Z.AI, a leading Chinese AI lab, has begun operating a 1-gigawatt data centre powered exclusively by domestic chips, marking a significant milestone in reducing reliance on Nvidia hardware blocked by US export controls.
- Z.AI runs clusters with more than 10,000 China-made AI chips each
- The 1GW centre rivals power of 750,000 average homes
- Domestic chip use bypasses US export restrictions on Nvidia GPUs
What happened
Z.AI, formerly Zhipu, has launched partial operations of a massive new data centre dedicated to training its General Language Models (GLM). This site, consuming 1 gigawatt of power, is one of the largest AI server facilities constructed by any Chinese lab. It hosts multiple computing clusters, each equipped with over 10,000 chips developed by Chinese manufacturers, notably excluding any Nvidia components.
The project demonstrates concrete progress in China’s quest to substitute foreign AI hardware with homegrown solutions. While the exact chip models remain undisclosed, contenders include Huawei, Cambricon, and Alibaba, known players in China's AI chip sector. This deployment represents a significant physical infrastructure build, not just design concepts or plans.
Why it matters
The centre’s exclusive use of domestic chips addresses long-standing questions about the capability of Chinese AI accelerators to handle training of advanced models at scale. US export controls have prevented Chinese labs from acquiring Nvidia’s top-tier GPUs essential for training frontier AI systems, making this a critical test of China’s chip development progress.
By successfully assembling and powering a gigawatt-scale data centre without Nvidia gear, Z.AI signals China's growing self-reliance in AI compute resources. This counters assumptions that sanctions might severely hinder China’s AI development pace and could reshape global competition dynamics in AI innovation.
What to watch next
The key evaluation will be whether Z.AI’s next generation GLM, trained on this domestic infrastructure, matches or surpasses performance levels of US AI models trained on Nvidia hardware. Technical challenges remain in achieving comparable efficiency and system stability when scaling thousands of newer Chinese chips together.
Additionally, monitoring the continued expansion of China’s AI data centres amid significant government and corporate investment, including Alibaba and China Telecom, will indicate how rapidly China can consolidate a robust ecosystem independent of Western semiconductor technology. Z.AI is also emerging as a major enterprise AI provider, aiming for $1 billion in recurring annual revenue following its recent financial milestones.