An independent, deterministic record of where AI infrastructure is, who operates it, what it draws, and what no instrument can reach. Every panel names its source and shows its status. No figure on this page is produced by a language model.
AI runs in buildings full of computers. Those buildings need enormous amounts of electricity, and electricity leaves paperwork: permits, grid filings, pollution reports, equipment orders. WAPR collects that paperwork every day and publishes what it adds up to.
Each panel below is one instrument. It shows a number, where the number came from, how fresh it is, and what is wrong with it. Most panels carry a lamp in their corner: Live when its feed reported within its window, Degraded when it is running with something wrong, or when the register or one-time snapshot it reads has a stated shortfall, Blind when its feed has not reported within its window, or when a register or a count is blind by design (the panel says which), Contested when the method itself is disputed. The key directly above the instruments shows each mark. Panels also list their own known defects: FATAL breaks the method, BIAS skews the answer, GAP means something needed is missing.
If an instrument cannot see, this page says so rather than printing a zero — a quiet instrument is not the same as a quiet world. Where a blind instrument still shows a figure, it is the last one it reported, with its date, drawn muted. Most bars on this page are shares of a total, and say which total; they are not error bars. Where a panel draws a range around a mark instead, the range is labelled with what it spans — on the compute-capacity panel, a 95% interval. No number on this page is produced or estimated by an AI.
WAPR is operated by BATECH LLC, a private company. It is not run for profit, carries no advertising, and nobody pays to appear here or to be left out. It is also young: several of the daily records below began collecting in September 2026. Where a panel speaks about change over time it is either drawing on an older published figure — labelled a — or saying plainly that it cannot yet. A long-run record has to start somewhere. How this is kept independent →
In plain terms: 5 is the least serious reading and 1 the most.
| Level | Rule — first match wins, top down |
|---|---|
| 1 · CRITICAL | Any open L3-published finding of an uncontrolled agent, or a verified multi-channel critical event (two independent channels agreeing, per MV-1.8 and the TRACKER §0 escalation rule), ≤ 90 days old. |
| 2 · SEVERE | ≥ 1 open L2 candidate on a critical claim, or ≥ 2 open L2 candidates of any kind. |
| 3 · ELEVATED | ≥ 1 open L2 candidate. |
| 4 · ATTENTION | ≥ 1 open L1 watchlist item, OR more than 2 panels blind — the index degrades when the instrument cannot see. Blindness is never “all clear”. |
| 5 · NOMINAL | None of the above. |
Ladder terms are protocol MV-1.8: L1 = watchlist (one channel at p<0.01, or a BH survivor) · L2 = candidate (two independent channels at p<0.01, in band, controls flat) · L3 = published (the full gate, registration hash cited). The registers live in the repository at data/wci/, are hand-curated with provenance, and scan output never enters them directly (MV-1.0). Watchlist deltas could at most separate level 5 from 4, and only with a null-expected count beside them (MV-1.1); no null model exists yet — the archive is too young — so today they are shown as context and move nothing.
In plain terms: these sentences are produced by fixed rules applied to the instruments’ numbers — not by someone deciding what to say, and not by an AI. A rule whose inputs are missing draws no sentence; where it records why, the reason is printed in its place. A sentence whose inputs are stale is still drawn, and the status line under it says so.
An instrument, not a report. A lamp says whether the instrument behind a figure reported within its window; the key below says what else each state can mean. Beside a figure, the data clause says what period the figure covers — a figure can be checked this morning and describe last year. A panel says blind rather than zero when it cannot see.
How many AI models are being published to the public model-sharing site this instrument tracks, and how many of those are packaged to run on ordinary hardware.
Quantization mechanically precedes deployment. A safety-stripped repository is a census dimension, not an accusation — abliteration is legal and published.
huggingface.co/api/models · keylessWhether the major AI companies still publish the address lists that let a website tell their web-crawling apart from an impostor.
OpenAI's Web Bot Auth directory recomputes exp = now + 7 days on every request. Hashing the document reports a key rotation on every fetch, always falsely. The detector hashes key material only. The 7-day window is not documented publicly.
5 operators · WBA JWKS + crawler CIDRsData-centre buildings that volunteer mappers have recorded, and how often the map also records who operates them.
A declared frame, not a world total. Open the Explorer →
OpenStreetMap telecom=data_center · ODbLVirginia’s regulator publishes air-pollution figures for individual data-centre sites — figures the . At a site whose permit has not changed, a rise means its ran.
The regulator already tracks the fuel — as NOx. The reported figures are fuel burned × emission factor, filed under penalty. Year-over-year NOx at a site with no permit change is “did the generators run.” The emissions table holds the current year only and overwrites itself; this record began 2026-09-14.
VA DEQ ArcGIS layers 294 / 298 / table 292 · keylessCampuses that intend to generate their own electricity instead of drawing it from the public grid. Announced capacity is what a permit or filing carries; operating capacity is what a document says is actually running — and where a document says a plant runs but none states the megawatts, the reading is left blank rather than written as zero.
Electricity demand at night across ’s Virginia territory, against the same months six years earlier. Night is when ordinary demand falls away and steady industrial loads stand out.
Jul–Dec 2019 vs 2025. Other PJM sub-regions serve as a built-in placebo set.
Power plants to the , and how much of that capacity has actually signed a .
Large new electricity users approved to connect in Texas, against how much of that demand ever showed up.
57% of approved load has never appeared. Phantom demand — the same project shopped across utilities — quantified from the operator's own figures. Grid planners cannot currently separate real from speculative.
Can a data centre’s waste heat be seen from space? Largely no. This panel records why the method fails, so that no claim gets built on it.
Still worth measuring: within-site change over time. Differencing a site against itself cancels the offset, and the waste-heat quadrant — night-positive, day-neutral — is discriminating. The off-site claim is bounded at 0.079 K; the on-site change is not.
Landsat C2 ST_B10 · ECOSTRESS via LP DAACTurning of power into a count of AI chips. The arithmetic is possible; the answer would be wrong by a factor of several — so the reason is published instead of the number.
A facility estimated at 30 MW has a 95% interval of 8 to 118 MW. Published only as a screening bound with the interval attached — never as a point estimate, never without this list beside it.
error propagation · wapr-system-design · interconnection + disclosure + Nvidia denominatorWhat kinds of power station are actually running on the grid where each company's data centres sit. This is the electricity physically flowing in those places, not what a company says it buys.
Unusual conditions happening right now near the mapped sites. These are things happening to the sites, not things the sites cause.
Deliberately not scored. This list is never totted up into a severity total — many things happening at once is what ordinary chance looks like, and level-setting belongs to the WAPR Condition rules alone.
Ireland is the only country that officially counts data-centre electricity as its own category. Everywhere else the sector disappears into general business use.
People expect a data centre to warm the land around it, and to see that heat spreading on a map. Drawn at true size, there is no plume to draw — this panel is the honest version of that picture.
Before a tall crane goes up, someone has to file a public notice with the aviation authority — so the paperwork appears months before a building does.
A watchlist signal, not a finding. It suggests where to look and settles nothing on its own.
Things no public instrument can currently read, each with the precise reason it is out of reach and the record that shows it.
The most defensible output of this instrument is a precise account of what no instrument can reach. It cannot be scooped by a commercial database, because the commercial databases are part of the finding.
| What is unknowable | Mechanism | Evidence |
|---|---|---|
| Loading the register… | ||