// blog · analysis · compute2026-08-08source: reporting and policy coverage

A gigawatt with nobody else's silicon in it

Export controls were designed to make frontier-scale domestic training impractical. A partially operating gigawatt says the binding constraint has moved from availability to efficiency.

Z.ai has completed a roughly 1-gigawatt data centre built entirely on Chinese-made accelerators, now partially operating, running multiple clusters of more than 10,000 chips each. Zhipu shares rose 37 percent.

What a gigawatt is and is not

It is roughly the draw of 750,000 homes, and it is a statement about power infrastructure rather than delivered training compute. Installed capacity is not throughput. Domestic accelerators still carry interconnect and software-maturity penalties that no facility tour reveals, and the site is described as partially operating — which is the honest state of most buildouts at announcement.

So the correct claim is narrow and still significant: the sufficiency question is settled. Whether domestic silicon can host frontier-scale training at all is no longer the interesting question. How efficiently it does so is.

The policy underneath

A draft $295 billion plan would build a national AI data centre network on 80 percent homemade chips by 2028, and nine domestic accelerators were added to the secure-and-reliable government procurement list for the first time.

The procurement listing is the more consequential of the two. A construction figure is a projection and projections slip. A certified list is a standing order that makes domestic parts the default for government buyers regardless of how they compare to restricted imports.

The constraint that decides the 2028 target is fabrication, not design or capital, and eighty percent domestic silicon at national scale needs wafer output that current local capacity does not obviously support.

Substitution, not denial

NVIDIA's own position has already moved. Jensen Huang has said the company has largely conceded China's AI accelerator market to Huawei. That reframes what export controls achieved: not denial of capability, but the construction of a parallel supply chain that will not go away when the policy does.

Tom's Hardware — Z.ai powers up a 1-gigawatt AI data center built entirely on Chinese chips → · South China Morning Post — Zhipu shares surge 37% as firm builds giant data centre powered by Chinese chips → · Tom's Hardware — China drafts $295 billion plan to build a national AI data center grid running on 80% homemade silicon →