Z.ai fires up 1GW data center built entirely on Chinese chips

A gigawatt is a lot of power, roughly what 750,000 homes pull at once, but that isn't the interesting bit about Z.ai's newly disclosed data center. The interesting bit is that the whole site is running on Chinese-made silicon, with no Nvidia inside, and it is being built to train frontier GLM models rather than just serve them. Bloomberg reported that Z.ai, the lab formerly known as Zhipu, has begun partially operating the hub, citing an unnamed source familiar with the matter.

For the last couple of years the working assumption in the industry has been that frontier training and Nvidia GPUs are effectively the same thing, which capped how far Chinese labs could scale under export controls. A 1GW cluster on domestic parts, running clusters of more than 10,000 chips each, is the clearest counter-datapoint yet. Zhipu was added to a US Commerce Department list over alleged ties to China's military, which cut the company off from Nvidia's H100 and A100 GPUs, and it then began collaborating with Huawei; earlier this year it said it had trained a multimodal model end-to-end on Huawei Ascend hardware. This is the same trajectory at much larger scale.

The context is Beijing's broader roughly 2 trillion yuan (about $295 billion) plan to build out domestic AI data centers, with Alibaba and China Telecom among the other big builders. Z.ai's site is the concrete proof point that plan needs, silicon that can actually carry a training workload rather than only inference.

The honest caveat is that the reporting is single-sourced and does not name which Chinese chip vendor actually powers the site (Huawei, Cambricon and Alibaba's accelerator arms lead the domestic field), nor does it disclose the total chip count or which GLM training runs have finished there. Partial operation is not the same as a completed frontier training run, and parity between GLM checkpoints trained on domestic stacks versus Nvidia-trained peers is something to verify on the leaderboards, not take on the company's word.

What is worth watching is downstream. If this cluster actually produces a competitive next-generation GLM, the export-control lever loses a lot of its bite, and vendors like Huawei stop being an inference story and start being a training story. That is the shift a US policymaker or an Nvidia account executive should be pricing in this week.