Huawei’s AI infrastructure ambition

Huawei’s recent deployments are a good starting point for understanding its plans for AI infrastructure. Throughout 2025, Huawei chose several data center sites across China to deploy its Ascend 910C, the most capable domestic AI chip to be produced in large volumes that year. We have identified four sites owned by Huawei itself, and four others owned by China Telecom, China Mobile, China Unicom, and the Shenzhen Longgang government.¹ All of these locations appear to have at least one CloudMatrix384 “supernode,” which consists of 16 server racks with a total of 384 Ascend 910C chips.

The Ascend 910C is a powerful chip, but it falls short of the best US-designed chips from Nvidia in both performance and efficiency. Huawei has not definitively reported its performance specifications, but evidence points to a theoretical peak performance of 1.504 × 10¹⁵ operations per second (OP/s) in 8-bit integer format.² That’s about 0.76 H100-equivalent (H100e) units, since the Nvidia H100 performs up to 1.979 × 10¹⁵ OP/s in 8-bit format. The H100e unit is based on the chip’s theoretical maximum operations per second in any 8-bit format, and thus doesn’t capture every aspect of performance for AI training and inference.³

Nvidia’s B200 and B300 chips perform far better in 8-bit format than the Ascend 910C, at 5 × 10¹⁵ OP/s, or about 2.5 H100-equivalents. They are also about 2.5 times more energy-efficient. So although China has abundant energy for data centers, the efficiency gap creates a relative disadvantage for their home-grown AI infrastructure, as a given amount of power capacity can only support 40% of the compute of leading US facilities.^4,5

Huawei is nevertheless capitalizing on China’s energy advantage to pursue an ambitious expansion of its AI infrastructure. We estimate that across all Huawei data centers, 12 buildings are likely to be running Ascend 910C chips today, with many more planned that could house next-generation chips. Early renderings of the Horinger campus show at least 27 data center buildings planned at that location alone, with one source reporting up to 50 buildings. Renderings of a similar Huawei campus in Wuhu show 45 buildings.

The Horinger campus: up to 286 megawatts of IT power capacity for AI

The Huawei Horinger campus is located within the Horinger New Area. Horinger County (Chinese: 和林格尔县) is in the Chinese autonomous region of Inner Mongolia, and the New Area is part of the “East Data West Computing” initiative launched in 2022. Satellite imagery shows that villages have been razed to make way for data centers there:

The Huawei campus is split into three parts near the perimeter of the New Area, a configuration dubbed “same-city three centers” (Chinese: 同城三中心). These serve as three availability zones for Huawei’s general cloud compute offering: West, South, and North. The campus is unusually specialized for AI, with “hypernode” servers that are primarily designed for large-model training and inference. Hypernode servers are only available in Huawei’s Southwest-Guiyang 1 (Gui-an), North China 3 (Horinger), and East China 2 (Wuhu) regions. It’s unclear how much of the campus is specialized for AI, but this suggests a large portion.

We acquired commercial high-resolution satellite images of these zones to help estimate the power capacity of the buildings. Our primary method is a cooling equipment model: based on the type, size, and number of cooling units visible, we estimate their heat rejection capacity and, in turn, the buildings’ IT power capacity.

The satellite imagery revealed that each zone contains nearly identical three-story buildings, which match the published renderings. However, the North zone has far fewer cooling towers than the West and South zones, suggesting that its buildings are unlikely to house powerful AI chips. Less cooling for the same-sized building suggests a lower power density, and modern AI servers tend to be more than 10 times as power-dense as the industry average.⁶

Excluding the low-density North zone, we were left with four apparently complete buildings with dense cooling infrastructure: two in the South and two in the West. Our cooling model estimates a total of 286 megawatts (MW) of IT power capacity across these four buildings. Estimates based on other sources are in a similar range: 304 MW from substation details and 230 MW from backup generator capacity.⁷

Even if the campus supports this power capacity, how can we tell whether the buildings are operational? This is hard to establish with certainty, but we’ve observed several clues. As a prerequisite, the buildings appear more or less complete from the outside, including cooling, backup generators, and carparks, with no signs of ongoing construction. Second, one satellite image shows what appears to be condensed water vapor above at least one evaporative cooling tower at one building, suggesting the building is operating. Third, multiple cooling tower fans have colored “ghosts” in another image, which indicates that the fans are moving.⁸ Further verification may be possible using thermal satellite imagery.

