Chinese AI company Z.AI has released a new large language model, GLM-5.2, that it says was trained without using any Nvidia chips.
According to Decrypt, the model rivals Anthropic's Claude Opus while using zero Nvidia chips. South Korean outlet 디지털투데이 reports that GLM-5.2 was trained using only Huawei chips, rather than the Nvidia hardware that has dominated cutting-edge AI development.
The claim matters because access to Nvidia's chips has become a flashpoint between the United States and China, with export controls limiting what Chinese firms can buy. A frontier-class model trained entirely on domestic Huawei silicon would suggest China is finding ways to build competitive AI without American hardware.
The release also fed a public debate about how fast China is catching up. According to 富途牛牛, a social media user asked when China's large models would reach "Fable-level" capability. Elon Musk responded that it was "Possibly Q1 next year." Tang Jie, the CEO of Zhipu AI, was more bullish, saying, "It won't take that long."
The sources do not detail benchmark figures, training costs, or the number of Huawei chips used, so the performance and efficiency claims rest on the companies' own framing rather than independent testing.
Why it matters: if a Chinese lab can train a top-tier AI model without Nvidia, the chip export controls meant to slow China's progress may be less of a barrier than Washington hopes.