What Was Released
On 30 September 2026, DeepSeek open-sourced programming infrastructure for Huawei's Ascend platform, developed with Huawei. Three components:
- TileLang - a programming language positioned as an alternative to Nvidia's CUDA, described as offering a simpler programming model
- DeepGEMM Ascend - a matrix multiplication kernel library for high-throughput operations
- DeepEP Ascend - efficient large-scale communication across multiple devices
The tools target Ascend 950 chips. DeepSeek is separately planning to deploy over 160,000 Huawei Ascend chips at a new data centre in Inner Mongolia.
Why This Is About Software, Not Chips
Nvidia's position has never rested mainly on having the fastest silicon. It rests on CUDA - roughly two decades of libraries, kernels, tooling and accumulated engineer familiarity. Competing chips have existed for years. Competing software ecosystems have not.
That is what makes this release more significant than another accelerator announcement. TileLang is an attempt at the layer that actually holds people in place, and DeepGEMM and DeepEP are the two pieces you need immediately after a language: fast matrix multiplication and multi-device communication. Between them, that is most of what training a large model requires from the stack.
Whether it works is a different question, and nobody can answer it yet.
No Benchmarks. At All.
DeepSeek says TileLang enables reaching the hardware's full performance potential. It has published no comparative figures against CUDA - no throughput numbers, no training time, no utilisation percentages.
For a release whose entire claim is that it is a viable CUDA alternative, that absence is the most important thing about it today. A simpler programming model that runs slower is a research project. A simpler programming model that matches CUDA on Ascend hardware changes the economics of Chinese AI training.
Until someone publishes numbers, treat this as a serious statement of intent rather than a demonstrated result. That is not scepticism about the engineering - DeepSeek has shipped genuinely strong work before. It is just what the available evidence supports.
The 160,000 Chips Are the Commitment
The Inner Mongolia deployment is the part that is not a claim. Ordering more than 160,000 Ascend accelerators is a capital decision that assumes this software stack will work, made by the company writing it.
Firms adopt their own tools all the time. But at that scale, the toolkit is not a side project - it is infrastructure DeepSeek itself depends on, which is a stronger signal than any announcement.
Two Attacks on the Same Moat, One Day Apart
This landed alongside AMD paying $8.2 billion for World Labs, partly to get a frontier model team developing on its ROCm stack.
Two very different companies, same week, same target - and neither is trying to out-engineer Nvidia's hardware. AMD is buying ecosystem credibility. DeepSeek and Huawei are building it in the open and giving it away.
The open-source route is slower and has a better track record.
What It Means For You
If you train outside China on Nvidia hardware, nothing changes this year. CUDA's advantage is cumulative and a first release does not dent it.
What is worth watching is whether anyone outside DeepSeek adopts TileLang, and whether independent benchmarks appear. An ecosystem is defined by its second and third users, not its first.
FAQ
What is TileLang?
A programming language released by DeepSeek for Huawei Ascend accelerators, positioned as a simpler alternative to Nvidia's CUDA.
What else was released?
DeepGEMM Ascend, a matrix multiplication kernel library, and DeepEP Ascend, for large-scale multi-device communication.
Which chips does it support?
Huawei Ascend, with optimisation specifically noted for the Ascend 950.
Is it faster than CUDA?
Unknown. DeepSeek has published no comparative benchmarks, only a claim that the tools allow reaching the hardware's full performance potential.
Is it free?
The tools have been released openly. A specific licence has not been confirmed in reporting.