5 Chinese AI models now training on homegrown chips, bypassing Nvidia
Synopsis
For the first time, a cluster of five Chinese AI models — including DeepSeek-V4-Pro and Alibaba's LongCat-2.0-Preview — are actively training on domestic chips from Huawei and Moore Threads, a move that could redefine the global AI hardware race even as China's silicon still trails Nvidia on raw performance.
Key Takeaways
Five Chinese AI models — including GLM-Image , DeepSeek-V4-Pro , EvoPhys-World , LongCat-2.0-Preview (Alibaba), and BitCPM-CANN (ModelBest) — are experimenting with domestic chips for AI training as of June 2026 .
Huawei's Ascend hardware and Moore Threads Technology GPUs are the primary domestic alternatives being tested at scale.
Domestic chips are already widely used for AI inference , but no Chinese top-tier model has publicly completed full pre-training on homegrown silicon.
Natixis economist Gary Ng said relying on indigenous suppliers means Chinese labs 'may not develop as quickly and efficiently as their US counterparts,' but noted China is building a domestic AI supply chain that is 'quite rare worldwide.' Washington's escalating export controls and Beijing's self-sufficiency drive are the primary forces accelerating the shift to domestic compute.
Meituan and Peking University are among the broader set of entities participating in China's domestic AI compute push.
A growing cohort of Chinese AI laboratories is training artificial intelligence models on domestically produced chips, marking a significant — if still incomplete — shift away from Nvidia hardware as Washington's export controls tighten and Beijing accelerates its push for technological self-sufficiency. As of June 2026, while domestic silicon remains widely used for model inference, none of China's top-tier models are publicly known to have completed full pre-training on homegrown processors — though that boundary is now being tested.
Understanding the three-stage AI pipeline
To grasp the significance of this shift, it is essential to understand how AI models are built. Pre-training is the most computationally intensive phase, in which a model ingests vast datasets to learn fundamental patterns. Post-training is a lighter process that fine-tunes the model to follow human instructions. Inference is the final, ongoing stage — running the finished model to respond to user queries in real time.
Domestic chips have made the most headway in inference, the least demanding of the three. The harder challenge — and the one now being actively pursued — is shifting pre-training and post-training workloads onto local silicon.
Five models pushing the frontier of domestic compute
GLM-Image, developed with ties to Peking University, is among the models experimenting with domestic hardware during training phases. DeepSeek-V4-Pro, one of China's most closely watched frontier models, is also reported to be exploring indigenous compute pipelines. EvoPhys-World, LongCat-2.0-Preview from Alibaba Group Holding, and BitCPM-CANN from ModelBest round out the cohort, with several leveraging Huawei's Ascend hardware and Moore Threads Technology GPUs. Meituan is also among the entities reported to be participating in this domestic compute push.
Each model represents a different segment of the AI landscape — from image generation to physics simulation to long-context reasoning — suggesting the experimentation is broad rather than confined to a single use case.
Why it matters: Building a full domestic AI supply chain
Relying on indigenous suppliers means Chinese AI labs 'may not develop as quickly and efficiently as their US counterparts,' according to Natixis economist Gary Ng. However, Ng added that in the long run, China is building 'an entire domestic AI supply chain,' which he described as 'quite rare worldwide.'
The strategic calculus is clear: short-term performance trade-offs are being accepted in exchange for long-term supply chain independence. Huawei's Ascend series and Moore Threads Technology processors are the primary domestic alternatives currently being stress-tested at scale.
The competitive backdrop
China's AI software layer has closed much of the gap with US peers, as evidenced by the global attention drawn by models like DeepSeek. The hardware gap, however, remains substantial. Nvidia's H100 and successor chips continue to outperform available domestic alternatives on raw throughput and energy efficiency for large-scale pre-training runs.
Yet the very act of training on domestic chips — even at reduced efficiency — yields critical engineering knowledge: labs learn how to optimise software stacks, compilers, and model architectures for non-Nvidia hardware, compounding gains over time.
What's next
The central question is whether any of China's flagship models will complete a full pre-training run on domestic silicon and publish verifiable benchmarks. If that threshold is crossed, it would represent a genuine inflection point in the global AI chip competition. Investors and policymakers tracking the US–China technology rivalry should watch Huawei Ascend adoption rates and any benchmark disclosures from DeepSeek-V4-Pro and Alibaba's LongCat-2.0-Preview in the months ahead.
Point of View
At-scale training on domestic hardware will compress the performance gap faster than export controls can widen it. Mainstream coverage focuses on benchmark comparisons; what it misses is the compounding software advantage that accrues when engineers spend years optimising for a specific non-Nvidia stack, as AMD's ecosystem history demonstrates. The participation of frontier labs like DeepSeek alongside platform giants like Alibaba signals this is a coordinated industrial strategy, not isolated experimentation. If even one flagship model publishes credible pre-training benchmarks on domestic silicon, it will shift the narrative — and market positioning — around the entire China chip-war thesis.
NationPress
2 Aug 2026
Frequently Asked Questions
Which Chinese AI models are being trained on domestic chips?
Five Chinese AI models are currently experimenting with domestic chips for training: GLM-Image , DeepSeek-V4-Pro , EvoPhys-World , LongCat-2.0-Preview from Alibaba Group Holding , and BitCPM-CANN from ModelBest . These models span image generation, physics simulation, and long-context reasoning tasks.
What domestic chips are Chinese AI labs using instead of Nvidia?
Huawei's Ascend hardware and Moore Threads Technology GPUs are the primary domestic alternatives being adopted by Chinese AI laboratories. These chips are already widely used for inference workloads, and labs are now pushing them into the more demanding pre-training and post-training phases.
Why are Chinese AI labs moving away from Nvidia chips?
Washington's escalating export controls have restricted Chinese firms' access to advanced Nvidia GPUs, while Beijing is actively pushing for technological self-sufficiency. Labs are accepting short-term performance trade-offs to build long-term supply chain independence, according to Natixis economist Gary Ng .
Can Chinese chips match Nvidia for AI training performance?
Not yet at the frontier level. As of June 2026 , none of China's top AI models are publicly known to have completed full pre-training on homegrown silicon. Natixis economist Gary Ng noted that relying on indigenous suppliers means Chinese labs 'may not develop as quickly and efficiently as their US counterparts.'
What is the significance of China building a domestic AI supply chain?
A fully domestic AI supply chain — spanning chips, compilers, and models — would insulate China's AI sector from future US export restrictions. Gary Ng of Natixis described such a supply chain as 'quite rare worldwide,' suggesting that even if slower, the long-run strategic value is substantial for Beijing's technology ambitions.
