Sulat.com / Writing on AIMonday, November 16, 2026
Chips/16 November 2026

Epoch AI estimates Huawei’s Horinger data centre has 7% of Colossus 2’s AI compute

Epoch AI puts Huawei's Horinger campus at about 81,000 H100-equivalents against Colossus 2's 1.1 million. Its 90% range is 36,000 to 141,000, and it expects the gap to hold through 2028.

Glossy red 3D Huawei logo sculpture on a dark table beside a glowing glass model of a four-building data centre campus with wind turbines, steppe at dusk behind

At a glance

  • Epoch AI's Ben Cottier estimates that Huawei's Horinger data centre in Inner Mongolia holds about 81,000 H100-equivalents of AI compute, roughly 7% of the 1.1 million at SpaceX's Colossus 2 in Tennessee.
  • Epoch says the estimate rests on satellite imagery and public sources, carries a 90% range of 36,000 to 141,000, and counts about 105,000 Ascend 910C and 8,000 Ascend 910B chips as running.
  • Epoch says a given amount of power supports 40% of the compute on Huawei's chips that it does in leading US facilities, because Nvidia's B200 and B300 are about 2.5 times more energy-efficient.
  • Epoch expects Horinger to reach about 141,000 H100-equivalents in early 2028 when a new building opens, still at about the same share or a smaller one of Colossus 2's capacity.
  • Epoch calculates that 10% of Horinger's compute could train a model at the scale of DeepSeek-V4-Pro in one month, though it says a campus's compute is often split across several clusters.

Epoch AI estimates that Huawei’s AI data centre at Horinger, in Inner Mongolia, holds about 81,000 H100-equivalents of computing capacity, roughly 7% of the 1.1 million at SpaceX’s Colossus 2 in Tennessee. The estimate comes from a 6 November report by Epoch senior researcher Ben Cottier, built from satellite imagery and public sources.

An H100-equivalent is Epoch’s yardstick for comparing chips: one unit is the peak 8-bit speed of Nvidia’s H100. Epoch says Colossus 2 has about 14 times Horinger’s capacity, and that Horinger is still probably one of the largest AI data centres in China. Only VNET’s Bayin Ulanqab site and Huawei’s own Wuhu campus are likely to match or slightly exceed it, Epoch says.

Two bars: Colossus 2 at 1.11 million H100-equivalents, Huawei Horinger at 81,000, about one in 14
Horinger holds about 7% of Colossus 2’s AI compute, Epoch estimates.

How Epoch sized the campus

Horinger is a three-zone Huawei campus, split into West, South and North, inside the Horinger New Area, which is part of China’s East Data West Computing programme launched in 2022. Epoch counts six buildings so far. It set aside the two in the North zone because they have far fewer cooling towers, which points to low-power racks rather than AI servers. That leaves four dense buildings, two in the South zone and two in the West.

From the number and size of the cooling units visible in satellite images, Epoch’s model puts those four buildings at up to 286 megawatts of IT power capacity. Two cross-checks land in the same range: 304 megawatts from substation details and 230 from backup generators. Epoch says three of the four buildings are already running and expects the fourth, in the West zone, to start by December. Its data page lists 170 megawatts of IT power for the site today.

To turn power into chips, Epoch assumed the same 76% rack occupancy reported for the whole Horinger area in September 2026. It then filled the racks using Huawei’s chip sales, with a four-month lag between sale and installation. That gives about 105,000 Ascend 910C chips and 8,000 Ascend 910B chips running, and the 81,000 figure carries a 90% range of 36,000 to 141,000. Epoch assumes the site holds no export-controlled Nvidia chips, because Horinger is a public cloud and Huawei has every reason to use its own silicon.

Why the same power buys less

By Epoch’s numbers, Huawei’s 910C scores about 0.76 H100-equivalents per chip, against about 2.5 for Nvidia’s B200 and B300. Epoch, citing SemiAnalysis, says the Nvidia chips are also about 2.5 times more energy-efficient. It says that means a given amount of power capacity supports 40% of the compute it would in a leading US facility. Epoch also notes that US AI data centres run at close to 100% rack occupancy, while Horinger’s area sits at 76%.

Ascend 910C scores 0.76 H100e per chip, Nvidia B200 and B300 about 2.5; same power gives 40% of the compute
Epoch says the same amount of power runs 40% as much compute on Huawei’s chips.
Epoch estimateHuawei HoringerColossus 2
AI compute81,000 H100e1,112,000 H100e
IT power170 MW946 MW
Capital costUS$6.4 billionUS$35.8 billion
ChipsAscend 910C and 910BNvidia B200 and B300
Epoch’s data pages put Horinger at 7% of Colossus 2’s compute on 18% of its power.

Epoch’s footnote adds that the 910C doesn’t natively support FP8, an 8-bit format useful for AI training, so the 0.76 figure arguably oversells the chip. Huawei hasn’t definitively reported the 910C’s specifications, so Epoch derives the number from a per-die figure in a Huawei research paper.

What 81,000 can train

Epoch says the compute in a campus is often split into several densely networked clusters, so the total can’t necessarily go into one training run. Even so, it calculates that 10% of Horinger’s compute, 8,100 H100-equivalents, could train a model at the scale of DeepSeek-V4-Pro in a month. If all of it were combined for one run, a model at the scale of GPT-6 Astra would take about a year, on Epoch’s median estimate of that model’s training compute.

What’s coming at Horinger

One more building is under construction in the South zone, and Epoch expects it to open in February 2028. A public announcement says it will support 100,000 accelerator cards and use Huawei’s next-generation Ascend 950 chips. Epoch assumes about 76,000 of the 950DT variant drawing 135 megawatts, which would lift Horinger by more than 70% to roughly 141,000 H100-equivalents, with a 90% range of 65,000 to 227,000. Epoch expects Colossus 2 to reach at least 1.8 million by 2028, so Horinger would stay at about the same share or a smaller one.

Five building icons: three running, one due December 2026, one new due February 2028; 81,000 H100e now, about 141,000 later
Three AI buildings run today; a fourth and a new one are due by early 2028.

Epoch says that doesn’t mean China’s frontier falls further behind. Huawei’s Wuhu campus is projected to reach 151,000 H100-equivalents by December 2027. Epoch’s conclusion is that China has the energy for data centres but that its compute constraints are unlikely to ease in the next two years.

What the estimate doesn’t show

Epoch doesn’t know who Horinger’s customers are. It says the campus likely serves a mix of users: Huawei itself is a major customer of Huawei Cloud, and Zhipu and iFLYTEK are possible customers, but Epoch doesn’t know where Zhipu’s chips are hosted. It withheld some sources to protect their availability and makes them available to researchers on request. The calculations are in a public spreadsheet, and Epoch’s data pages are published under a CC-BY licence.


Discover more from Sulat AI & Tech News

Subscribe to get the latest posts sent to your email.