A year ago, Nvidia was essentially the only serious option for AI compute in China. That’s no longer true. According to Huawei Central, Huawei’s Ascend AI chips have crossed the 50% market share threshold in China, while Nvidia has fallen from roughly 95% down to just 8%. For anyone tracking the AI hardware space, that number deserves a second read.
This didn’t happen overnight. US export restrictions on advanced semiconductors, which have been tightening since 2022, effectively forced Chinese AI companies to find alternatives. Huawei stepped into that gap with its Ascend 910B and more recently the 910C, chips designed to compete with Nvidia’s A100 and H100 series. They aren’t perfect substitutes, but they’re good enough, and for Chinese buyers operating under sanctions, good enough is the only option on the table.
The practical impact is significant. Major Chinese AI developers including Baidu, ByteDance, and various state-backed research institutions have been shifting workloads to Ascend hardware. This creates a compounding effect: more usage means more software optimization, better tooling, and a maturing ecosystem around Huawei’s CANN compute framework, which is the rough equivalent of Nvidia’s CUDA. CUDA’s dominance has always been as much about software lock-in as raw hardware performance. Huawei is now building that same kind of stickiness.
For Nvidia, the China revenue loss is real but somewhat contained. The company has been clear that export-restricted markets represent a shrinking portion of its total addressable market, and demand from US hyperscalers and enterprise customers remains intense. Still, ceding 95% to 8% in any major market is a signal worth taking seriously, especially as other governments watch closely and consider their own chip dependency strategies.
The bigger picture here is about supply chain bifurcation. The AI chip market is splitting into two tracks: one built around Nvidia and its CUDA ecosystem, and one increasingly built around Huawei’s Ascend stack inside China. These two tracks are not compatible, and the gap between them is widening by design. Companies building AI infrastructure now face a genuine architectural fork depending on which markets they intend to serve. That’s a sourcing and strategy question, not just a technical one, and it’s arriving faster than most organizations planned for.
