Z.ai, the Beijing-based artificial intelligence company formerly known as Zhipu AI and spun out of Tsinghua University, has brought a 1-gigawatt AI data center online using exclusively Chinese-made chips — no Nvidia hardware included. The facility, partially operational as of July 2026, is the first confirmed case of a frontier AI lab scaling its training infrastructure to this magnitude on a fully domestic semiconductor stack.
Key Highlights
- Z.ai's 1GW facility houses multiple clusters of more than 10,000 Chinese accelerators each
- The entire operation runs on Huawei Ascend chips using Huawei's MindSpore software framework
- US export controls placed Zhipu on the entity list in January 2025, cutting off legal Nvidia access
- GLM-5.2, released June 2026, was trained on this infrastructure and topped open-weight leaderboards within a week
- Daily token usage for GLM-5.2 jumped 27x during its first week after launch
- Z.ai stock climbed nearly 20% on the day the facility was reported
A Milestone Built by Exclusion
When the US Commerce Department added Zhipu to its export blacklist in January 2025, cutting off the lab's legal access to Nvidia's most advanced silicon, the company made a consequential decision: rather than slow down, it scaled up on what was available domestically.
The result is a 1-gigawatt compute cluster — a facility consuming as much electricity as 750,000 homes — built entirely on Huawei's Ascend accelerators and running Huawei's MindSpore machine learning framework. First reported by Bloomberg on July 20, 2026, the facility operates multiple independent computing clusters, each containing more than 10,000 chips. Z.ai has not officially commented, but the scale of the infrastructure is confirmed through reporting from sources close to the company.
The data center represents both a technical achievement and a political statement: proof that frontier-scale AI compute does not require Nvidia, even if the road there is longer and more expensive.
GLM-5.2: The First Proof
The data center's first major output is GLM-5.2, an open-weight language model released in June 2026 under the MIT license. With 744 billion parameters in a Mixture-of-Experts architecture, the model topped open-weight leaderboards within a week of its release. Z.ai describes it as the first major frontier model built entirely on a domestic Chinese compute stack — no foreign silicon at any stage of training.
GLM-5.2 offers a one-million-token context window, selectable reasoning modes, and enterprise-grade natural language capabilities, while costing roughly a fifth of comparable US models per token on third-party APIs. The model's rapid adoption — 27x growth in daily token usage in its first week — suggests enterprise demand for a sovereign, open-weight alternative to Anthropic, OpenAI, and Google is real and growing.
The Performance Caveat
Industry analysts are quick to note that a Chinese-chip gigawatt and an Nvidia gigawatt are not equivalent measures of compute. Huawei's Ascend accelerators lag behind Nvidia's current-generation Blackwell chips on performance per watt. Huawei shipped roughly 812,000 AI chips total in 2025, which means filling a full gigawatt of installed capacity with domestic silicon is a slower and more expensive undertaking than building with Nvidia hardware.
That caveat, however, does not diminish the significance of what Z.ai has demonstrated. The facility proves that frontier-scale AI training is achievable on an entirely domestic Chinese hardware stack, at a magnitude that would have seemed implausible just 18 months ago. The key question is no longer whether it can be done — it has been done.
The Bigger Picture
Z.ai's data center sits within a far broader national effort. China is reportedly planning around 2 trillion yuan — approximately $295 billion — over five years to expand domestic AI data center capacity. Z.ai itself is on track for $1 billion in annual recurring revenue and recently listed on the Hong Kong Stock Exchange, positioning itself as an enterprise AI supplier in the same tier as Anthropic or Cohere.
The deeper irony of the story is that US export controls, designed to slow China's AI progress, may be producing the opposite effect at the infrastructure layer. By cutting Chinese labs off from Nvidia, the restrictions have incentivized — and in some cases state-funded — the development of a parallel hardware ecosystem. The 1GW facility is the clearest evidence yet that this ecosystem is now capable of delivering frontier-scale results.
What's Next
The real test, as analysts note, is the next GLM model: can infrastructure built on domestic chips sustain a trajectory that keeps pace with US labs shipping on the latest Nvidia hardware? Z.ai has not announced a timeline for its next major release. Beijing's five-year data center investment plan suggests the build-out will accelerate regardless of the performance-per-watt gap. The next 18 months will show whether China's chip independence strategy produces models that can close — or widen — the gap with the US frontier.
Source: The Next Web