For a year, one question has hung over China’s rapidly improving AI models: what are they actually working on? ZAI has just given part of the answer. It built a huge data center using only Chinese-made chips.
The company formerly known as Zhipu has begun partially managing the site. Bloomberg reported on thisciting a person familiar with the matter. It is a 1 gigawatt center built to train Z.AI’s GLM models. ZAI did not respond to a request for comment.
What is built?
Libra is the title. One gigawatt is equivalent to powering approximately 750,000 homes at any given time. This puts the site among the largest server centers built by any Chinese AI lab.
The chip count is equally surprising. Z.AI now operates several computing clusters with more than 10,000 chips each. Most importantly, none of them are Nvidia’s.
Bloomberg did not name the exact chips. China’s leading designer of AI accelerators is Huawei, which is competing with domestic firms such as Cambricon and Alibaba to close the gap on Nvidia.
Why is it important?
US export controls have prevented Chinese labs from buying Nvidia’s most powerful chips. The open question was whether home-grown parts could carry the burden of training border models, not just driving them.
A 1GW cluster built entirely on native silicon is a real answer to this question. This suggests that Chinese labs can scale, even if they are cut off from the chips that the rest of the industry defaults to.
Timing sharpens the point. The construction comes days after the building in Beijing Kimi K3 from Moonshot Matched with the top models in the US, then the calculation fell short and stopped new registrations. Z.AI competes with the same opponents and bets to own the equipment.
Bigger construction
Z.AI does not do it alone. China is set to spend about 2 trillion yuan, about $295 billion, on data centers across the country over five years. Cloud giants Alibaba and China Telecom remain the biggest builders by far.
The company also has money to compete. Fresh from Hong Kong list and with the subsequent share sale, Z.AI is on track to generate $1 billion in annual recurring revenue, which makes it the first Chinese AI company to reach that mark. Its shares rose nearly 20% on the day.
It positions itself as an enterprise AI supplier, which invites comparisons with Anthropic.
Catch
A few caveats are worth keeping. The chip details come from an anonymous source and have not been confirmed by Z.AI. Creating a cluster is not the same as proving that it can train a boundary model as efficiently as Nvidia’s model.
Native chips can still lag behind in raw performance, and it’s difficult to integrate thousands of them into stable systems. The real test is the next GLM model and how it compares to what the US labs sent.
Again, the direction is clear. effort to establish a Chinese alternative Nvidia has moved from a slide program to a working, gigawatt-scale data center. That alone is changing the debate over how long export controls can last.
