An analysis says Huawei's artificial intelligence (AI) data centre in Horinger, Inner Mongolia has computing power equivalent to about 7 percent of SpaceX's "Colossus 2", one of the world's largest AI data centres.
On Nov. 19 (local time), online media outlet Gigazine reported that research organisation Epoch AI estimated the Horinger data centre's actual computing power based on satellite images and the scale of Huawei's semiconductor transactions.
The analysis is drawing attention as a gauge of China's AI infrastructure expansion and the technology gap with the United States. Due to U.S. and Dutch restrictions on semiconductor exports to China, China faces constraints in accessing Nvidia's top-end data centre chips and ASML's key equipment. China has built AI infrastructure using export-approved semiconductors and domestically produced chips.
Epoch AI analysed the scale of the Horinger data centre focusing on Huawei's domestic semiconductor, the Ascend 910C. It identified the type, size and number of cooling facilities in high-resolution satellite images and used them to estimate the buildings' heat dissipation capacity and IT power capacity.
The satellite images show the Horinger complex is divided into three zones: north, south and west, each with similar three-storey buildings. The north zone has significantly fewer cooling towers than the south and west zones, and was analysed as unlikely to house large numbers of high-performance AI semiconductors. Signs of ongoing construction were also identified in some buildings.
By contrast, four buildings with dense cooling equipment were identified in the south and west zones. Of those, three buildings were estimated to be complete, with cooling equipment, emergency generators and parking lots. At least one building's evaporative cooling tower showed traces of condensing water vapour, and signs that cooling tower fans were operating, raising the possibility it is actually running.
Epoch AI estimated the number of chips installed in each building by reflecting Huawei's semiconductor sales volume and data centre operating conditions. It assessed that about 105,000 Ascend 910C chips and about 8,000 older Ascend 910B chips are in operation. Converted to Nvidia H100 equivalents, that is about 81,000 units. That compares with about 1.1 million H100 equivalents for Colossus 2, or about 7 percent.
It also compared computing power needed for model training. Under an optimistic assumption that the Horinger data centre's total computing power can be concentrated on a single training job, Epoch AI estimated it would take about 1 year to train a GPT-6 Astra-class model. It assessed that a DeepSeek-V4-Pro-class model could be trained in about 1 month even using only 10 percent of total computing power.
Even after the data centre is completed, it is expected to be difficult to narrow the gap with the United States in a short period. Epoch AI said that given the current pace of construction, additional buildings at the Horinger data centre are expected to begin operating in February 2028. If those buildings have AI computing equipment at a level similar to existing facilities, total computing power could increase by more than 70 percent.
In that case, the Horinger data centre's computing power is projected to reach about 141,000 H100 equivalents. But Colossus 2 is also expected to expand over the same period and reach at least 1.8 million H100 equivalents. That would mean a significant gap in computing power would remain despite Horinger's expansion.
A gap also emerged in power infrastructure. Based on the cooling facilities, Epoch AI estimated the total IT power capacity secured by Horinger's four completed buildings at 286 megawatts (MW). In the United States, there are data centres with power capacity reaching 1,000 MW.
Epoch AI assessed that it is notable China has developed high-performance AI models over the past 2 years despite semiconductor export restrictions. It said it is unlikely that China's constraints on computing resources will ease over the next 2 years, and it expects the AI infrastructure gap with the United States to persist. It said even if China continues to catch up through domestic semiconductors and data centre expansion, it will not be easy to narrow the computing gap with the United States in a short period.