The Pulse

Epoch AI Estimates Huawei Horinger Has Just 7% of Colossus 2’s AI Compute

Epoch AI estimates Huawei’s Horinger data center in Inner Mongolia has 81,000 H100-equivalents of AI compute, just 7% of SpaceX’s Colossus 2.

AI.info Team · Written with AI · fact-checked · 2026-11-07

China has abundant energy for data centers, but Epoch AI estimates that US AI data centers support 2.5 times more AI compute per unit of power. Its report estimates Huawei’s Horinger campus in Inner Mongolia has 81,000 H100-equivalents of AI compute, just 7% of SpaceX’s Colossus 2 in Tennessee. Epoch AI puts Colossus 2 at about 1.1 million H100-equivalents.

Epoch AI estimates 286 megawatts across four Horinger buildings

Epoch AI estimates 286 megawatts of IT power capacity across four Horinger buildings with dense cooling infrastructure. The researchers’ primary method uses visible cooling equipment to estimate heat rejection capacity and, from that, each building’s IT power capacity. Other estimates cited are 304 megawatts based on substation details and 230 megawatts based on backup-generator capacity. The North zone has fewer cooling towers than the West and South zones, so Epoch AI excludes it from the four-building estimate.

Epoch AI estimates that three of the four buildings are operating, with the remaining West zone building expected to become operational by December 2026. The report cites signs of operation including condensed water vapor above a cooling tower and colored “ghosts” from moving cooling-tower fans in satellite images. It estimates that most of the campus’s capacity is installed and that most installed chips are Ascend 910C.

The estimate assumes about a 76% rack occupancy rate

Epoch AI estimates that about 105,000 Ascend 910C chips and 8,000 Ascend 910B chips are operational at Horinger. The report converts that deployment to 81,000 H100-equivalents, with a 90% confidence interval of 36,000 to 141,000. Its calculation assumes about a 76% rack occupancy rate, based on the reported occupancy rate for the Horinger area in September 2026. The researchers allocate occupied-rack power between Huawei chips using sales estimates and an assumed four-month lag between sale and deployment.

Epoch AI’s evidence points to a theoretical peak of 1.504 × 10¹⁵ operations per second for the Ascend 910C in 8-bit integer format, or about 0.76 H100-equivalents. It says Nvidia’s B200 and B300 perform at 5 × 10¹⁵ operations per second in that format and are about 2.5 times more energy-efficient. Epoch AI says the difference means a given amount of power capacity can support 40% of the compute of leading US facilities.

Huawei Horinger could add over 70% more computing power by early 2028

Renderings and satellite imagery show at least three more Horinger buildings planned beyond the six already built, according to Epoch AI. One new building is under construction in the South zone, and the report expects it to become operational in February 2028. Assuming that building is as AI-focused as the existing ones, Epoch AI estimates that the campus could grow its AI compute by over 70% by early 2028. The report estimates Horinger would then reach about 141,000 H100-equivalents, with a 90% confidence interval of 65,000 to 227,000.

A public announcement said the new building would support 100,000 AI accelerator cards and deploy Huawei’s Ascend 950 chips. For its early-2028 estimate, the report assumes 76,000 Ascend 950DT chips at a 76% occupancy rate, requiring 135 megawatts of IT power. Epoch AI expects Colossus 2 to reach at least 1.8 million H100-equivalents by 2028. The report notes that Horinger’s AI compute may be split among separate clusters, which cannot necessarily be combined to train one model.