Chinese Firm Trains 744-Billion Parameter Model on Huawei Chips Without Nvidia

Reports indicate that Chinese company Zhipu AI completed the training of its 744-billion-parameter language model, GLM-5.2, entirely on Huawei Ascend chips, serving as real-world proof that China is now capable of building frontier-scale models without needing Nvidia hardware.

This achievement holds immense strategic importance amid escalating U.S. export restrictions on advanced AI chips to China, demonstrating that the restrictions have not halted Chinese technological progress, but rather pushed companies to build alternative, domestically produced supply chains.

Huawei Ascend chips hold a central position in China's strategy for technological independence, though they were previously considered less efficient than their Nvidia counterparts for training massive models. Zhipu AI's success breaks this prevailing belief, according to reports.

This breakthrough could impact the global AI chip market by weakening the argument that U.S. hardware dominance constitutes an insurmountable barrier, which may increase pressure on policymakers in Washington to tighten restrictions or adjust their approach.

What do these terms mean?

Frontier Model: The most advanced and largest model in its category globally at a given time — representing the front edge of what is technologically possible in artificial intelligence.