Zhipu AI is building a 1-gigawatt data center powered entirely by domestic Chinese chips, according to Pandaily. The Tsinghua University spin-off also acquired compiler startup Zhongke Jiahe and is exploring custom AI chip development.
Key facts
- 1GW data center powered entirely by domestic Chinese chips
- Acquired compiler startup Zhongke Jiahe
- Exploring custom AI chip development
- Zhipu AI competes with Anthropic in foundation models
- GLM-5.1 and GLM-5.2 models gaining traction
The Beijing-based AI company, which develops the GLM series of large language models, is accelerating its decoupling from the Nvidia ecosystem with three parallel infrastructure moves.
1GW All-Domestic Data Center
Zhipu AI is constructing a 1-gigawatt data center that will run entirely on domestic Chinese chips. The 1GW facility would rival the largest U.S. AI data centers from Google and Microsoft [per Pandaily]. The company did not disclose the facility's location, construction timeline, or which domestic chip suppliers it is using.
Compiler Acquisition and Custom Chips
The acquisition of Zhongke Jiahe, a compiler startup, signals Zhipu AI's intent to optimize its software stack for non-Nvidia hardware. Compilers translate high-level AI frameworks into efficient machine code for specific chips — a critical piece of infrastructure that Nvidia's CUDA ecosystem has long dominated.
Zhipu AI is also exploring custom AI chip development, though no timeline or tape-out target has been disclosed. The company did not name any chip design partners or fabrication plans.
Strategic Context
Zhipu AI competes with Anthropic and other foundation model labs [per knowledge graph]. Its GLM-5.1 and GLM-5.2 models have gained traction in China's enterprise market, which is standardizing on mixture-of-experts architectures with wide expert parallelism [per recent reporting].
Full decoupling from Nvidia's CUDA ecosystem remains years away for most Chinese AI labs. The H100 and Blackwell architectures still dominate training workloads globally, and Nvidia shipped hundreds of thousands of Grace standalone servers as recently as July 2026 [per knowledge graph]. But Zhipu AI's infrastructure bet parallels moves by other Chinese AI players to build domestic supply chains amid ongoing export controls.
The Unique Angle
The 1GW figure is what matters most: it signals that Zhipu AI is not merely hedging against Nvidia export restrictions but is planning for a scale of compute that would make it a hyperscaler-class operator in its own right. Most Chinese AI labs run on far smaller clusters. A 1GW facility, if realized, would put Zhipu AI in the same infrastructure tier as the largest U.S. cloud providers.
What to watch
Watch for the data center's construction timeline and which domestic chip supplier Zhipu AI selects. Also track whether Zhipu AI discloses a custom chip tape-out date — that would signal a multi-year engineering commitment. Competitor response from other Chinese labs like Baidu and Alibaba on their own infrastructure plans.
Source: pandaily.com
