What happened
Huawei Technologies announced that a significant portion of AI model training in China will migrate to its Ascend-based SuperPoD and SuperCluster systems in 2027. Rotating chairman Eric Xu Zhijun stated that Ascend chips have overtaken Nvidia in domestic market share, although specific figures were not provided. The company unveiled the Atlas 960 SuperPoD and Peerium Computing Architecture, while noting that production constraints will keep these systems primarily within China.
During the Huawei Connect 2026 conference in Shanghai, rotating chairman Eric Xu Zhijun told reporters that starting next year, a large volume of AI model training will be based on the SuperPoD or SuperCluster systems powered by Ascend 950DT processors. Xu asserted that Huawei’s proprietary Ascend chips have overtaken Nvidia in domestic market share in China, though he did not provide specific percentage figures to substantiate this claim.
The announcement comes as Nvidia confirmed last month that it had resumed shipments of its H200 processors to China under a new US licensing arrangement. However, these sales represented less than 1 percent of Nvidia's second-quarter data-centre revenue. SCMP reported last month, citing sources at major Chinese AI developers, that China’s most advanced AI models were still being trained on Nvidia chips, suggesting a lag between market share claims and actual high-end model training infrastructure.
Huawei unveiled its new Atlas 960 SuperPoD computing cluster on Thursday, which is powered by Ascend 960 chips and claims double the training and performance of its predecessor. The company also introduced the Peerium Computing Architecture, designed to interconnect one million processors to function as a single computer. While Peerium is already deployed on Atlas 950 systems, it is currently undergoing testing on the new Atlas 960 hardware.
Despite these technological advancements, Xu cautioned that international availability for Huawei’s latest hardware remains constrained. He stated that production limits would keep the systems mostly within China, noting that the company does not have enough capacity to satisfy domestic demand, let alone expand to the international market in a fully fledged way. However, Huawei has been shipping small quantities overseas for testing in high-demand regions.
Huawei reported that it has deployed over 1,000 Atlas SuperPoD units across more than 370 enterprise customers. Xu emphasized that many clients chose Ascend due to geopolitical risks and a preference for multi-vendor strategies, stating that pushing full chip self-sufficiency is the certain path forward for China. He projected that global AI computing supply and demand would reach equilibrium around 2029, with China reaching a balance slightly later in 2030.
Why it matters
This marks a pivotal shift in the global AI hardware landscape, indicating that US export controls may be accelerating China's domestic semiconductor self-sufficiency rather than solely restricting its capabilities. The transition to Huawei infrastructure suggests a decoupling of Chinese AI development from Nvidia's ecosystem, with practical implications for supply chain resilience and the competitive dynamics between US and Chinese AI hardware providers.
The shift to Huawei infrastructure signals a structural change in China's AI ecosystem, reducing reliance on US semiconductor technology. This development is a direct response to US export controls designed to restrict China's access to advanced chips, suggesting that these controls may be inadvertently fostering a robust domestic alternative.
The claim that Ascend has overtaken Nvidia in domestic market share is significant, but it lacks independent verification or specific data points. The discrepancy between this claim and recent reports that top Chinese models still train on Nvidia chips highlights the complexity of the transition and the potential gap between general market share and high-performance training infrastructure.
For the global AI industry, this move indicates a bifurcation of the hardware market. Chinese AI developers may increasingly standardize on Huawei's ecosystem, creating a separate technological trajectory that could impact the interoperability and competitive landscape of AI models developed in China versus the rest of the world.
The production constraints mentioned by Huawei suggest that while the strategic shift is underway, the practical implementation is limited by manufacturing capacity. This bottleneck could delay the full realization of the predicted 2027 shift, keeping Nvidia relevant in the Chinese market for high-end applications until Huawei scales its production.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
What is the best response when AI Models Explained makes a mistake in production?
What to watch next
Monitor the actual deployment rates of Ascend 950DT and 960 chips among major Chinese AI developers to verify the claimed market share shift. Watch for any changes in US licensing policies regarding Nvidia H200 shipments to China, as well as Huawei's ability to scale production to meet domestic demand before 2027.
Independent verification of the market share claims is crucial. Look for data from third-party analysts or disclosures from major Chinese AI labs regarding their actual training infrastructure composition in 2026 and 2027.
Monitor the performance benchmarks of the Atlas 960 SuperPoD and Ascend 960 chips compared to Nvidia's latest offerings. Real-world performance in training large language models will determine if the 'double performance' claim holds up under practical workloads.
Watch for updates on US-China semiconductor negotiations. Any further easing or tightening of export controls could accelerate or hinder the adoption of Huawei's chips, as well as the availability of Nvidia's H200s in China.
Track Huawei's production capacity expansions. The company's ability to meet domestic demand will be a key indicator of whether the 2027 shift is feasible or if supply constraints will continue to limit the pace of adoption.