Methodology
How we verify AI data center capacity
Nobody on the open internet maintains a canonical, sourced, dated record of announced gigawatts versus actually energized gigawatts. This is that record. A campus with headline announcements but no queue position, no permits, and no turbine or transformer orders gets the honest status the evidence supports, not the one the press release implies.
Verification signals, strongest to weakest
1. Energization evidence (strongest)
Utility interconnection filings, grid-operator large-load records, and official confirmation that load is actually being served. This is the only evidence that moves a campus into the energized series.
2. Construction evidence
Building permits, air permits for on-site turbines from the EPA or a state environmental agency, county permit dockets, dated satellite imagery, and verified groundbreakings. Enough to record under-construction, not energized.
3. Financing evidence
SEC filings, loan documents, and official prospectus data. Strong for the financial-structure ledger (backstops, guarantees, equity, capex), but a signed check is not a switched-on transformer.
4. Paper evidence (weakest)
Press releases and earnings-call statements. Enough to record an announced figure at the claimed tier, and nothing more.
How current-best capacity resolves
Each of the four capacity series (announced, contracted, under construction, energized) is an append-only ledger of dated, sourced claims. We never overwrite a claim; we supersede it, so the history can be replayed at any past date. The current best value for a series is the claim with the highest evidence tier (verified, then reported, then claimed), and among equal tiers, the most recent as-of date. Superseded claims are excluded from the current value but stay on the record. The rule is deterministic and inspectable, not a matter of editorial taste.
Original claimant, not the outlet that repeated it
When a source reports an earlier claim, we record the original claimant and the original claim date as the claim itself, and record the reporting article as the access source. A 2026 news story that mentions a figure first announced in 2024 dates that claim to 2024, not 2026. We do not cite aggregator or directory sites for any headline figure: every announced capacity, capex, or chip-count claim on a campus page traces to a primary release, a filing, or tier-one reporting. Where only an aggregator carries a number, we treat the figure as unverified and record the absence rather than launder an estimate into a citation.
Absence of evidence is recorded
When a campus has an announced figure but nothing built, the energized series reads zero and says why. When a campus depends on a power plant that is itself unbuilt, we track that plant as its own project with its own lifecycle, so a campus can never show energized ahead of its power source. The empty cells are not gaps in our coverage; they are the finding. This candor is the point.
Derived figures, and how we compute them
A few figures on a campus page are computed, not sourced, so that a reader or an assistant has a plain-language comparison to quote. Each is labeled and its formula lives here.
- Homes-equivalent. One gigawatt powers roughly 810,000 US homes, from the EIA average household use of about 10,800 kWh per year (about 1.23 kW continuous): 1,000,000 kW divided by 1.23 kW. A data center runs at a high load factor and a home does not, so this is a size comparison, not an equivalence.
- Cohort standing. Rank and percentile are computed across the reviewed campus cohort by announced capacity, energized capacity, and the announced-to-energized gap, recomputed as claims change. A percentile is the share of the cohort at or below that campus.
- Compute-class estimate. Where a campus has a sourced chip-count claim, we multiply it by the chip's published FP16 throughput for an order-of-magnitude training-compute figure, expressed relative to a GPT-4-class run. It is derived from claims, not measured, and stated to one significant figure.
Satellite construction evidence
Some campus pages carry a dated strip of satellite imagery of the site. It is Sentinel-2 true color from the European Copernicus program, 10 m resolution, cloud-free scenes, one per year, oldest to newest. Each frame is captioned with its date and the required attribution, "Contains modified Copernicus Sentinel data." It shows site-scale change (bare land, grading, building mass), not fine detail.
We render imagery for a campus only after a person has verified its precise site coordinate against the picture. Most records carry a city-level location, which would point the camera at the wrong place, so those campuses show no imagery rather than a misleading frame. Absence of a strip is not a claim that nothing is being built; it means we have not yet verified where to look.
Hard rules
- Primary sources over aggregators. We cite the filing, the permit, or the prospectus, not a site that repackages them.
- We never reproduce paywalled article text. We extract the fact and cite the source.
- Every rendered number traces to a source row. If you can see a figure, you can see where it came from.
- Estimates are labeled estimates, with the method disclosed.
- When sources conflict, we show the conflict rather than pick one silently.
Back to the campus index.