Almost every headline-making AI model outside China runs on the same hardware: Nvidia GPUs. That dependence is also a vulnerability, and in 2026 it became a strategic problem. Export controls limited the most powerful Nvidia chips from reaching China, and Beijing responded by betting heavily on a domestic alternative: Huawei's Ascend (昇腾) series.
If you are not in hardware, the comparison can feel like alphabet soup. Here is what actually matters.
Key takeaways
- Nvidia dominates global AI compute with its GPU lines (datacenter parts such as the H-series and B-series) and the CUDA software ecosystem that almost everything is built on.
- Huawei Ascend (昇腾) is China's flagship AI-accelerator family — parts like the 910 series — paired with CANN, Huawei's software stack, and packaged into CloudMatrix supernodes.
- The real gap is not only silicon. It is software: CUDA's depth versus CANN's relative youth, and the model libraries that assume Nvidia.
- China's answer is a broad domestic compute buildout — government and enterprise clusters meant to run on local chips.
The chokepoint: export controls
The story starts with policy. The most capable Nvidia datacenter GPUs are restricted from sale to China, pushing Chinese buyers toward reduced-spec exports or domestic parts. For labs and cloud providers, that means either accepting weaker hardware or switching ecosystems entirely.
That constraint is why "build our own" stopped being a preference and became a plan. China's 15th Five-Year Plan (十五五, 2026–2030) and provincial compute programs funnel investment into local accelerators, with the explicit goal of reducing reliance on foreign GPUs.
What Ascend actually is
Huawei's Ascend family is a line of AI processors designed for training and inference. The 910-series parts are the workhorses; they are assembled into large systems such as CloudMatrix supernodes — clusters that pack many accelerators with high-speed interconnects so they train big models together.
Ascend's pitch is sovereignty and scale: a Chinese lab can build a training cluster without waiting on export licenses. The cost is a younger software stack and a smaller library of ready-made tools.
The software gap: CUDA vs CANN
Here is the part outsiders miss. Nvidia's true moat is not the chip — it is CUDA, the software layer almost every AI framework targets. Models, libraries, and engineer habits assume CUDA. Swapping to Ascend means porting code to Huawei's CANN stack and re-validating performance.
CANN has improved fast and Huawei ships tools to ease porting, but depth takes years of ecosystem use. For many Chinese teams, the pragmatic path is to write models that run on both — keeping options open as the domestic stack matures.
China's compute buildout
The response is not one chip but a national buildout. Reports through 2026 describe compute centers and "compute voucher" programs — public credits that let startups and research labs train on domestic hardware. One program cited roughly ten regional compute centers issuing vouchers, lowering the barrier for smaller teams to experiment on Ascend rather than scarce Nvidia parts.
The bet is that volume and policy will do for CANN what market dominance did for CUDA: make the local stack the default by sheer usage.
The honest caveats
Ascend is competitive on paper and at scale for many workloads, but benchmarking varies by task, and the newest Nvidia parts still lead on peak training throughput for some models. Porting pain is real, and not every framework is Ascend-ready. "Self-reliance" is a direction, not a finished state — hybrids (both ecosystems in one shop) are the near-term reality.
Honest limitations
This is an explainer based on public reporting and company disclosures, not a benchmark the author ran. Export-control specifics change frequently and are summarized, not quoted from primary regulation. Performance comparisons between Ascend and Nvidia parts depend on model, precision, and interconnect, and published figures vary by source. The "compute voucher" and regional-center details reflect 2026 coverage and may not represent the full national program.
Sources
Huawei Ascend and CANN product information; Nvidia datacenter GPU disclosures; 2026 reporting on export controls and China's domestic compute buildout, including CloudMatrix supernodes and regional compute-voucher programs.