Epoch AI Estimates Huawei's Xinjiang Data Center Is 14x Weaker Than Colossus 2, Using Satellite Images
Epoch AI estimated the capacity of one of China's largest AI data centers, Huawei's Horingel campus, by counting the cooling towers on its rooftops in satellite images. The result is about 81,000 H100-equivalents, or 7% of Colossus 2 in Tennessee. Even with a new building, the campus is expected to reach only 141,000 by 2028, while Colossus 2 will exceed 1.8 million. That suggests Chinese labs will remain short of compute for at least two more years.
AI-processed from Epoch AI; edited by Hamidun News
When people debate whether China will catch up to the US in AI, they usually argue about models and benchmarks. But models grow on compute, and that compute can literally be seen from space. The Epoch AI research group examined commercial satellite images of Huawei's campus in Horingel (Inner Mongolia) and calculated how much AI capacity is actually running there. The conclusion: it is one of China's largest AI data centers, but it delivers only about 7% of the capacity of Colossus 2 in Tennessee, the largest in the world.
How the Rooftop Capacity Was Calculated
Huawei has not published data on its chip count, so the analysts worked backward from cooling. Every megawatt that servers consume turns into heat, and that heat has to be removed. Based on the type, size, and number of cooling towers and chillers on a roof, one can estimate how much heat a building can dissipate, and therefore how much electrical power is available for IT equipment inside.
The campus has three zones: West, South, and North. The North zone has noticeably fewer cooling towers, while modern AI servers draw more than ten times the power of conventional ones. For that reason, the analysts excluded this zone from the calculation. Four finished buildings with dense cooling remained, and according to the model they have up to 286 MW of IT capacity combined.
Whether the buildings were operating was judged from indirect signs: no construction activity, presence of generators and parking, visible steam above cooling towers, and colored "ghosts" on the fan blades in the image. These appear because the sensor captures the red, green, and blue channels with a slight time offset, so a spinning fan produces a glare in one color. According to Epoch's estimate, three buildings are already operating and the fourth will come online by December.
From Megawatts to Chips: 81,000 H100s
Next, the analysts used data on Huawei's chip sales and assumed a four-month gap between sale and installation in racks. The result was about 105,000 Ascend 910C and 8,000 Ascend 910B chips. Publicly disclosed information suggests the 910C delivers roughly 0.76 of an H100, while Nvidia's B200 and B300 deliver about 2.5 H100s. Converted, this comes to 81,000 H100-equivalents (90% confidence interval: 36,000 to 141,000).
To check plausibility, they compared this with total sales: by the end of March 2026, Huawei had sold 907,000 910C and 600,000 910B chips. Horingel is estimated to have used about 11% of the 910C supply as of August, a reasonable share for a single site.
Colossus 2 is roughly 14 times more powerful. But 81,000 H100e is still a lot. Epoch estimates that 10% of this capacity, about 8,100 H100e, is enough to train a model on the scale of DeepSeek-V4-Pro in a month. In a very optimistic scenario, if the entire campus were combined into a single cluster, training a model on the scale of GPT-6 Astra would take a year. The caveat: a data center's capacity is usually split across several clusters, so training one model on the full total is not always practical.
What Comes by 2028
Construction is ongoing. Besides the six buildings already erected, at least three planned ones are visible in the imagery. One is going up in the South zone, and at its current pace it is expected to launch in February 2028. If it is equally AI-focused, the campus's capacity will grow by more than 70%, to roughly 141,000 H100e (range: 65,000 to 227,000). By then, Epoch projects Colossus 2 will reach at least 1.8 million H100e, so Horingel's share will not grow.
This does not mean China as a whole will fall further behind. Huawei's sister campus in Wuhu is growing slightly faster and is projected to reach 151,000 H100e by December 2027. But the overall picture is this: Chinese AI data centers are being built at hundreds of megawatts and filled with domestic chips, while in the US each unit of energy yields 2.5 times more computing power, and the largest sites have already reached 1,000 MW.
What This Means in Practice
For developers and product managers, the takeaway is simple: Chinese models from the past two years were built on noticeably smaller compute resources, which makes their results impressive. But Epoch estimates that export restrictions on Nvidia chips will not ease over the next two years. Chinese labs will therefore most likely have to keep competing on algorithmic efficiency rather than scale. This matters when choosing models and planning: price pressure and open weights from Chinese teams will persist, while leadership in the largest training runs will probably remain with the US.
This remains an estimate. It rests on satellite imagery, a cooling model, and assumptions about delivery timelines, which is why the confidence interval is so wide.
- Nov 21, 2026Nvidia doubled profit, but shares fell 2% over the search for a new CEO
- Nov 21, 2026Nous Research fine-tuned China's Qwen3.8, and propaganda in its answers dropped from 22% to 1%
- Nov 21, 2026Samsung Wants to Buy Modine, a Data Center Cooling Supplier, for Nearly $6 Billion
- Nov 21, 2026Reuters: China's $3.2B AI Data Center Buildout Relies on Nvidia Chips as Huang Steps Down
Key Questions Answered
How the Rooftop Capacity Was Calculated?
Huawei has not published data on its chip count, so the analysts worked backward from cooling. Every megawatt that servers consume turns into heat, and that heat has to be removed. Based on the type, size, and number of cooling towers and chillers on a roof, one can estimate how much heat a building can dissipate, and therefore how much electrical power is available for IT equipment inside. The campus has three zones: West, South, and North.
What are the key details on From Megawatts to Chips: 81,000 H100s?
Next, the analysts used data on Huawei's chip sales and assumed a four-month gap between sale and installation in racks. The result was about 105,000 Ascend 910C and 8,000 Ascend 910B chips. Publicly disclosed information suggests the 910C delivers roughly 0.76 of an H100, while Nvidia's B200 and B300 deliver about 2.5 H100s. Converted, this comes to 81,000 H100-equivalents (90% confidence interval: 36,000 to 141,000).
What Comes by 2028?
Construction is ongoing. Besides the six buildings already erected, at least three planned ones are visible in the imagery. One is going up in the South zone, and at its current pace it is expected to launch in February 2028. If it is equally AI-focused, the campus's capacity will grow by more than 70%, to roughly 141,000 H100e (range: 65,000 to 227,000). By then, Epoch projects Colossus 2 will reach at least 1.8 million H100e, so Horingel's share will not grow.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.