Z.ai completes a 1-gigawatt AI data center built exclusively on Chinese processors, sidestepping US export-control restrictions.

Z.ai, the Chinese AI developer formerly known as Zhipu, has completed construction of a data center running 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 approximately one gigawatt of power—roughly the supply needed for 750,000 homes—and will train the company's GLM models without the Nvidia hardware that US export controls keep out of its reach.

According to the same person, speaking anonymously, Z.ai now operates or has built several computing clusters, each holding more than 10,000 chips. Neither the location of the gigawatt hub nor the specific accelerators it houses have been disclosed, and the account relies on a single source rather than a permit, filing, or signed power contract. Nonetheless, the direction is clear: one of China's strongest model builders is standing up hyperscale training capacity that bypasses American silicon entirely.

A gigawatt represents a significant scale, placing this campus in the same category as the largest AI sites under construction in the United States. At this level, power—not chips—becomes the binding constraint, the same pressure pushing the entire industry toward more efficient computing approaches. Standing up that magnitude of electricity, cooling, and networking infrastructure in one location represents the hardest challenge; sourcing the processors is comparatively straightforward.

However, raw megawatts tell only part of the story. China's domestic accelerators—led by Huawei's Ascend line, with competitors including Cambricon and Moore Threads—lag behind Nvidia's current Blackwell generation in performance per watt. Consequently, a gigawatt of Chinese silicon delivers less usable training compute than a gigawatt drawn by Nvidia systems: the facility must consume more power and wire together more chips to achieve the same effective throughput. Clusters of 10,000-plus chips imply a fleet encompassing tens of thousands of accelerators. At that scale, the bottleneck shifts from the chips themselves to the networking and memory infrastructure binding them together—precisely where Chinese components fall furthest behind. This penalty represents the true cost of domestic sourcing.

Z.ai has had little choice. The US Commerce Department added Zhipu to its export blacklist in January 2025, cutting the company off from legally purchasing advanced Nvidia components. Although Washington has since eased certain restrictions, permitting Nvidia to sell scaled-down chips to China while barring its most powerful systems, none of that relief reaches a blacklisted laboratory. An all-Chinese build remains the only lawful path available to Z.ai.

The company has been moving in this direction for months. Zhipu has disclosed that it trained recent GLM models on Huawei's Ascend accelerators using Huawei's MindSpore software, work described as the first major open model built on an entirely domestic technology stack. Rather than slowing Chinese AI development, Washington's sanctions have instead accelerated the country's drive for chip self-reliance—and the new data center represents the industrial-scale manifestation of that commitment.

The project aligns with broader policy objectives. Beijing has pressed state-owned operators to equip new AI data centers with domestically produced chips, part of a self-sufficiency initiative aimed at sourcing most of the country's AI compute domestically within years. A gigawatt of domestic capacity at a flagship laboratory provides the kind of validation this policy has sought.

The timing coincides with Chinese models crowding the frontier. Z.ai's GLM systems and rival open models from labs such as Moonshot have garnered attention for matching Western systems at substantially lower cost. Whether China can train and serve these models on its own hardware represents the computational half of its broader effort to shape the terms of the AI era.

For now, critical details remain unconfirmed. No independent party has verified the chip count, identified the specific accelerators, or confirmed the power source—the details separating a functioning AI factory from a headline figure. Should Z.ai's clusters operate at the scale described, China will have demonstrated it can build frontier-scale compute infrastructure without Nvidia. If not, the gap between a completed shell and a gigawatt of live training capacity represents precisely where such projects tend to stall.