Satellite Images Put Huawei Data Center at 7% of Colossus 2
Epoch AI counts about 81,000 H100-equivalents at Huawei's Horinger campus, one of China's largest, yet SpaceX's Colossus 2 still holds 14 times more AI compute.
Hakkı ALKANPublished 5 min read
Huawei's flagship AI data center in Horinger, Inner Mongolia, has only about 7% of the AI computing power of SpaceX's Colossus 2 in Tennessee, according to a satellite imagery study by research group Epoch AI. Epoch estimates the Huawei campus runs the equivalent of about 81,000 Nvidia H100 chips today, while Colossus 2 has roughly 14 times as much compute.
The Epoch AI report, published on November 6 and condensed on the group's Substack on November 12, is one of the most detailed public looks at how much AI hardware China's biggest chipmaker has actually switched on. The short version: China has the power plants and the land, but not enough efficient chips to close the gap with the largest US sites any time soon.
How big is the Huawei data center in Horinger?
The campus has three zones: West, South and North. Epoch bought commercial satellite images captured by Vantor on August 18, 2026, and counted the cooling units on each roof to estimate how much electricity the servers inside can use, a figure known as IT power capacity.
The North zone has far fewer cooling towers, so Epoch thinks it does not hold powerful AI chips. The four dense buildings in the South and West zones add up to an estimated 286 megawatts (MW) of IT power. Two other methods landed nearby, at 304 MW based on substation details and 230 MW based on backup generators.
Three of those four buildings appear to be running, with water vapor over a cooling tower in one image and visibly spinning fans in another. The fourth, in the West zone, is expected to come online by December. Reported rack occupancy across the Horinger area was about 76% in September.
What is an H100-equivalent?
An H100-equivalent (H100e) is a yardstick that compares any AI chip to one Nvidia H100, based on peak 8-bit operations per second. It does not capture every detail of training performance, but it lets researchers compare very different hardware on one scale.
By Epoch's count, Horinger runs about 105,000 Huawei Ascend 910C chips and 8,000 older Ascend 910B chips. That works out to about 81,000 H100e, with a wide 90% confidence range of 36,000 to 141,000. Epoch considers the use of export-controlled Nvidia chips at the site possible but unlikely, since it is a public cloud and Huawei has every reason to showcase its own silicon.
| Data center | Estimated AI compute | Timing |
|---|---|---|
| Huawei Horinger (today) | About 81,000 H100e | 3 of 4 AI buildings running, 2026 |
| Huawei Horinger (next phase) | About 141,000 H100e | Early 2028 |
| Huawei Wuhu | About 151,000 H100e | December 2027 projection |
| SpaceX Colossus 2 | About 14 times Horinger today; at least 1.8 million H100e | Today; 2028 projection |
All figures are Epoch AI estimates; the table is our summary of the report.
Why is Huawei so far behind Colossus 2?
The main reason is chip efficiency. Epoch puts the Ascend 910C at about 0.76 H100e, while Nvidia's B200 and B300 reach about 2.5 H100e and are roughly 2.5 times more energy efficient than Huawei's chip.
That means the same amount of electricity buys a Chinese data center only about 40% of the compute of a leading US facility. Scale is the second factor: the largest US AI sites have already reached 1,000 MW, several times Horinger's estimated AI power.
Supply is the third limit. Epoch's chip sales estimates suggest Huawei had sold 907,000 Ascend 910C and 600,000 Ascend 910B chips by the end of March 2026, and Horinger alone had absorbed about 11% of the 910C stock by August. US export controls, government rules that block sales of advanced chips and chipmaking tools to China, keep Huawei from buying the most efficient hardware or the newest manufacturing equipment.
Will the gap close by 2028?
Not at Horinger. A new South zone building is under construction and should be ready around February 2028, and a public announcement says it will hold 100,000 AI accelerator cards using Huawei's next Ascend 950 chips. Epoch estimates that could lift the campus by more than 70%, to about 141,000 H100e.
Colossus 2, however, is expected to reach at least 1.8 million H100e by 2028, so Horinger's share would stay flat or shrink. Epoch also notes the Ascend 950 appears to be less energy efficient than the 910C, so the growth will come from adding power rather than better chips. Huawei's sister campus in Wuhu is growing slightly faster, and Epoch's data center directory tracks both sites.
What can a site this size actually train?
Quite a lot, by normal standards. Epoch says just 10% of Horinger's compute, about 8,100 H100e, is enough to train a DeepSeek-V4-Pro scale model in a month. Putting all of today's capacity into one run would take about a year for a GPT-6 Astra scale model, though compute on a campus is usually split into clusters that cannot always be combined.
Likely users include Huawei itself, which had allocated about 60,000 H100e to driver-assistance software as of April 2026. Zhipu (Z.AI) reportedly served its GLM-5.3-Flash model on 100,000 Chinese chips in August, but Epoch does not know where those chips were hosted.
What does it mean for the US-China chip race?
The findings land in a tense week for AI chips. Earlier this month Nvidia began restricting cloud clients that serve Chinese firms, and Treasury Secretary Scott Bessent has given Nvidia 10 days to explain whether it pushed "seemingly illicit" chip sales to China, as the company searches for a new CEO.
Epoch's numbers help explain why access to Nvidia hardware matters so much to Chinese buyers. Huawei is building an all-domestic stack, from Ascend chips to the first HarmonyOS desktop, yet its best AI campus still delivers a fraction of what a single US site can. Epoch concludes that China's compute constraints are unlikely to ease over the next two years, which makes the strong Chinese models of the past two years look even more impressive.