Z.ai Builds Gigawatt Data Center on Chinese Chips Alone

Z.ai, the Chinese AI developer formerly known as Zhipu, has finished building a data center that runs entirely on Chinese-made chips and has begun operating part of it, according to a person familiar with the project cited by Bloomberg. The site is built to draw about a gigawatt of power — roughly what it takes to supply 750,000 homes — and to train the company’s GLM models without the Nvidia hardware that US export controls keep out of its reach.

The same person, who spoke anonymously, said Z.ai now operates or has built several computing clusters that each hold more than 10,000 chips. Neither the location of the gigawatt hub nor the specific accelerators inside it were disclosed, and the account rests on a single source rather than a permit, a filing, or a signed power contract. The direction is clear enough: one of China’s strongest model builders is standing up hyperscale training capacity that skips American silicon altogether.

What a gigawatt actually buys

A gigawatt is a serious figure. It puts this campus in the same weight class as the largest AI sites under construction in the US, and it makes power, not chips, the binding constraint — the same squeeze that is pushing the whole industry toward more efficient ways to compute. Standing up that much electricity, cooling, and networking in one place is the hard part; sourcing the processors is almost the easy half.

Raw megawatts also flatter the comparison. China’s domestic accelerators — led by Huawei’s Ascend line, with challengers such as Cambricon and Moore Threads — trail Nvidia’s current Blackwell generation on performance per watt. A gigawatt of Chinese silicon therefore yields less usable training compute than a gigawatt drawn by Nvidia systems: the lab has to burn more power, and wire together more chips, to reach the same effective throughput. Several clusters of 10,000-plus chips implies a fleet in the tens of thousands of accelerators, and at that scale the bottleneck shifts from the chips themselves to the networking and memory that tie them together — precisely where Chinese parts lag furthest behind. That penalty is the real price of going domestic.

Why Z.ai has no Nvidia to buy

For Z.ai, there is little choice in the matter. The US Commerce Department added Zhipu to its export blacklist in January 2025, cutting the company off from buying advanced Nvidia parts through legal channels. Washington has since eased some restrictions, again clearing Nvidia to sell scaled-down chips into China while barring its most powerful ones — but none of that reaches a blacklisted lab. An all-Chinese build is the only lawful path Z.ai has left.

The company has been heading this way for months. Zhipu has said it trained recent GLM models on Huawei’s Ascend accelerators using Huawei’s MindSpore software, work it described as the first major open model built on an entirely domestic stack. Washington’s sanctions, meant to slow Chinese AI, have instead accelerated the country’s drive for chip self-reliance — and the new data center is the industrial-scale version of that bet.

Beijing’s domestic-silicon bet

The build fits a wider policy push. Beijing has urged state-owned operators to fill new AI data centers with home-grown chips, part of a self-sufficiency drive that aims to source most of the country’s AI compute domestically within a few years. A gigawatt of domestic capacity at a flagship lab is the kind of proof point that policy has been reaching for.

It also lands as Chinese models crowd the frontier. Z.ai’s GLM systems and rival open models from labs such as Moonshot have won attention for matching Western systems at far lower cost, and whether China can train and serve them on its own hardware is the compute half of its broader push to set the terms of the AI era.

For now, the load-bearing details are unconfirmed. No independent party has verified the chip count, named the accelerators, or shown where a gigawatt of power is coming from — the questions that separate a working AI factory from a press-ready number. If Z.ai’s clusters are running at the scale described, China has its clearest sign yet that it can build frontier-scale compute without Nvidia. If they are not, the distance between a finished shell and a gigawatt of live training is exactly where projects like this tend to stall.