Z.ai Reportedly Launches 1GW Data Center with Domestic Chips

According to Bloomberg, Z.ai has completed and partially launched a 1-gigawatt (1GW) AI data center in China. The facility is said to rely entirely on domestically produced chips, avoiding Nvidia, and will be used to train next-generation GLM models. This development is highly notable not only for its massive scale—equivalent to powering roughly 750,000 households—but also as a sign that domestic chips are entering the actual training phase for frontier models.

Controversies and Doubts

Questioning the staggering "1GW" figure, user @zacharynado pointed out that if estimated based on "multiple 10,000+ chip clusters," the actual chip load would likely be only 75 to 100MW. Therefore, the 1GW might only represent the approved power capacity or overall planning scale of the park, rather than reflecting currently deployed hardware.

Leaked Cluster Architecture Details

User @bookwormengr shared suspected architecture details of the massive center. According to them, Z.ai built an Atlas 950 SuperCluster containing approximately 524,288 Ascend chips. Architecturally, one SuperCluster consists of 64 SuperPods. Due to its massive scale and potential chip shortages, the supercluster is currently still in a "partially operational" state.

Infrastructure and Compute Deployment

Regarding the construction progress of domestic compute centers, user @teortaxesTex believes that the "completed" facilities seen by the public might not be sitting empty. Because China has strong rapid deployment capabilities in power supply and civil engineering, data centers can be built first, waiting for chips to arrive and be installed later. Additionally, recent leaked information about Huawei's Ascend 950DT chip has further drawn attention to the actual deployment progress and hardware strength of China's domestic compute.

2026-07-21 ~ 2026-07-21 · 10 related posts

Primary sources

2 near-duplicate retellings: kimmonismus · kimmonismus