Everything on the board is sourced. This page is the inverse: the measurements that should exist and do not — the ones a trillion-dollar buildout runs on without disclosing. Each entry says who would have the number, what we tried, and the nearest sourced stand-in. When one of these gets published, it moves to the board and this page says so in the changelog.
All four report total capex and describe the growth as AI- and cloud-driven in commentary. None breaks out an AI figure. Every "AI capex" number you have ever read is an estimate.
Nearest sourced stand-in: total capex as filed — $513.8B in the four quarters to Jun 2026. The table
A $500B four-year commitment was announced in January 2025. No party has since reported dollars actually spent, on any schedule.
Nearest stand-in: the commitment itself, and the first Abilene buildings going live Sep 2025 (Crusoe). The board's $3,961/sec Stargate counter is the commitment spread evenly — labeled as exactly that.
Google published the energy of one median Gemini text prompt (0.24 Wh, Aug 2025). Nobody publishes how many queries the world runs, so every "AI uses X% of energy per question" headline multiplies a real number by an invented one. We refused to print that multiplication.
Nearest stand-in: total data-center electricity (IEA: 415 TWh in 2024, all workloads). Sources
How hard do the GPUs actually run? No operator publishes fleet utilization. It is the single number that would settle the overbuild debate, and it does not exist publicly in any audited form we could find.
Nearest stand-ins: vacancy 1.4% and 80.4% of construction pre-leased (CBRE, H1 2026) — occupancy of buildings, not usage of chips — and the rising open-market H100 rent (+17% in a year).
ERCOT is the only US grid operator publishing a classified large-load queue (233+ GW, >70% data centers). LBNL's national queue study explicitly excludes loads. Nobody can currently say how much data-center demand is waiting to connect in America.
Nearest stand-in: the ERCOT figure, labeled Texas-only. Context
List prices are public; the prices large customers actually pay — volume discounts, committed-use deals, caching economics — are not. The real price of intelligence is a private number.
Nearest stand-in: list prices, labeled as such.
Company sustainability reports mix withdrawal and consumption, on-site and off-site. No defensible global figure exists, so the board shows one rigorous regional study instead (the Potomac basin: ~9% of annual consumptive use).
Nearest stand-in: the ICPRB study.
The existing estimates rest on AI-detector tools with known false-positive problems. We cut the metric rather than print detector output as measurement.
Nearest stand-in: bot share of traffic — 51% in 2024 on Imperva-protected sites — which measures automation, not authorship.
Satellites detect every routine flare on earth nightly. The site-level catalog is licensed, not open, so our map carries the Bakken's outline and North Dakota's own figure (173 MMcf/d flared at 95.3% capture) instead of the points. One license answer away from the full overlay.
Nearest stand-in: the flaring sources.
There is no free, redistributable US equity price feed a public site can legally display: vendor free tiers forbid it, exchange delayed data requires signed agreements, and the one exchange that permits redistribution of delayed data has no web API. This is why the board's watchlist shows sourced events, not prices.
Nearest stand-in: the event-based watchlist, from SEC filings.
An entry means: we attempted to source this number against primary documents and licensed terms, and the attempt is recorded in the methodology. "Not published" claims are falsifiable — if you can point to a primary, licensed source for any entry, it graduates to the board and gets logged. That is the standing offer.
Part of BRAIAIN — the AI buildout, measured. © 2026 BRAIAIN, a dnilgis project. Not affiliated with any company named.