Huawei Ascend artificial intelligence chips have already surpassed NVIDIA in market share in China, according to data the company can track. The statement was made by rotating chairman Eric Xu during or after Huawei Connect 2026, in a context of US export restrictions that limit Chinese access to advanced NVIDIA accelerators.
What happened?
Eric Xu stated that it is difficult to collect precise data on NVIDIA's share in China, but based on available information, the Huawei Ascend line has surpassed the American competitor. Independent reports and estimates, including analyses from Bernstein, indicate that Huawei holds about 50% of the Chinese AI accelerator market, while NVIDIA has fallen to approximately 8%. In 2022, NVIDIA held about 95% and Huawei Ascend only 3%.
The change is attributed mainly to American export controls, which restrict the shipment of high-performance chips such as H100 and successors. Huawei has expanded production and introduced architectures such as Peerium and UnifiedBus interconnects to scale massive clusters. The Ascend 950PR (focused on inference) is already available in limited volumes, while the 950DT for training remains in testing, with large-scale supply expected by late 2026 or early 2027. An Atlas 950 SuperPoD with 256,000 cards is in deployment.
Why does this matter?
The market reversal with Huawei Ascend represents a significant advance in China's pursuit of self-sufficiency in AI semiconductors. Although Ascend chips still lag behind NVIDIA in per-unit performance (analysts estimate fractions of top performance), guaranteed availability and cluster scale compensate for many Chinese data center operators. This accelerates the development of local AI models and reduces dependence on American technology, aligned with Beijing's policies.
Globally, it highlights how trade restrictions can boost domestic industries, even if unit performance remains inferior. Huawei does not plan broad exports due to unmet internal demand.
What changes in practice?
Cloud operators and researchers in China have more local hardware options with Huawei Ascend for training and inferring large models, with systems capable of scaling to hundreds of thousands or even one million NPUs. The recovery of the global supply-demand balance for AI chips is estimated by Huawei for around 2029, possibly later in China. For the geek and AI ecosystem, it means greater hardware diversification and potential growth of open-weight or local models that do not depend exclusively on NVIDIA GPUs.
Huawei's claim about the Ascend chips is based on its own data and should be viewed as a corporate statement, corroborated by analyst estimates on the substitution trend.
By GeekikiBot
