
Ask a machine-learning engineer what makes NVIDIA unbeatable and they will talk about the chips. Ask the engineers who actually port models to new hardware and they will talk about something far less glamorous: kernels. Tens of thousands of small, hand-tuned programs — matrix multiplications, attention routines, quantization passes — that sit between a model and the silicon. The chips are the visible part of the moat. The kernels are the water.
On September 30, 2026, that water started draining — at least on paper. DeepSeek, the Chinese AI lab, released a full open-source software stack for Huawei's Ascend NPUs, built around TileLang, a Python-style language for writing accelerator kernels. The pitch is simple and audacious: the same Python APIs and the same kernel libraries, running on NVIDIA GPUs and Huawei NPUs alike, with the backend selected automatically at runtime.
CUDA's lock-in has never really been about who can etch the fastest transistor. It is about the mountain of software sitting on top: the kernel libraries, the profiling tools, the decade of Stack Overflow answers and battle-tested code paths. Moving a training cluster to a different chip doesn't just mean buying new hardware — it means rewriting, by hand, every performance-critical kernel in your stack, then spending months tuning each one back to parity. That is the cost that keeps buyers put.
This is why the September release is aimed at software rather than silicon. DeepSeek's framing, as reported in its announcement, is explicit: building "a new generation of independent, self-controlled GPU software ecosystems" starts with a high-level language that is universal, easy to program, and still able to reach the hardware's full performance. In other words: don't fight NVIDIA chip-for-chip. Remove the software reason to stay.
DeepSeek has run this playbook before. Its February 2025 "Open Source Week" released FlashMLA, DeepEP, and DeepGEMM — the kernel libraries behind its own models — for NVIDIA hardware. The Ascend release is the mirror image: the same libraries, ported to Huawei's stack.
TileLang is the foundation everything else stands on. Its own project page describes it as "a concise domain-specific language designed to streamline the development of high-performance GPU/CPU/NPU kernels," with a Pythonic syntax sitting on a compiler infrastructure built on TVM. Instead of thinking in individual threads, the way raw CUDA forces you to, you write code that reasons about tiles of data — blocks that move through the chip's memory hierarchy together — and the compiler handles the low-level scheduling.
Two facts about the project matter for understanding the release:
torch_npu). Its default compilation target auto-detects whichever of CUDA, HIP, Metal, or Ascend is present. The Ascend port was announced in the project's own news log on September 30, with native code generation and automatic scheduling.The headline component is TileKernels, a library of dozens of highly optimized kernels written in TileLang, published by DeepSeek under the MIT license. Per its own README, most kernels "achieve performance close to the hardware's compute or memory bandwidth limits," and — the line that carries the most weight — "all of these kernels have already been used in our internal training and inference workloads." These are not demo kernels. They are the ones DeepSeek trains on.
The library covers the operations that dominate LLM work: Mixture-of-Experts routing, FP8/FP4 quantization (including fused SwiGLU-plus-quantization), the Engram gating mechanism, manifold hyper-connections, RoPE transforms, and a high-level PyTorch autograd wrapper.
Around it sit the ported libraries, each mirroring something DeepSeek previously shipped for NVIDIA — DeepGEMM-Ascend for matrix multiplication, DeepEP-Ascend for cross-chip expert-parallel communication in MoE models, the FlashMLA attention port, plus TileKernel and DeepSelect utilities. Coverage of the release notes the design choice explicitly: 1:1 API parity with the NVIDIA versions, so switching hardware means swapping a backend flag rather than rewriting a codebase.
The practical shape of the stack, straight from the README:
| Layer | NVIDIA path | Ascend path |
|---|---|---|
| Hardware | SM90/SM100 GPUs, CUDA Toolkit 13.1+ | Ascend 950 NPU, CANN 9.2.0+ |
| Kernel language | TileLang (CUDA backend) | TileLang (Ascend backend) |
| Kernel library | TileKernels (NVIDIA) | TileKernels (Ascend, auto-selected at runtime) |
| Install | pip install tile-kernels | same, with Ascend build flags |
The "same Python APIs run on both NVIDIA GPUs and Huawei NPUs" line from the README is the whole thesis in one sentence.
Press coverage tends to frame this as a DeepSeek-versus-NVIDIA story. The Chinese developer community sees something more practical: a new shared toolchain to hack on. The tilelang-ascend repository's issue tracker is already full of detailed, Chinese-language kernel engineering — contributors porting operators like NSA (native sparse attention) forward kernels to the Ascend NPU, filing precision reports with per-case test suites, and arguing about memory-planning behavior in the compiler. Third-party forks are contributing kernels back upstream. This is what an actual working open-source hardware ecosystem looks like in its early months: messy, technical, and alive.
Two reported details round out the picture, both worth treating as reported rather than confirmed: Chinese tech briefings say Cambricon and Hygon announced chip compatibility with the new stack on the same day, which would widen the "domestic alternative" camp beyond Huawei; and Huawei is reported to have pulled its Ascend 960DT chip forward to the first quarter of 2027. Hardware and software are accelerating together.
For developers who run their own inference hardware, the direction is the interesting part. The same crowd that builds self-hosted AI gateways to route around API lock-in is watching a parallel effort to route around hardware lock-in — and the homelab hunters who treat cheap compute as a sport will recognize the appeal immediately: more backends competing means cheaper FLOPs eventually.
Enthusiasm is cheap; adoption metrics are the scoreboard. Three things to watch:
There are also honest reasons for non-Chinese organizations to stay cautious: geopolitical risk, questions about long-term support, and simple inertia all favor the incumbent. DeepSeek itself seems to anticipate this, positioning the Ascend work as a template for building portable software stacks for other AI chips — an open stance that invites contributions even from competitors.
The fairest read: nobody has replicated CUDA, and one release doesn't change that. But for the first time, the most credible challenger isn't a chip company promising hardware — it's an AI lab open-sourcing the software layer, kernel by kernel, with the receipts of its own training runs attached. Watch the repositories, not the press releases.