What happened
Epoch AI published a case study of Huawei’s AI data center in Horinger, Inner Mongolia, dated November 6, 2026. The report examines the campus using satellite imagery and public information, with a focus on how much AI compute it may support and how that compares with the largest AI data centers.
The campus is divided into West, South and North zones, which Huawei uses as availability zones for its cloud offering. Epoch says Huawei’s hypernode servers, designed primarily for large-model training and inference, are available in Horinger and two other Huawei regions. The report estimates IT power capacity from visible cooling equipment: four apparently complete, high-density buildings in the West and South zones yield an estimate of 286 megawatts. Other methods produce estimates in a similar range: 304 MW based on substation details and 230 MW based on backup-generator capacity. The North zone has less visible cooling equipment and was excluded from the AI-focused estimate.
Epoch says satellite images show signs that three of the four high-density buildings are operating. One building had not shown signs of operation; a public disclosure indicated it was complete but not operational as of September 2026. The report estimates Huawei Horinger had about 105,000 Ascend 910C chips and 8,000 Ascend 910B chips operational, equivalent to 81,000 H100s. Its 90% confidence interval is 36,000 to 141,000 H100-equivalents. To reach that estimate, Epoch assumes a 76% rack-occupancy rate for Horinger, based on the reported occupancy rate for the wider area, and allocates occupied capacity among chip types using Huawei chip sales with a four-month deployment lag.
The comparison is not simply a count of accelerators. Epoch estimates the Ascend 910C’s theoretical peak performance at about 0.76 H100-equivalents and says Nvidia’s B200 and B300 deliver about 2.5 H100-equivalents in the cited 8-bit measure and are around 2.5 times more energy-efficient. The report estimates Horinger has about 7% of the compute capacity of Colossus 2 in Tennessee, which it identifies as the global frontrunner. It also cautions that compute across a campus may be split among clusters and may not be available for a single training run.
Epoch describes further expansion at Horinger. At least three additional buildings are planned beyond the six already built, and one new South-zone building is under construction. Based on construction pace, the report expects that building to be operational in February 2028. Assuming it is similarly AI-focused, Epoch estimates campus compute could rise by more than 70% by early 2028, to about 141,000 H100-equivalents, with a 90% confidence interval of 65,000 to 227,000. The report says this projected growth comes from added power capacity, not improved chip efficiency.
What the archive adds
New Runtime’s earlier coverage described Envision’s planned 3 GW Galaxy Campus in Ulanqab, also in Inner Mongolia, targeting completion by the end of 2028 (October 31 note). That was a separate project and does not validate Epoch’s Horinger estimates. The new report adds a quantified assessment of a specific operating Huawei campus to the broader regional buildout story.
Trend signal
This does not yet fit a pattern New Runtime tracks. The report is a site-specific estimate of data-center capacity, not evidence of a durable shift in agent economics, software interfaces or runtime architecture.
Why now
The report’s timing reflects a campus that is partly operational and still expanding. Epoch combines imagery of existing buildings with construction evidence and a projected schedule for a new building, making the comparison a snapshot of current capacity alongside a near-term buildout estimate.
Why it matters
For teams assessing AI infrastructure, megawatts are not a reliable stand-in for model-training capacity. The Horinger estimate depends on the accelerators installed, their performance and efficiency, the share of power assigned to AI, and how fully the racks are occupied. Those factors can produce very different compute totals from facilities with similar power figures.
The comparison also shows why headline campus size needs a clear denominator. Epoch’s 7% figure is an estimate against Colossus 2’s compute, not a measurement of usable capacity for any particular customer or training job. Epoch notes that large campuses can contain separate clusters; aggregate compute does not necessarily combine into one system. Builders comparing cloud regions or planning workloads should therefore ask what hardware is available, how it is networked, and what portion can actually be reserved for their work—not just how large the site is.
What happens next
- The new South-zone building becomes operational around the report’s projected February 2028 date (12 months, 55%). Watch for: satellite evidence of completed cooling infrastructure and signs of operation.
Limits and uncertainty
The campus figures are estimates, not a disclosed Huawei inventory of installed and operating accelerators. Epoch infers power capacity from satellite-visible cooling equipment and uses assumptions about rack occupancy, chip mix and the lag from sales to deployment. Its confidence interval is wide, and signs of operation do not establish the exact hardware or workload in each building. The H100-equivalent comparison uses a theoretical performance measure and does not capture every aspect of training or inference performance. The report also says campus-wide compute may be divided among clusters, limiting what can be combined for an individual run.
Skill to build now
When comparing AI data-center claims, make a short assumptions table before using the headline number: power capacity, accelerator type, occupancy, share assigned to AI, performance basis and cluster availability. Mark which figures are disclosed and which are inferred. This week, use that checklist to distinguish a facility’s theoretical ceiling from compute your team could actually access.
Reader takeaway
Horinger is a major Chinese AI data center by Epoch’s estimate, but its power capacity does not translate directly into frontier-scale compute. The useful comparison is the hardware and operating capacity behind the megawatts—and how much of that capacity can be brought to bear on a real workload.