What happened
Epoch AI’s directory entry, updated October 30, 2026, describes VNET Bayin Ulanqab as an operational AI data center in Ulanqab, Inner Mongolia, China. Epoch estimates current capacity at 123,000 H100-equivalents, or 2.43 × 10^20 8-bit operations per second, supported by 221 MW of IT power, with a capital cost of $8.4 billion. It estimates that the campus uses Huawei Ascend 910C chips. The page labels ByteDance as the speculative chip owner and user, rather than presenting those relationships as confirmed.
The estimate is not a count reported by the operator. Epoch says its timeline for IT power and AI compute capacity is based on satellite imagery, a model of cooling-equipment power, chip availability and chip efficiency. It describes the campus as large-language-model oriented and as handling training-related requests, citing VNET’s description. The page identifies the site as part of VNET’s broader “10 GW Hyperscale 2.0” project to build three gigawatt-class clusters by 2036.
Epoch’s listed timeline starts with land clearing in late March 2024 and a signing ceremony involving VNET, the Ulanqab municipal government and Shandong Hi-Speed Holdings. Satellite imagery showed five buildings under construction by February 2025. Epoch estimates that building 1 became operational in July 2025 after a roughly three-month fit-out following structural completion; by November, it estimates buildings 2 and 5 were operational, bringing the campus estimate to 94,000 H100-equivalents and 171 MW. It places building 4’s operation in December 2025, at which point the estimate reached 123,000 H100-equivalents and 221 MW.
The directory also details the cooling equipment behind its power inference. It lists 394 air-cooled chillers with a combined current capacity of 208.5 MW, although the individual chiller manufacturers are marked unknown. Epoch lists 64 units at 94.2 MW, 240 at 80.6 MW, 24 at 18.3 MW, 60 at 11.6 MW and six at 3.7 MW. Those figures are part of the cooling-equipment model rather than an operator disclosure of installed server capacity. The distinction matters because the H100-equivalent figure is a normalized estimate of compute performance, not a claim that the campus contains 123,000 Nvidia H100 GPUs.
As of imagery dated May 6, 2026, Epoch reports that building 9 had completed all its chillers but still lacked a roof. It says the prior building timelines would imply operation by the end of 2026. The directory projects the nine-building section to be fully operational by late 2027 or early 2028, with 388,000 H100-equivalents, 700 MW of IT power and a $26.5 billion cost. That projection assumes completion timelines inferred from satellite observations: the eastern cluster of buildings 1 through 5 took roughly a year each from clearing to operation, while the western cluster was cleared in late 2024 but was not structurally complete as of mid-2026. Buildings 10 and 11 are expected later, with no specific completion date listed.
What the archive adds
New Runtime’s earlier coverage on November 11, 2026, examined what Huawei’s Horinger data center reveals about China’s compute gap, citing an estimate that Horinger had about 7% of Colossus 2’s compute capacity and could add 70% more by early 2028. This is a new fact on the broader story of China’s large-scale AI compute infrastructure: Epoch’s entry estimates capacity and expansion at a different site, VNET Bayin Ulanqab.
Trend signal
This does not yet fit a pattern New Runtime tracks. The entry documents the capital and power footprint required to supply frontier-scale compute, but it does not establish how that infrastructure changes the cost per completed task, agent interoperability, runtime security, or organizational deployment.
Why it matters
The estimate gives teams and infrastructure watchers a way to compare the scale of Chinese AI compute sites using a common H100-equivalent measure, while also making visible the power and capital behind that capacity. It is particularly useful as a structured comparison point: the entry separates estimated current compute, IT power, capital cost, construction status and projected capacity instead of treating a campus as one undifferentiated number.
The caveats are important. H100-equivalents translate estimated performance into a comparison unit; they do not mean the site contains that number of H100 GPUs. A campus using Ascend chips may differ from an H100-based system in hardware, software stack and realized workload performance even where an equivalence calculation is useful for scale. The projected expansion is also contingent on construction and deployment assumptions, not a confirmed delivery schedule. For buyers, developers and policymakers, treating such estimates as precise inventories could lead to false comparisons. The more durable lesson is to track capacity, power, cost and confidence separately rather than collapsing them into a single headline number.
Limits and uncertainty
Epoch’s figures are estimates, and the page does not establish an independently verified chip count or confirm that ByteDance owns or uses the chips. Its projected 2027 capacity depends on inferred construction timelines and assumptions about chip availability and efficiency. The $8.4 billion current cost and $26.5 billion projected cost are Epoch estimates; the page does not establish what has been spent or disclose a project budget from VNET. Actual compute utilization, delivered training throughput, network capacity and the workload mix are not reported. The comparison unit also cannot show whether the site’s usable performance matches another campus on a particular model-training or inference workload.
Reader takeaway
Epoch AI estimates that Bayin Ulanqab is already a very large AI compute campus, with a projected expansion that could more than triple its H100-equivalent capacity. Treat those figures as modeled estimates, not a verified GPU count or proof of who operates the facility. For infrastructure comparisons, keep the reported power, modeled compute and future buildout distinct.