The biggest lock-in was never the chips
When people talk about NVIDIA's grip on AI, they usually mean the chips. But the deeper moat is software: CUDA, the programming framework AI developers have used for over a decade. Switching away from NVIDIA hardware means rewriting mountains of CUDA code — a migration cost so high that even China, with its own chips, stayed dependent on NVIDIA.
On September 30, 2026, DeepSeek attacked that dependency directly. The Chinese AI lab announced it was open-sourcing a complete toolkit for programming Huawei's Ascend AI accelerators, described as "fully supported" by Huawei. This is the most concrete step yet toward a Chinese AI stack that does not need American hardware or software.
What is actually in the toolkit
The release ports to Ascend the same internal software DeepSeek built for NVIDIA hardware, centered on TileLang — a tile-based programming language for writing high-performance AI kernels that Bloomberg has called China's answer to CUDA. TileLang had already been used in many of the operators that trained DeepSeek's V4 models; what is new is the Ascend edition.
Around it come five libraries rebuilt for Huawei silicon: DeepGEMM for optimized matrix multiplication (the core operation of neural networks), DeepEP for communication between chips at cluster scale, TileKernels for standard vector operations, FlashMLA for long-context attention, and DeepSelect for data filtering. The two companies also tuned a "supernode" cluster of 128 Ascend 950 chips, signaling this is meant for serious training, not just inference.
The killer detail: DeepSeek says every TileLang operator used to train its V4-series models now has a high-performance Ascend version, and the Python API stays identical on either backend. Write once, run on NVIDIA or Huawei — no rewrite tax.
Why this is Jensen Huang's nightmare scenario
Back in April 2026, NVIDIA CEO Jensen Huang said on a podcast that if DeepSeek were the first to launch on Huawei chips, it would be a "bad outcome" for the US. His fear was never just about Huawei's chips catching up — it was about AI models starting to optimize around Huawei's architecture instead of NVIDIA's. Once developers find that models run well by default on non-American hardware, the advantages of American chips quietly fade.
DeepSeek has been walking that exact path. In April 2026 it adapted its V4 model — a 1.6 trillion-parameter system reported to match GPT-5 and Claude Opus on benchmarks — to run on Huawei hardware, and Huawei has said its chips took part in V4 training. The September toolkit generalizes that one-off achievement: first prove the frontier model works on domestic silicon, then hand every developer the software to follow.
The money, the metal, and the master plan
The software is only half the strategy. Reports say DeepSeek expects new Huawei training chips as early as the fourth quarter of 2026, and it is planning a data center in Inner Mongolia designed for at least 160,000 Huawei Ascend accelerators — a scale of commitment that only makes sense if the company is confident the software stack can carry real workloads. The lab is also reportedly raising around $7.5 billion, which would fund exactly this kind of infrastructure build-out.
None of this happened in a vacuum. US export controls were designed to keep Chinese labs dependent on NVIDIA; instead, each restriction has given Chinese companies another reason to build a stack that needs no American approval to run. NVIDIA's H20 — the cut-down chip built for the Chinese market — has faced its own licensing troubles, which only accelerates the search for alternatives.
What this means for the rest of us
For developers outside China, the toolkit is a signal, not yet a tool: the question is whether a community forms around Ascend software the way one formed around CUDA. For NVIDIA investors, it is the first real crack in the moat. And for anyone watching the AI race, it is the clearest test yet of a big question: do export controls slow China down — or just speed up its independence?
Do you think the future of AI runs on whoever writes the best software, or whoever builds the best chips? Let us know in the comments.
Comments
Post a Comment