Huawei's Horinger data centre has about 7% of the AI compute of SpaceX's Colossus 2
Epoch AI estimates the campus at about 81,000 H100-equivalents, among China's largest. Colossus 2 has about 14 times as much.
Epoch AI studied Huawei’s AI campus in Horinger, Inner Mongolia, from commercial satellite images and public sources. Four dense buildings, three running and a fourth expected by December 2026, are estimated at up to 286 MW of IT power from the cooling equipment on their roofs. Epoch counts about 105,000 Ascend 910C and 8,000 Ascend 910B chips in operation, equal to about 81,000 Nvidia H100-equivalents (90% interval: 36,000 to 141,000).
That makes Horinger one of the largest AI data centres in China. Epoch found only two likely to match it, VNET Bayin Ulanqab and Huawei’s own Wuhu campus. The world’s largest, SpaceX’s Colossus 2 in Tennessee, has about 14 times as much compute.
The gap comes from the chips. An Ascend 910C delivers about 0.76 H100-equivalents, against about 2.5 for Nvidia’s B200 and B300, which also use about 2.5 times less energy for the same work. A fifth building, expected in early 2028, would lift Horinger to about 141,000 H100-equivalents. Epoch expects Colossus 2 to reach at least 1.8 million by then, so the share would stay the same or shrink.
What it means
China has the electricity and the buildings. Chips are the limit, and US export controls keep Nvidia’s best out of reach. Epoch concludes that Chinese compute constraints are unlikely to ease within two years. Ten percent of Horinger, 8,100 H100-equivalents, is still enough to train a model of DeepSeek-V4-Pro’s size in a month, by Epoch’s calculation.
Caveats
- Everything is an estimate from satellite images, a cooling-equipment model and public sources. The 90% interval runs from 36,000 to 141,000 H100-equivalents, and Epoch’s data-centre directory lists 170 MW of IT power today against up to 286 MW in the report.
- Huawei has not published the 910C’s specifications. The 0.76 conversion rests on a research paper’s figures and counts 8-bit operations, which flatters a chip that lacks native FP8.
- The chip mix is assumed from Huawei’s estimated sales with a four-month lag, and rack occupancy is assumed to match the area’s reported 76%. No chip was seen directly.
- The campus likely serves several users, including Huawei’s own driver-assistance software, so not all of it trains AI models. Its compute may also be split across clusters that cannot train one model together.
- Some of Epoch’s sources are withheld to protect their availability.