Huawei's AI Cluster Is 14 Times Smaller Than SpaceX's Colossus 2

Researchers at Epoch AI measured Huawei's biggest AI supercomputer. It sits in Inner Mongolia and goes by the name Horinger. The system runs on Huawei's Ascend 910C and 910B chips. Right now it holds about 105,000 Ascend 910C chips and 8,000 Ascend 910B chips. Added together, that comes out to roughly 81,000 H100 equivalents. H100 is an Nvidia chip, and it serves as the common yardstick for comparing AI chips across vendors. Compare that to Colossus 2, the cluster SpaceX runs in Tennessee. Colossus 2 has 14 times more compute power, so Huawei's supercomputer falls well behind it. On top of that, the Nvidia B200/B300 chips used in Colossus 2 are also about 2.5 times more energy-efficient than Huawei's Ascend 910C.

You may have heard that China is closing the gap with the US on AI compute. Here is some context for that claim. Even in China, the largest AI data center still trails a single American cluster by a wide margin. What's more, several companies, among them SpaceX, Microsoft, and Google, already operate multiple clusters at this scale. A model's compute power is a basic limit on how reliable any AI agent built on top of it can be.