Z.AI, the Chinese AI developer formerly known as Zhipu, has completed a roughly 1-gigawatt data center that runs exclusively on domestic accelerators, giving one of the country’s leading model builders a frontier-scale training site with no Nvidia hardware inside it.
The facility has begun partial operations and is built to train and develop Z.AI’s GLM family of models. It houses several computing clusters, each with more than 10,000 Chinese-made chips. At roughly one gigawatt, its continuous power draw is comparable to that of about 750,000 homes, placing it among the largest AI campuses now under construction anywhere.
The build is a direct response to US export controls. Z.AI was added to the Commerce Department’s Entity List in January 2025, cutting off access to Nvidia’s most advanced chips. Rather than wait for that access to return, the company has trained its GLM series on Huawei Ascend accelerators using Huawei’s MindSpore software stack, and has now assembled a gigawatt of domestic compute to scale that work. The specific accelerators inside the new facility have not been identified.
The data center is among the clearest physical signs yet that China’s top labs are building a fully domestic training stack rather than routing around controls chip by chip. It follows Huawei’s push to close the hardware gap with new scaling techniques and its move to sell Ascend systems abroad, and it arrives as Beijing weighs limits on overseas access to its most capable models. It also fits a broader national push: Chinese authorities are drafting plans to spend on the order of 2 trillion yuan, about 295 billion dollars, over five years on a nationwide network of AI data centers built largely with domestic chips.
The timing is pointed. Z.AI’s completed campus surfaced the same week Moonshot’s Kimi K3 triggered a sharp selloff in AI and semiconductor stocks, with Chinese model developers caught in the downdraft; shares in Z.AI, listed in Hong Kong, fell about 28 percent on July 17, their steepest single-day drop since the company went public. A model priced to undercut Western labs and infrastructure built to run without Western chips point at the same strategy from two directions: reduce dependence on a US supply chain that Washington has shown it will restrict at will, even as it continues to approve some sales to keep Chinese firms buying American.
For US policymakers, the facility complicates the theory behind the current export regime. That approach assumes withholding advanced chips slows Chinese frontier development. A gigawatt of domestic silicon dedicated to a frontier model line is evidence that the constraint is instead accelerating a homegrown alternative.
Sources: TechNode, Tom’s Hardware, Bloomberg
–
By the Control Plane Editorial Team