Z.AI Builds Large-Scale AI Training Data Center Using Entirely Domestic Chips
Taylor Wilson
Z.AI has built a 1-gigawatt AI training data center running entirely on Chinese-made chips, now partly operational for training its frontier GLM models — the first real-world proof that China's 'de-Nvidia' push can work at scale.
How big is this data center?
Installed capacity hits 1 gigawatt — enough to power roughly 750,000 homes at once.
The facility is already partly operational, training Z.AI's frontier large language model series, GLM.
Z.AI now runs multiple compute clusters, each packing more than 10,000 chips. At this scale, the center ranks among the largest built by any Chinese AI lab.
Why does this count as a milestone?
Every chip inside is domestically designed — zero reliance on Nvidia. This means → a Chinese AI lab has, for the first time, run the full training pipeline for a frontier model on homegrown hardware.
In plain terms = the biggest open question was whether China could train top-tier models without Nvidia's best GPUs. Z.AI just delivered the first working answer.
Alibaba and China Telecom remain China's largest compute-infrastructure builders. Z.AI's move signals that AI labs building their own compute is becoming a distinct trend.
Who else is in the domestic-chip race?
Huawei is the leading designer of Chinese AI accelerators, competing with Cambricon and Alibaba to fill the gap Nvidia left.
Z.AI rival Moonshot AI recently released Kimi K3, seen as competitive with frontier models from OpenAI and Anthropic.
Moonshot then paused new subscriptions, citing a need to "secure compute for existing users." This reflects a shared bottleneck: compute scarcity across Chinese AI companies.
Where is Z.AI getting the money?
Z.AI raised billions of dollars through its Hong Kong IPO and follow-on offerings.
Bloomberg reports Z.AI hit its 2026 sales target early, in July, with annualized recurring revenue on track to top $1 billion.
This means → Z.AI is accelerating on both the technology and the commercial side, giving it the funding base to keep investing in compute buildout.
What comes next?
China plans to spend roughly ¥2 trillion (about $295 billion) on nationwide data-center construction over the next five years.
Whether Z.AI's new center can keep GLM competitive with global frontier models will be the key test of the domestic-chip substitution strategy.
In plain terms = building the center is step one. The real exam is whether models trained on homegrown chips can hold their own on performance.
Content is for reference only, not financial advice.