Huawei Rotating Chairman: Ascend AI Chip Sales in China Have Surpassed NVIDIA
nashnova research
Huawei rotating chairman Xu Zhijun says Ascend AI accelerators have surpassed Nvidia in Chinese sales and market share; U.S. export controls are pushing China's AI compute demand toward domestic suppliers faster than expected.
What is the actual evidence that Ascend outsells Nvidia in China?
Xu's claim rests on unit sales and market share — Ascend has overtaken Nvidia inside China.
Goldman Sachs data offers a cross-check: Ascend held roughly 20% of China's AI-chip shipment market in 2026, making it one of the country's leading domestic AI-accelerator suppliers.
This means → Ascend is no longer just "a usable alternative" — it is becoming China's default option.
How much presence does Nvidia still have in China?
Nvidia has resumed limited shipments of the H200 to China, but the resulting revenue accounts for less than 1% of its data-center business.
This means → For Nvidia, China has become marginal in revenue terms. For Chinese buyers, Nvidia's supply is both limited and uncertain.
In plain terms = Even though some Nvidia products have trickled back in, Chinese customers no longer dare stake their compute lifeline on a source that could be cut off at any time.
If each Ascend chip trails Nvidia, how does Huawei close the gap?
Huawei's answer is system-scale: its AI compute architecture is designed to scale up to one million NPUs.
Over 1,000 Ascend 910C super-nodes are already deployed; the Ascend 950 super-node has entered mass commercial use. More than 300 CloudMatrix 384 super-nodes run at over 20 customer sites.
In plain terms = If a single chip can't outrun yours, wire tens of thousands together — what matters is not one runner's speed but the whole team's throughput.
Software ecosystem — how is Ascend's biggest weakness changing?
HiSilicon chief scientist Liao Heng acknowledged that Ascend's biggest bottleneck was never hardware — it was the CUDA ecosystem.
Over roughly the past 18 months, AI models have increasingly adopted a "mega-kernel" programming paradigm — a coding approach that decouples model code from any single chip-software platform — reducing dependence on CUDA. Tools recently released by DeepSeek have further lowered the barrier for developers migrating to Ascend.
This means → Nvidia's deepest moat — CUDA software lock-in — is being eroded by an industry-wide trend, not just by Huawei attacking it alone.
Will Huawei challenge Nvidia head-on overseas?
Xu stated plainly: current production capacity cannot serve both massive domestic demand and large-scale overseas expansion at the same time. There are no near-term plans to challenge Nvidia globally.
Present demand is concentrated among Chinese cloud providers, government bodies, and large enterprises, plus selected overseas buyers seeking sovereign AI infrastructure or an alternative to Nvidia.
This reflects a clear-eyed read of Huawei's own pace: lock down the Chinese market first; overseas expansion comes after capacity and ecosystem have both matured.
What exactly are export controls accelerating?
Xu's core argument: U.S. export controls forced China to confront the risk of depending on external compute — supply security now outranks the pursuit of cutting-edge chip specs.
This means → The deeper effect of the controls is not just restricting chip flows — it is driving China to build an independent AI compute stack.
The key test ahead: whether Ascend can continue narrowing the gap with Nvidia in ecosystem maturity, supply-chain stability, and commercial scale outside China.
市场有风险，内容仅供研究参考，不构成投资建议。