.
The centerpiece is TileLang, a high-level programming language pitched as a simpler alternative to Nvidia’s CUDA framework.
What DeepSeek and Huawei actually released #
The software package includes six modules. Together they cover compute, communication and programming capabilities for Ascend hardware.
TileLang is designed to make development easier than it is with CUDA. It is also built to extract full performance from the underlying hardware.
The hardware commitment behind the software is substantial. DeepSeek is preparing to deploy at least 160,000 Huawei Ascend 950DT chips in a new data center in Inner Mongolia.
The Ascend 950DT is set to launch in Q4 2026. The research also indicates the chips will be partially online by late 2027 or early 2028.
DeepSeek also plans to move further away from Nvidia hardware for training its models.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
How DeepSeek got here #
This announcement builds on work already in motion. DeepSeek previously optimized its V4 AI model for Huawei’s Ascend architecture.
Huawei’s Ascend 950 chips were used in parts of V4’s training. That made V4 something of a test run for the hardware transition DeepSeek is now formalizing with tooling and infrastructure.
Nvidia has noticed. CEO Jensen Huang has voiced strong concerns about what the DeepSeek and Huawei partnership means for US interests, and the company views the collaboration as a threatening competitive move.
What this means for Nvidia, Huawei and developers #
The real contest here is over developer habits. Hardware can be swapped out on a purchase order. Years of code, tutorials and engineering muscle memory cannot.
TileLang is a direct attempt to lower that switching cost. If developers can write code more easily and still get full performance from Ascend chips, one of the strongest reasons to stay on Nvidia hardware weakens, at least inside China.
Releasing the tools as open source is also a deliberate choice. It invites outside developers to build on, test and improve the stack.
For DeepSeek, the bet ties its future model training to domestic supply. That could reduce exposure to export controls, though it also concentrates dependence on a single hardware partner. There are clear questions still to answer. The most important is whether TileLang can match CUDA on real-world performance benchmarks, not just on ease of use.
Timing is the third. With the Ascend 950DT launching in Q4 2026 and the Inner Mongolia deployment expected to come partially online by late 2027 or early 2028, the full picture of how this ecosystem performs will take time to emerge.
Investors and industry watchers should track developer engagement with TileLang, adoption rates for Ascend chips, and any shifts in market share between Nvidia and Huawei in China.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our