Huawei, China’s largest AI chip developer, is building large AI-focused data centers that leverage China’s abundant energy, but U.S. export controls have constrained its chips’ performance and efficiency relative to U.S. competitors such as Nvidia. Analysts examined the Huawei facility in Horinger, Inner Mongolia using satellite imagery and public data to estimate its power and compute capacity, and to project near-term growth.
Campus layout and cooling-based capacity estimates
The Horinger campus sits in the Horinger New Area and is organized into three zones: West, South, and North. Satellite imagery shows villages were removed to make way for the data centers. Huawei documentation describes the use of “hypernode” servers intended for large-model training and inference, indicating a significant portion of the campus is dedicated to AI workloads.
The investigation acquired high-resolution commercial satellite images and used a cooling-equipment model: the visible type, size, and number of cooling units were used to estimate heat-rejection capacity and then infer IT power capacity. The buildings observed are nearly identical three-story structures, but the North zone has far fewer cooling towers than the West and South zones, implying lower-density (less power-dense) servers in North.
Ignoring the lower-density North zone, the analysts identified four apparently complete, high-density buildings (two in South, two in West). Their cooling model estimates these four buildings together have roughly 286 megawatts (MW) of IT power capacity.
Evidence of operation and timeline
Several indicators suggest the buildings are operational: their exteriors show completed cooling systems, backup generators, and carparks with no active construction; one satellite image showed condensed water vapor above at least one evaporative cooling tower; and in other images the cooling-tower fans produced colored ‘‘ghosts,’’ an optical effect indicating movement. One of the four buildings (in the West zone) had not shown operational signs; a public disclosure indicated it was complete but not operational as of September 2026. The analysts estimate that building would become operational by December 2026, while the other three were already running.
Estimating compute from chip deployments
To convert physical capacity into compute, the researchers combined the facility timeline with Huawei chip-sales information, assuming roughly a four-month lag from chip sale to deployment. This produced an expected mix of Ascend chips: the first building likely contains a mix of Ascend 910C and 910B (mostly 910C), and the other three buildings are modeled as 100% Ascend 910C.
Given that one building was not yet operational, the estimate is that about 105,000 Ascend 910C chips and 8,000 Ascend 910B chips were deployed and operational to date.
For international comparison, compute is often expressed in H100-equivalents (H100e), based on the performance of Nvidia’s H100. Huawei disclosures suggest the 910C performs at around 0.76 H100e, while Nvidia’s B200 and B300 chips perform at about 2.5 H100e. The analysts’ chip estimate corresponds to roughly 81,000 H100-equivalents (90% confidence interval: 36,000 to 141,000 H100e).
The chip-sales backdrop: by the end of March 2026, an estimated 907,000 Ascend 910C and 600,000 Ascend 910B chips had been sold. Satellite imagery indicates the third AI building at Horinger was operational by August 2026; with a four-month sale-to-deployment lag, the estimate implies Horinger used about 11% of the total 910C stock by August 2026.
How Horinger compares globally
At roughly 81,000 H100e, Huawei Horinger is large within China and likely among the country’s biggest AI data centers. A broad review of Chinese AI data centers found few contenders — VNET Bayin Ulanqab and Huawei Wuhu are the closest rivals — but Horinger remains far smaller than the world’s largest site.
SpaceX’s Colossus 2 in Tennessee is currently the largest AI data center in the world; based on available estimates, Colossus 2 has about 14 times the computing capacity of Huawei Horinger, meaning Horinger holds roughly 7% of Colossus 2’s compute.
Data-center campuses often split compute into multiple densely networked clusters, so total campus capacity does not always translate directly into a single trainable cluster. Still, analysts note that 10% of Horinger’s estimated compute (~8,100 H100e) would suffice to train a DeepSeek-V4-Pro–scale model in one month. In an optimistic scenario where all current Horinger compute were pooled for a single run, training a GPT-6 Astra–scale model would take about one year.
Planned expansion: ~70% more compute by early 2028
The Horinger buildout is ongoing. One of the four buildings had not yet started operation, and satellite imagery and renderings show at least three more buildings planned beyond the six already built. Construction of one of these new buildings is underway in the South zone, and based on the construction pace it is expected to be operational in February 2028. If that building is as AI-focused as the existing ones, the campus could increase its AI compute by over 70% by early 2028.
With that growth, Horinger would reach roughly 141,000 H100e (90% CI: 65,000 to 227,000) by early 2028. Colossus 2 is expected to reach at least about 1.8 million H100e by 2028, so Horinger would still represent a small fraction of the frontrunner’s capacity. Other Chinese sites may grow at different rates — for example, Huawei Wuhu is projected to reach about 151,000 H100e by December 2027.
Implications for China’s AI trajectory
The analysis underscores that China has abundant power and is building large AI-optimized facilities populated with domestic chips. However, U.S. data centers appear to deliver roughly 2.5 times more AI compute per unit of power, and the largest U.S. facilities already operate at scales of hundreds to a thousand MW. Horinger’s example — large domestically but only about 7% of Colossus 2’s compute — highlights the structural gap. That gap helps explain how China has nonetheless produced capable AI models in the past two years, but it also suggests China’s compute constraints are unlikely to disappear in the next two years, meaning overall AI capability and deployment could remain behind the United States for the near future.
Short technical note on satellite imagery artifacts
The analysts explain that the colored ‘‘ghosts’’ from cooling-tower fans arise because satellite sensors capture red, green, and blue wavelengths at slightly different times; when a moving fan reflects light during just one of those captures, the reflection can appear as a bright color in a single channel.
