VNET Bayin Ulanqab
This site’s record, built from what each source attests and blank where none does.
TypeAI TrainingStatusOperationalWhereUlanqab, Inner Mongolia, CNIndexed fromEpoch AI Frontier Data Centers
Reading the record…
This site’s record, built from what each source attests and blank where none does.
Reading the record…
Control of a site decomposes into legs the record keeps apart — who holds it, who runs it, whose hardware is inside it — and they disagree often enough that collapsing them into one owner line is how a record starts being wrong. Everything below is drawn from what this row and the organisations above it actually carry; blanks are rendered as blanks, with what is known about why.
This record enters the platform from Epoch AI Frontier Data Centers.
Three legs, recorded separately because they answer different questions and because at least one of the sources says so in its own field definition: the catalog that names an AI-hardware owner states that the organisation it names is not necessarily the owner or operator of the facility. Where a leg is filled, the walk above it is drawn; where it is empty, the empty column is drawn too.
Title
Which organisation holds the site?
Operation
Which organisation runs it?
AI hardware · per catalog
Which organisation owns the AI hardware inside it?
Where an organisation recorded here has a control record of its own, that record — not this page — is where its full ledger lives. This page states the leg and links out; restating a company’s control claims on every site it touches would create as many copies of them as it has facilities, and they would drift.
Tenant
ByteDanceinferred
Nothing below is asserted at the facility row. Each constraint attaches to an organisation somewhere above this site, and the number of hops is part of the finding — no export-control designation appears anywhere on the walked chain.
Jurisdiction on the chain
A hop on the chain sits in China
Regimes reaching an entity there: Chinese Export Control Law (2020), MOFCOM Unreliable Entity List authority. Not a Wassenaar participant; operates its own Export Control Law (2020) and Unreliable Entity List; MOFCOM and the Ministry of Commerce administer dual-use catalogues.
This site is classified Tier 4 in the compute-visibility ladder — Linked org named, no ancestors documented. The ladder measures how much of a site’s control is documented, not how much compute it holds.
What the record holds as figures, with the estimate flag on the number rather than in a footnote. A derived figure and a measured one are indistinguishable downstream once the label is separated from the value.
Power capacity
Derived, not measured — the record flags this figure as an estimate.
IT load
The share of power reaching compute is not recorded separately.
Accelerators
No hardware inventory is recorded for this site.
A blank on this record can mean two very different things, and until they are separated they render identically. The plate below separates them: a field the source that produced this row has never reported for any record, and a field it reports routinely and did not report here.
This record was contributed by Epoch AI Frontier Data Centers, which accounts for 70 of the 4,257 facilities tracked. Each track below is that source’s measured rate for the field — how often it reports the field at all — set beside what this record holds.
3 of 6 fields held · 1 outside this source’s reach · 2 unresolved
221 MW sits at the 88th percentile of the 382 AI Training facilities that report a power figure, on a scale running to the largest measured value in that class (5,000 MW), unpadded. Read it against its universe: 73% of the 521 AI Training facilities tracked report one, so the ranking covers most of the class rather than a reporting subset.
Title, operation and AI-hardware ownership are three columns in the substrate and three different questions. They were separated because the field naming a hardware owner carries a published definition disclaiming the other two, and a page that renders them as one Owner: line reproduces the conflation in the presentation layer, where no migration can reach it.
A leg is never filled from a neighbouring leg. If title is unrecorded and operation is not, this record says exactly that.
From each organisation recorded on this site, the ownership graph is walked upward through documented edges. Legs naming the same organisation share one walk. Each hop carries the source class of the edge that reached it, and a walk that reaches no documented ancestor is recorded as terminating rather than as a one-entity chain.
Constraints are attached to hops, never to the facility. The hop number is reported with every finding because a designation on the direct operator and one on a great-grandparent are different facts.
An empty leg draws an empty column rather than disappearing, so the reader can see the question was asked. A constraint whose attachment point could not be placed on the walk is listed as unplaced and is never drawn at the site’s own position, which would overstate its proximity.
The observation plate above separates a field outside a source’s reach from a genuine gap, using that source’s measured rate across all of its records.
Some ownership edges and sovereign-capital linkages here were established with reference to subscription corporate-data sources that cannot be republished. Where that applies, what renders is the relationship class and the public primary that establishes it; no subscription-sourced record, identifier or value is served by this site.
Capacity for this row comes from Epoch AI’s catalog, whose published methodology derives estimates from satellite imagery, permitting documents and corporate disclosures. Their authority tier and source URL are stored on the row itself, so the chain back to the originating work is auditable from here.
Other sites are in the facilities directory, and the organisations above this one have records of their own.