For years, the assumption underlying US semiconductor export controls was that cutting off China’s access to Nvidia’s most advanced chips would slow its AI development in a meaningful way. A new analysis by the South China Morning Post, published on June 17, 2026, challenges that assumption with concrete, verifiable evidence: five major Chinese AI models, trained entirely on domestic chips, that are performing at or near the frontier of global AI capability.
The five models are DeepSeek-V4, Alibaba’s Qwen3, Moonshot’s Kimi K2, Zhipu AI’s GLM-5.2, and Baidu’s ERNIE 5.0. Together, they represent a cross-section of China’s AI landscape, a pure-play AI research lab, a cloud hyperscaler, a well-funded startup, an enterprise AI company, and a search giant. The fact that all five have successfully trained frontier-class models on domestic chips is not a coincidence; it is the result of a sustained, coordinated effort to build a chip ecosystem that does not depend on American hardware.
The Hardware Behind the Models
The chips used to train these models are primarily Huawei’s Ascend 910B and 910C series, supplemented by accelerators from Cambricon, Biren Technology, and Moore Threads. None of these chips match the raw performance of Nvidia’s H100 or H200 on standard benchmarks, but the SCMP analysis argues that the performance gap has narrowed to the point where it no longer constitutes a fundamental barrier to training competitive AI models.
The key insight is that training efficiency is not solely a function of peak chip performance. It also depends on the software stack, the compilers, libraries, and distributed training frameworks that determine how effectively a model can utilize the available hardware. Chinese AI companies have invested heavily in optimizing their software stacks for domestic chips, and the results are reflected in the models they have been able to produce.
DeepSeek’s approach is the most extensively documented. The company has published detailed technical reports describing how it achieved training efficiency on Huawei Ascend hardware comparable to what other labs achieve on Nvidia hardware, through a combination of custom CUDA-equivalent kernels, modified attention mechanisms, and novel parallelization strategies. Those techniques are now being adopted by other Chinese AI labs, accelerating the transition away from reliance on Nvidia.
The Competitive Implications
The practical significance of this development extends beyond the question of whether Chinese AI can survive without Nvidia. It raises a more fundamental question about the long-term structure of the global AI hardware market.
If Chinese AI companies can train frontier models on domestic chips, the addressable market for Nvidia’s products in China, already substantially reduced by export controls, shrinks further. The companies that were once Nvidia’s largest customers in China are now its competitors’ customers, and the transition appears to be durable rather than temporary.
For Huawei, the success of Chinese AI labs on Ascend hardware is a powerful commercial endorsement. The company is projecting 62% market share in China’s AI accelerator market in 2026, a figure that would have seemed implausible two years ago. The five-model evidence base provides the kind of concrete validation that enterprise customers require before committing to a hardware platform at scale.
The SCMP analysis also notes that the transition to domestic chips has been accompanied by a shift in the economics of AI development in China. Training costs on Ascend hardware are lower than on equivalent Nvidia hardware, partly because the chips themselves are cheaper, and partly because the absence of export control premiums eliminates the black-market markup that has historically inflated the cost of restricted chips in China.
What Remains Difficult
The SCMP analysis is careful not to overstate the case. Inference, running trained models at scale to serve user requests, remains more challenging on domestic chips than on Nvidia hardware, particularly for the largest models. The memory bandwidth and interconnect performance of current domestic chips are not yet at the level required to support frontier models at the latency and throughput demanded by commercial applications.
The chip design tools and manufacturing processes required to produce the next generation of domestic AI accelerators also remain significant constraints. EastFrontier has previously reported on Huawei’s Tau Scaling Law and LogicFolding architecture, which represent one approach to closing the manufacturing gap without access to EUV lithography. Whether that approach can deliver the performance improvements required to close the inference gap, not just the training gap, will be one of the defining questions for China’s AI hardware ecosystem over the next two to three years.
For now, the evidence from five frontier models trained on domestic chips is a meaningful data point in a debate that has often been conducted in the abstract. The chips are not perfect substitutes for Nvidia’s best products. But they are good enough, and getting better.