QTS Richmond 3
This site’s record, built from what each source attests and blank where none does.
TypeAI TrainingStatusOperationalWhereSandston, Virginia, USIndexed 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.
No organisation is recorded in any of the three control legs for this site. That is a statement about this row, not about the place: the columns below are drawn empty rather than omitted so the reader can see which questions were asked and came back with nothing.
Title
Which organisation holds the site?
Operation
Which organisation runs it?
AI hardware · per catalog
Which organisation owns the AI hardware inside it?
Power provider
No organisation is recorded in any control leg for this site, so there is no ownership chain to walk. The reach analysis was not attempted for this record — nothing is claimed either way about designations, sovereign linkages or jurisdictions above it, because there is no recorded “above” to ask about.
This site is classified Data Gap in the compute-visibility ladder — No operator, owner, or hardware-owner record linked to this facility. 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.
2 of 6 fields held · 1 outside this source’s reach · 1 genuine gap · 2 unresolved
144 MW sits at the 85th 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.