China's Z.AI has finished building a data center designed to draw roughly one gigawatt of power and has begun switching parts of it on, a facility notable less for its size than for what is inside it: only Chinese-made chips, and not a single Nvidia accelerator. The site, reported by Bloomberg on July 20 citing a person familiar with the matter, is among the largest ever stood up by a Chinese artificial intelligence lab, and it is meant to train the company's GLM family of models entirely on domestic silicon.
For a company cut off from American chips by U.S. export controls, the message is pointed. Z.AI, the Beijing developer formerly known as Zhipu, is demonstrating that a frontier-scale training run can, at least in principle, be assembled without any of the Nvidia hardware that underpins nearly every leading AI system in the West.
What was actually built
A gigawatt is enough electricity to power roughly 750,000 homes at any given moment, which is the yardstick that puts the facility among the biggest developed by any Chinese model maker. According to the person who described the project to Bloomberg, Z.AI has now built or operates several computing clusters, each holding more than 10,000 chips. The data center has started partial operations rather than running at full capacity, and it is dedicated to developing the company's most advanced GLM systems.
The source did not name the chip supplier, but Z.AI's recent history points squarely at Huawei. In June the company released GLM-5.2, an open-weight model it says was trained entirely on Huawei's Ascend accelerators, using Huawei's MindSpore software framework, with no Nvidia hardware involved. That model topped the open-weight leaderboards within a week of release. Huawei is China's leading designer of AI accelerators, competing domestically with Cambricon Technologies and Alibaba as local suppliers race to narrow the performance gap with Nvidia.
Why an all-domestic build
The all-Chinese hardware is not simply a patriotic preference; it is what remains after the doors closed. Z.AI has been on the U.S. Commerce Department's entity list since January 2025, a designation that cuts off legal access to advanced Nvidia silicon and leaves domestic parts as effectively its only supply line. Seen that way, the gigawatt site is the infrastructure response to a supply problem Washington created.
That framing matters for how the milestone should be read. Export controls were designed to slow Chinese AI progress by denying access to the fastest chips. Z.AI's data center is evidence that the policy has instead accelerated a parallel domestic stack, from Ascend accelerators up through Huawei's software, that no longer depends on American approval.
The performance catch
Raw power draw flatters the comparison with U.S. sites of similar size. Chinese accelerators such as Huawei's Ascend line trail Nvidia's current-generation Blackwell parts on performance per watt, so a gigawatt of domestic silicon delivers meaningfully less usable training compute than a gigawatt consumed by Nvidia systems. Matching a Western rival gigawatt-for-gigawatt therefore requires more chips, more floor space, and higher operating cost.
Supply is the harder constraint. Scarce domestic high-bandwidth memory limits how many Ascend-class accelerators Huawei can assemble, and the company shipped only around 812,000 AI chips last year. SMIC's most advanced stable manufacturing node, a roughly 7nm-class process, is already running above 93% utilization. As one blunt way of putting it goes: China can put up a one-gigawatt shell far faster than it can produce the chips needed to fill it.
The national picture
Z.AI's project fits inside a much larger state effort. Beijing is drafting a plan to spend roughly 2 trillion yuan, about $295 billion, over five years on a nationwide grid of AI data centers, with a mandate that at least 80% of the underlying technology come from Chinese suppliers. Alibaba and China Telecom remain among the country's largest builders of computing infrastructure, and the government's target makes clear that self-sufficiency, not efficiency, is the organizing principle.
The competitive pressure inside China is intensifying at the same time. Z.AI is expanding compute as rival Moonshot, whose Kimi K3 model has challenged systems from OpenAI and Anthropic, suspended new subscriptions on Sunday to prioritize computing resources for existing users, a sign of how tight compute has become even for the country's strongest labs. Z.AI, for its part, is on track for $1 billion in annual recurring revenue after hitting its 2026 sales target in July, and has raised billions of dollars through a Hong Kong initial public offering and a follow-on share sale.
What to watch
The open question is no longer whether China can build gigawatt-class AI facilities; Z.AI has answered that. It is whether the country can manufacture enough capable chips, and enough high-bandwidth memory, to keep those facilities running near capacity. Watch three things: how quickly Z.AI ramps the new site beyond partial operation, whether the next GLM release trained on domestic silicon can stay competitive with Nvidia-trained frontier models, and how fast SMIC and Huawei can lift chip output against a utilization rate already brushing its ceiling. Export controls were meant to buy time. Z.AI's data center is a measure of how China is spending it.