← Back to feed

Z.ai, the Beijing-based AI developer formerly known as Zhipu, has completed construction and begun partial operations at a 1-gigawatt data center built exclusively on Chinese-made AI accelerators. Bloomberg first reported the activation on July 20. The facility houses multiple compute clusters, each containing more than 10,000 domestic chips, and is used to train the GLM family of large language models.

No NVIDIA hardware is in the building.

What Exists

The 1GW facility is one of the largest compute deployments built by a Chinese AI lab. Its significance is less the raw power figure than the silicon profile: at the frontier of AI compute, every comparable facility outside China runs on NVIDIA H100s, H200s, or Blackwell GPUs. Z.ai’s facility runs on domestic accelerators across every cluster.

The company has not publicly identified the chip vendor. China’s leading domestic AI accelerator suppliers include Cambricon, Biren Technology, Huawei (Ascend 910B and 910C), and Enflame. Huawei’s Ascend chips have the strongest documented performance data and are the primary candidate for large-scale cluster deployments at this size.

Z.ai is not the first Chinese lab to run models on domestic silicon. But a dedicated 1GW facility — purpose-built for frontier model training with no NVIDIA components — represents a different kind of infrastructure commitment than mixing domestic chips into existing Nvidia-based clusters.

Strategic and Corporate Context

Z.ai recently acquired Zhongke Jiahe, a heterogeneous computing software firm. The acquisition suggests Z.ai is not simply using domestic chips as they come but building the software infrastructure to extract competitive performance from them — the same investment NVIDIA made in CUDA over two decades that created its current software moat.

The company is also in early talks with Chinese chip designers about co-developing custom AI chips specifically for GLM workloads. If those talks progress, Z.ai would become one of the few AI labs outside OpenAI and Google that both trains frontier models and influences the design of the hardware those models run on.

Z.ai stock (ticker: 02513.HK) rose approximately 20% following the Bloomberg report. The company is on track for $1 billion in annual recurring revenue.

The Export Control Backstory

The facility’s existence is a direct result of US export controls that have restricted NVIDIA A100, H100, and H200 shipments to Chinese entities since October 2022, with successive tightening rounds since. The intent was to slow Chinese AI development by limiting access to the most capable training hardware.

The Z.ai data center represents the scenario US export control designers wanted to avoid: Chinese labs not slowed but redirected into building domestic hardware capacity at scale. A 1GW facility running domestic chips at commercial density is not a workaround — it is a parallel infrastructure.

Perplexity AI CEO Aravind Srinivas recently observed that export controls may be creating a more capable long-term competitor: “By forcing them to go out there and build all this, you are converting them into a far more potent competitor.” The Z.ai activation is evidence of that dynamic in operational form.

The GLM Record

Z.ai’s GLM model family has posted consistent benchmark results over the past 18 months. GLM-5.2 reached #10 on Agent Arena, matched GPT-5.5 on real-world agent tasks in the open-weights Intelligence Index at score 51, and was clocked at 2,626 tokens per second on AMD MI355X hardware. GLM-5.1 hit #4 on Arena Search globally. The lab’s models have been used by Perplexity as a base for its own tuned orchestration system.

The question the 1GW activation raises is what GLM-5.3, or whatever follows, is being trained on — and at what scale. The facility is described as partially activated, with remaining clusters coming online over an unspecified timeline.