Huawei's Horinger data center holds about 81,000 H100 equivalents, one fourteenth of Colossus 2

How big is China's compute, really? Epoch AI published one of the clearest answers yet on November 6, 2026. Its analysis of Huawei's Horinger data center, estimates about 81,000 H100-equivalents of AI compute, with a 90% interval of 36,000 to 141,000. That is roughly 7% of Colossus 2, the SpaceX cluster in Tennessee that Epoch puts at about 1.1 million.

The comparison is the news. China's largest AI campuses are real and growing, and they are a long way from the biggest American ones.

What Epoch measured

Epoch's author Ben Cottier estimates that Horinger is likely one of China's largest AI data centers. Only the VNET Bayin Ulanqab site and Huawei's own Wuhu campus are likely to match or slightly exceed it. By Epoch's reckoning, Colossus 2 has about 14 times Horinger's compute.

The campus has three zones, North, South and West, described as "same-city three centers" serving as availability zones. Six buildings are built and at least three more are planned, with renderings showing at least 27. Of four high-density buildings, three are operating. Rack occupancy across the area was about 76% as of September 2026.

Power is the cross-check. Epoch estimates campus IT capacity at up to 286 megawatts from its cooling model, with 304 megawatts implied by the substation and 230 megawatts by backup generators. Three independent routes landing in the same range is why the estimate is more than a guess.

The chips

Epoch estimates about 105,000 Ascend 910C and 8,000 Ascend 910B chips are operating at the campus. By its figures, one Ascend 910C has a theoretical peak of 1.504 × 10¹⁵ operations per second at INT8, about 0.76 of an H100. An Nvidia B200 or B300 delivers about 5 × 10¹⁵, roughly 2.5 H100-equivalents and about 2.5 times more energy-efficient than the 910C.

That efficiency gap shows in the power figures. The 910C draws about 1.56 kilowatts of IT power per chip. Epoch also reports that Huawei had sold about 907,000 910C and 600,000 910B chips by the end of March 2026, and that Horinger had used about 11% of the 910C stock by August. Most Huawei chips are therefore somewhere other than this campus.

What it could train

Epoch gives two reference points. Just 10% of Horinger's compute, about 8,100 H100-equivalents, could train a model at the estimated scale of DeepSeek V4 Pro (about 9.7 × 10²⁴ floating-point operations) in one month. An optimistic training run of a GPT-6 Astra scale model, which Epoch's median estimate puts at 1 × 10²⁷ operations, would take about 361 days on all of Horinger.

Two caveats from Epoch matter. Compute within a campus may be split across clusters that cannot be combined for one training run. And the chip mix is estimated, not confirmed; Epoch considers use of export-controlled Nvidia chips possible but unlikely.

Power is not the constraint

The surprise in the report is energy. Cottier writes that "China appears to have no shortage of energy for data centers," and Epoch notes that China's data centers average a power usage effectiveness of 1.40 against 1.49 for US ones in 2025, per the Uptime Institute. US AI data centers support about 2.5 times more AI compute per unit of power, because they run better chips, and the largest US sites have reached 1,000 megawatts.

In other words, the limit is silicon quality and volume, not electricity. That sits oddly beside the American narrative in which power is the scarce resource, and it is the same story our readers will recognize in the Anthropic gas turbine reports.

Where it goes next

Epoch projects Horinger will reach about 141,000 H100-equivalents by early 2028, with a 90% interval of 65,000 to 227,000. A new South zone building, expected in February 2028, is reportedly set to support 100,000 accelerator cards with Ascend 950 chips, an estimated 76,000 950DT chips needing about 135 megawatts. Huawei's sister campus at Wuhu is projected to reach 151,000 H100-equivalents by December 2027. Epoch expects Colossus 2 to reach at least 1.8 million by 2028.

That would still leave the gap at roughly an order of magnitude in single-campus terms. The picture changes if you count all of China's compute. Epoch's separate chip-import work, which we summarized in our piece on the Epoch chip imports explorer, tracks where AI chips are landing.

Why it matters for the export debate

The estimate lands in the middle of an argument. Al Jazeera reported on November 10 that Epoch estimates chips equivalent to 1.6 million Nvidia H100s reached China in violation of restrictions through the end of 2025, mainly via Southeast Asia. Epoch's Horinger work and its smuggling estimate answer different questions, and neither tells you how much compute China can marshal for one frontier training run.

Who uses the compute

Epoch reports that Huawei allocated about 60,000 H100-equivalents to driver-assistant software as of April 2026, linked to its Qiankun brand, and that Zhipu (Z.AI) reportedly served GLM-5.3-Flash on 100,000 Chinese chips in August 2026, though where that capacity sits is unknown. iFLYTEK has reportedly set up operations in the Horinger New Area. The campus is therefore not a single frontier lab's training cluster. It is shared infrastructure, which is another reason the 361-day figure for one huge training run is an upper-bound thought experiment and not a forecast.

Our take

Cottier's final point is the one to keep: that Horinger being so far from the largest data center in the world "makes it even more impressive that China has produced such capable AI models in the past two years." We would read the ratio as a ceiling on single-site frontier training, not a ceiling on Chinese models, and we would watch the 2027 Ascend 950 deployments and Epoch's updates for the first real narrowing.