Only one of the four buildings, in the West zone, has yet to show any signs of operation. A public disclosure implies it was complete but not yet operational as of September 2026. We estimate it will become operational by December, while the other three are already up and running.

As we’ll discuss in the next section, it’s unlikely that 100% of the 286 MW power capacity is devoted to Ascend 910C chips, given the limited availability of AI chips in China. High power density alone does not guarantee that AI chips are present. Nevertheless, our best estimate is that most of the capacity is installed, and most of the installed chips are Ascend 910C.

AI compute: 7% of the global frontrunner

The physical scale of the Huawei Horinger campus is impressive, but its compute capacity is more important for assessing China’s AI capabilities. Calculating compute based on power is straightforward: divide the power capacity by the power per server, then multiply that by the performance per server. However, we need to account for how many racks are occupied, and how power capacity is allocated to different types of servers.

Filling an entire data center campus with one type of chip is not unheard of,⁹ but a mixture of chips is more common, especially for public clouds serving a variety of customers and applications.¹⁰ Furthermore, some portion of the Huawei Horinger campus may remain idle while the facilities await additional chips. The reported rack occupancy rate of the entire Horinger area was about 76% as of September 2026. In the absence of more specific information, we assumed Huawei Horinger has the same rack occupancy rate.¹¹

It is possible but unlikely that Huawei Horinger uses export-controlled Nvidia chips. As the champion of China’s domestic AI chip industry, Huawei has a strong incentive to deploy its own chips rather than Nvidia’s. Horinger is also a public cloud deployment, making it more open to scrutiny than private deployments and increasing the risk that any use of export-controlled chips would be discovered.

Assuming no export-controlled chips are involved, we allocated the power capacity of the occupied racks based on Huawei chip sales over time, assuming a four-month lag from sale to deployment. This resulted in a mix of Ascend 910C and 910B chips in the first building (mostly 910C), and 100% 910C in the other three buildings.¹² Given that one building is not operational yet, we find that about 105,000 Ascend 910C chips and 8,000 Ascend 910B chips are operational to date. That corresponds to 81,000 H100-equivalents (90% CI: 36,000 to 141,000). Full calculations are available in this spreadsheet.

We believe this estimate is plausible given what we know about total chip production and deployments at other data centers. According to the chip sales estimates, 907,000 Ascend 910C and 600,000 Ascend 910B chips had been sold by the end of March 2026. Satellite imagery indicates that the third AI building at Horinger was operational by August 2026. With a four-month lag between sale and deployment, our estimate suggests Huawei Horinger had used 11% of the total 910C stock by August.

We know of three other Huawei campuses reportedly deploying CloudMatrix384 systems (which contain Ascend 910C chips): Wuhu, Gui’an, and Ulanqab. There are also at least four reported non-Huawei campuses using the chips. The Huawei Gui’an campus had reportedly deployed over 110,000 Ascend chips as early as November 2024, likely 910Bs, which demonstrates that Huawei deploys large portions of its own chip supply to its data centers. Even if the Ascend 910C deployment is spread over many more unreported sites, it’s plausible that Horinger is one of the largest deployments.

At 81,000 H100-equivalents, Huawei Horinger is a large AI data center by today’s standards, and probably one of the largest in China. We have done a broad search of Chinese AI data centers across major hubs, and reviewed several top contenders in depth. Of those, only VNET Bayin Ulanqab and Huawei Wuhu are likely to match or slightly exceed Huawei Horinger.

However, Huawei Horinger is much further from the largest AI data center in the world. That title goes to SpaceX’s Colossus 2 in Tennessee. Colossus 2 has about 14 times as much computing capacity as Huawei Horinger.

