zAI completed a 1-gigawatt AI data center in China using zero Nvidia chips, per Bloomberg. The facility runs exclusively on Chinese-made silicon and has begun partial operations.
Key facts
- 1-gigawatt power capacity for AI training
- Zero Nvidia chips used in the facility
- Runs exclusively on Chinese-made chips
- Partial operations already begun
- Supports development of GLM frontier models
zAI has reportedly completed construction of a 1-gigawatt AI data center in China without a single Nvidia chip, according to Bloomberg via @kimmonismus. The facility runs exclusively on Chinese-made chips and has begun partial operations, supporting development of the company's frontier GLM models.
zAI also operates several computing clusters containing more than 10,000 chips each, per the same report. The scale of this data center — 1 gigawatt of power capacity — rivals the largest U.S. AI clusters, such as those operated by Microsoft and OpenAI, but with entirely domestic hardware.
Chip Independence as Strategic Priority
The buildout underscores China's push for chip independence amid U.S. export controls. Nvidia's A100 and H100 chips are banned from sale to China; the company's H800 and B200 variants, designed to comply with U.S. rules, were later restricted. zAI's data center demonstrates that Chinese AI labs can train frontier models without access to Nvidia's hardware, potentially reducing the effectiveness of U.S. sanctions.
Unique Take: Infrastructure Over Model Releases
While the AI press obsesses over the next frontier model — GPT-5, Gemini 2, Claude 4 — this data center represents a more structural shift. China's ability to deploy 1GW of compute without Nvidia means that future model performance gaps may narrow not because of algorithmic breakthroughs, but because of unrestricted access to domestic silicon. The U.S. export control regime assumed that cutting off Nvidia chips would slow China's AI progress; zAI's facility suggests the strategy may be failing.
Scale Comparison
For context, the largest U.S. AI clusters — such as Microsoft's planned $100B data center — target similar power levels but rely on Nvidia GPUs. zAI's 1GW facility is comparable in scale to the entire power draw of a small city. The company did not disclose the exact chip type, fabrication process, or cost, according to the Bloomberg report.
Key Takeaways
- zAI built a 1GW AI data center in China with no Nvidia chips, using only domestic silicon.
- It supports frontier GLM model development and has begun operations.
What to watch
Watch for zAI's next GLM model release and whether its benchmark scores close the gap with GPT-4 or Claude 3.5. Also monitor U.S. export control updates — if the Commerce Department tightens restrictions on chip design tools or foundry access, it would signal concern over this development.
[Updated 21 Jul via pandaily]
zAI has also acquired compiler startup Zhongke Jiahe, a move that strengthens its ability to optimize software for domestic chips, and is exploring custom AI chip development, according to a Pandaily report. These steps accelerate its decoupling from the Nvidia ecosystem [per Pandaily].
