For a year, one question has hung over China's fast-improving AI models: what are they actually running on? Z.AI has just given part of the answer. It has built a huge data centre that uses only Chinese-made chips.
The company, formerly known as Zhipu, has started partially operating the site, Bloomberg reported, citing a person familiar with the matter. It is a 1-gigawatt hub built to train Z.AI's GLM models. Z.AI did not respond to a request for comment.
What was built
The scale is the headline. A gigawatt is roughly the power draw of 750,000 homes at any one moment. That puts the site among the largest server hubs any Chinese AI lab has built. The chip count is just as striking. Z.AI now runs several computing clusters, each holding more than 10,000 chips, the person said. Crucially, none of them are Nvidia's.
Bloomberg did not name the exact chips. China's leading designer of AI accelerators is Huawei, which competes with local firms such as Cambricon and Alibaba to close the gap on Nvidia. Other domestic players include Hygon (a joint venture with AMD) and Startech (a spin-off from the Chinese Academy of Sciences). These companies have been developing chips that can handle AI workloads, but until now, their ability to scale to frontier model training was unproven.
Why it matters
US export controls have blocked China's labs from buying Nvidia's most powerful chips, such as the H100 and B200. The open question was whether homegrown parts could carry the load for training frontier models, not just running them. A 1GW cluster built entirely on domestic silicon is a real answer to that question. It suggests Chinese labs can keep scaling even while cut off from the chips the rest of the industry treats as default.
The timing sharpens the point. The buildout lands days after Beijing-based Moonshot's Kimi K3 matched top US models, then ran short of compute and paused new sign-ups. Z.AI is racing the same rivals, and it is betting on owning the hardware. Moonshot's temporary shortage highlights the fragility of relying on imported chips; Z.AI's move shows a deliberate strategy to secure supply chains.
The bigger buildout
Z.AI is not doing this alone. China is preparing to spend around 2 trillion yuan, about $295bn, over five years on data centres across the country. Cloud giants Alibaba and China Telecom remain the biggest builders so far. The government's push for 'self-reliance' in semiconductors has accelerated since 2022, when the US Department of Commerce imposed new export controls. This has led to a wave of domestic chip investment and R&D.
Z.AI also has the money to compete. Fresh from a Hong Kong listing and a follow-on share sale, Z.AI is on track for $1bn in annual recurring revenue, which would make it the first Chinese AI firm to reach that mark. Its shares jumped almost 20% on the day. The company is positioning itself as an enterprise AI supplier, a role that invites comparison with Anthropic. It is targeting sectors such as finance, healthcare, and manufacturing, where Chinese companies are eager to adopt AI but wary of foreign dependencies.
Technical challenges and limitations
Building a 1GW cluster with domestic chips is not straightforward. Chinese chips still lag on raw performance compared to Nvidia's latest offerings. For instance, Huawei's Ascend 910B is roughly comparable to Nvidia's A100 in some benchmarks, but the A100 is already two generations behind Nvidia's Blackwell architecture. Moreover, stitching thousands of chips into a stable training system requires sophisticated networking and software. Nvidia's CUDA ecosystem is a huge advantage; Chinese alternatives like Huawei's CANN or Baidu's PaddlePaddle are maturing but still have gaps in libraries and developer tools.
The person familiar with the matter noted that Z.AI's cluster is partially operational, meaning full-scale training may not yet be underway. The real test will come when Z.AI trains its next GLM model on this hardware and compares its performance to models trained on Nvidia chips. If the results are competitive, it will be a major blow to US export control strategy.
Historical context and industry implications
The path to this data centre began in 2020, when the US first restricted exports of certain AI chips to China. At that time, Chinese companies relied heavily on Nvidia's V100 and A100 chips for training. The restrictions forced a pivot: companies like Alibaba, Baidu, and Huawei began developing their own accelerators. Zhipu (now Z.AI) was founded in 2019 as a spinoff from Tsinghua University, focusing on large language models. It received backing from Alibaba, Tencent, and others, allowing it to invest heavily in compute.
The 1GW cluster is not the only such facility in China. Earlier this year, ByteDance (owner of TikTok) built a cluster with nearly 10,000 Huawei chips, and Alibaba Cloud operates dozens of smaller clusters. However, the Z.AI project is the largest single-site, domestic-only data centre reported so far. It indicates that Chinese companies are not just copying foreign designs but are innovating in hardware and cooling technologies. The data centre likely uses advanced liquid cooling to manage heat from so many chips.
Another angle is software. Training large models requires not just hardware but also efficient algorithms and parallelization techniques. Chinese AI labs have developed techniques like Model parallelism and ZeRO optimization that reduce memory requirements. They have also pioneered sparsity and mixture-of-experts (MoE) architectures, which can reduce compute demands. These innovations may partially compensate for the weaker performance of domestic chips.
Competitive landscape and future outlook
Z.AI's main competitors in China include Baidu (ERNIE models), Alibaba (Qwen models), Moonshot (Kimi), and DeepSeek. All are racing to achieve parity with US models like GPT-4 and Claude. The shortage of advanced Nvidia chips has forced them to collaborate or share resources. For instance, a consortium of Chinese AI labs recently proposed a shared compute network, but Z.AI has chosen a different path: vertical integration.
If Z.AI succeeds, it could become the de facto platform for enterprise AI in China, similar to how Anthropic positions itself for enterprise use in the West. It would also weaken the argument that China cannot compete in frontier AI without US chips. However, the risk remains that US export controls will tighten further: the Biden administration has considered restricting memory bandwidth of exported chips and limiting software support. China is also racing to develop its own chip fabrication capabilities, but TSMC's latest nodes remain out of reach.
In the short term, the 1GW cluster will likely be used to train Z.AI's next-generation GLM model, expected within months. The results will be scrutinized by analysts worldwide. If the model matches or exceeds the performance of existing open-source models like Llama 3, it will validate China's domestic hardware strategy. If not, it may show that software and scale alone cannot overcome hardware gaps.