This comparison comes with some caveats. First, the AI compute in a data center campus is often separated into multiple densely-networked clusters, and the total capacity cannot necessarily be combined to train one AI model.¹³ However, large R&D experiments could be run even if the compute at Huawei Horinger is split into many clusters. Just 10% of Huawei Horinger’s estimated AI compute, at 8,100 H100e, is enough to train a DeepSeek-V4-Pro scale model in one month.¹⁴ In a very optimistic scenario, if all of Huawei Horinger’s current compute was combined for a single training run, a model the scale of GPT-6 Astra would take one year to train.¹⁵

Next phase: 70% more computing power by early 2028

The buildout at Huawei Horinger is far from complete. As discussed above, one of the four buildings has yet to become operational, and at least three new buildings are planned beyond the six already built, based on renderings and construction sites seen in satellite imagery.

One of these new buildings is already underway in the South zone, and based on the pace of construction, we expect it to be operational in February 2028.¹⁶ Assuming the new building is as AI-focused as the existing buildings, the campus could grow its total AI compute by over 70% by early 2028.

This compute growth will come from adding power capacity, not from improvements in chip efficiency. One public announcement said the new building will support 100,000 AI accelerator cards and deploy Huawei’s next-generation Ascend 950 chips. However, the Ascend 950 appears to be less energy-efficient than the Ascend 910C.¹⁷ Huawei plans to ship hundreds of thousands of Ascend 950 chips by the end of 2026, starting with the 950PR variant before ramping up production of the 950DT. Since we expect this building to become operational by early 2028, we assume it will have about 76,000 950DT chips (assuming 76% occupancy rate), requiring 135 MW of IT power.¹⁸

The electric grid appears capable of handling this expansion. Plans call for five transformers to be added to the existing three at each zone substation. Assuming N+1 redundancy, this means the power capacity of the whole campus could eventually increase by up to 3.5 times.

With over 70% more computing power by the start of 2028, Huawei Horinger would reach about 141,000 H100e (90% CI: 65,000 to 227,000). But this would still leave it at about the same share, or a smaller one, of Colossus 2’s capacity, which we expect to reach at least 1.8 million H100e by 2028.¹⁹ This does not necessarily mean the overall frontier of Chinese data centers will fall further behind. For instance, Huawei’s sister campus in Wuhu will grow slightly faster, reaching a projected 151,000 H100e by December 2027.

Huawei Horinger likely serves a diverse range of users

Although we believe the Huawei Horinger data center is primarily filled with AI chips, it likely has a diverse set of end users and applications. One local announcement describes the second building at the West zone as enhancing capabilities for data aggregation and fintech service hosting. Huawei itself is a major customer of Huawei Cloud,²⁰ and allocated about 60,000 H100e to driver-assistant software as of April 2026.²¹

Huawei Horinger likely also supports the training or deployment of large language models. One possible customer is Zhipu (Z.AI), which reportedly served its GLM-5.3-Flash model on 100,000 Chinese chips in August 2026. Huawei is a leading domestic AI chipmaker, making its chips a likely choice for Zhipu. However, we don’t know where these chips were hosted. Another potential customer is iFLYTEK, which has reportedly set up operations in the Horinger New Area.

What this means for Chinese AI

China appears to have no shortage of energy for data centers. Many AI-optimized facilities are being built with hundreds of megawatts of capacity and filled with domestically-produced chips. However, AI data centers in the US support 2.5 times more AI compute per unit of power, and the largest among them have already reached 1,000 MW. The Huawei Horinger data center is a key example of this gap, with just 7% of the computing capacity of Colossus 2 in Tennessee. This makes it even more impressive that China has produced such capable AI models in the past two years. But China’s compute constraints are unlikely to ease in the next two years, which suggests that their AI capabilities and deployment capacity will continue to lag behind those of the US.

Note: some sources are withheld from this article to protect their availability. Those sources are available to researchers on request at info@epoch.ai.