Z.AI powers 1GW AI hub without Nvidia chips

For the past year, a basic question has hovered over China’s fast-improving AI models: what hardware are they actually using? Z.AI, formerly known as Zhipu, has now offered part of the answer. According to Bloomberg, citing a person familiar with the matter, the company has partially started operating a 1-gigawatt data center built to train its GLM models — and it uses only Chinese-made chips.

That scale is notable on its own. 1 gigawatt is roughly the power draw of 750,000 homes at a given moment, putting the site among the largest server hubs built by any Chinese AI lab. Bloomberg also reported that Z.AI runs several computing clusters, each with more than 10,000 chips. None of them are Nvidia parts, according to the source, though the report did not identify the exact chips.

The likely contenders are domestic suppliers led by Huawei, alongside firms such as Cambricon and Alibaba, all trying to narrow the gap with Nvidia. That matters because US export controls have blocked Chinese labs from buying Nvidia’s most powerful accelerators. The key unknown was whether Chinese chips were good enough not just to run models, but to train frontier systems at scale.

Recommended reading

OVH patched Januscape with mass reboots and a Sydney test

A domestic 1GW cluster does not settle that question completely, but it is the clearest sign yet that Chinese labs may keep scaling even without the industry’s default hardware. The timing is especially sharp: the buildout comes just days after Beijing-based Moonshot’s Kimi K3 matched top US models, then paused new sign-ups after running short on compute.

China data center spending and Z.AI revenue

Z.AI is expanding into a much larger national push. China is preparing to spend around 2 trillion yuan, or about $295bn, over five years on data centers across the country. So far, Alibaba and China Telecom have been the biggest builders.

Z.AI also appears to have fresh capital behind it. After a Hong Kong listing and a follow-on share sale, the company is on track for $1bn in annual recurring revenue, which would make it the first Chinese AI company to hit that mark. Its shares rose almost 20% on the day. The company is also positioning itself as an enterprise AI supplier, drawing comparisons with Anthropic.

There are still important caveats. The chip details come from an anonymous source, and Z.AI did not respond to a request for comment. Building a giant cluster is also not the same as proving it can train a frontier model as efficiently as an Nvidia-based system. Domestic chips may still lag on raw performance, and linking thousands of them into a stable training platform is difficult. The real test will be Z.AI’s next GLM model — and how it stacks up against what US labs release.