Z.AI, a Beijing-based AI company that grew out of Tsinghua University, just finished building what may be the largest AI data center ever completed by a Chinese AI lab. It uses one gigawatt of power, which is enough electricity to run 750,000 homes at once. Every processing chip inside it is made in China.
The company did not build this because it preferred Chinese chips. The US Commerce Department placed Z.AI on its export blacklist in January 2025, cutting off its legal access to Nvidia hardware entirely. When you cannot buy the tools you need, you build alternatives. That is what happened here.
The facility houses several computing clusters, each containing more than 10,000 chips. The chips appear to be Huawei's Ascend processors, based on Z.AI's own training history. The company has said its recent AI models, including GLM-5 and GLM-5.2, were trained on 100,000 Huawei Ascend 910B chips with no Nvidia hardware at any stage of the process.
The results have been competitive. GLM-5.2, released in June 2026, ranked fourth on one of the main independent AI model leaderboards and scored within roughly one percentage point of Anthropic's flagship model on a closely watched coding benchmark. It costs about five times less per query than leading Western models. Z.AI's stock crossed HK$1 trillion in market value shortly after the launch.
There are real limits to Chinese chips that should not be glossed over. Huawei's best chips currently deliver around 60% of the computing throughput of an Nvidia H100 per chip. Training Z.AI's models on Huawei hardware required about 15% more computing time than equivalent Nvidia-based runs, and the chips generate fewer output tokens per second during actual use. The performance gap on the hardest reasoning tasks is still meaningful: on one demanding benchmark specifically designed to test genuine problem-solving, the best Chinese models scored 11.8% against much higher scores from leading US labs.
But here is the key point that gets buried in the technical discussion: China is not trying to match Nvidia chip for chip. It is building enough capacity, at sufficient quality, to keep training and running competitive AI models without any American hardware in the loop. Z.AI just proved that is possible at gigawatt scale.
The broader context makes this more significant. Chinese domestic chip suppliers led by Huawei are now expected to hold 56% of China's AI chip market in 2026, up from 46% the previous year. Foreign suppliers, including Nvidia, are forecast to fall to 21% market share in China. Nvidia's own CEO has described the company's China market share as effectively zero. Meanwhile, Beijing is drafting a national plan to spend $295 billion over five years on AI data centers, with targets requiring at least 80% domestic sourcing.
The sanctions were designed to slow Chinese AI development by cutting off the best hardware. The practical effect has been to accelerate China's push to build hardware it does not need permission to use. Z.AI is on track to hit $1 billion in annual recurring revenue, having already reached its 2026 sales target in July.
For any business currently using or evaluating AI tools, the practical implication is this: a Chinese AI model that is nearly as capable as the market leaders, fully open to download and use, and significantly cheaper per query now exists and was built without a single chip that the US government approved. The supplier environment for AI services is getting more competitive, not less.