WAPR

Virginia’s regulator publishes air-pollution figures for individual data-centre sites — figures the federal government does not collect. At a site whose permit has not changed, a rise means its generators ran.

The regulator already tracks the fuel — as NOx. Reported tonnes 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.

Campuses 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.

• Only projects with a document behind them are listed here. An announcement on its own does not qualify.— register projects trace to a permit, filing, or transcript. A document can prove a project exists without proving its size, so a documented project is not a documented megawatt figure.
• The claim that these campuses install about three times as much generation as they draw from the grid does not hold up.The "3× generator-to-grid ratio for AI campuses" is refuted — JLARC 1.0–1.25×, AWS New Carlisle and Microsoft Fairwater ~1.07×. The only ~3× cases are prime turbines at off-grid sites.

Electricity demand at night across Dominion’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. 15 of 18 other PJM sub-regions flat or negative — a built-in placebo set.

• Some of this rise is weather. Hot nights push electricity use up on their own, and the later period was the warmer one — that has not been corrected for.Not weather-normalised. 3 a.m. load on a 26 °C night exceeds a 14 °C night, and Jul–Dec 2025 was warmer in the mid-Atlantic.
• The comparison regions are meant to show what normal looks like, and they do not: two of them have the same kind of load, and the rest are falling for unrelated reasons, which makes the gap here look bigger than it is.Control pool is contaminated. AEP and DAY are both Ohio and both treated; the flattest remaining controls carry behind-meter solar and manufacturing decline, which pushes the null down and inflates the effect.
• Two snapshots six years apart cannot show whether this region was already pulling ahead of its neighbours — and it was.Two six-month windows cannot test parallel trends. Dominion was already PJM's fastest-growing zone before 2019.
• Nobody has yet checked this figure against the utility’s own published data-centre load. Until it lands near that number, treat it as untested.Not yet anchored. Dominion publishes contracted data-centre load in its IRP. Until this is within ~20% of that figure, the instrument is unvalidated.

Power plants queued to connect to the Texas grid, and how much of that capacity has actually signed a final connection agreement.

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.

• You cannot currently reproduce these two numbers from a public source. The report the grid operator publishes covers power stations, not large electricity users.Not re-derivable from public MIS. Report 15933 is the generator-interconnection queue, not large load; the large-load series awaits NPRR1267 or LLWG materials. These figures stand as a dated research anchor.

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.

• At a site cooled by evaporation the heat leaves as water vapour, and vapour carries no surface temperature for a satellite to read.Latent partition. A 300 MW evaporatively cooled site rejects ~120 kg/s as latent heat — no surface temperature, thermodynamically invisible to a radiometer.
• The roof is insulated, so only one or two percent of the heat reaches the surface a satellite can see. Over most of the building the image cannot tell whether the servers are running.Roof decoupling. R-30+ insulation passes 3–15 W/m², 1–2% of the 650 W/m² rejected. Most of the polygon is blind to whether the servers are on.
• Cooling equipment is bare metal, which reflects the cold sky instead of showing its own heat. Hot machinery can read cooler than the asphalt beside it, so even the direction of the effect is not guaranteed.Emissivity. Cooling gear is finned aluminium, LWIR ε 0.05–0.30. It reflects cold sky; a 50 °C coil bank can read at or below adjacent asphalt — the sign is not guaranteed.
• Each site is compared against nearby ground that is never quite like it, and that mismatch is about the size of the signal being looked for. Measuring more sites does not shrink it.Systematic offset. Per-campus control mismatch is ±0.5–1.5 K and does not shrink with sample size — comparable to the whole-parcel signal.
• One of the two sensors covers only Los Angeles in the archive used here. Asking it about a Virginia site returns an empty result that looks exactly like a completely cloudy day.ECOSTRESS on Earth Engine holds only the LA metro. A bounds filter over Ashburn returns an empty collection with no error — indistinguishable from total cloud.
• The other sensor’s coverage band stops short of northern Europe, so the European sites named here have no free night-time thermal option at all.52°N–52°S coverage. Dublin, Amsterdam and the Nordic belt have no free night-thermal option.

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.

Turning megawatts 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.

Read this as a range, not a filled bar. The mark at the centre is the best estimate, the shaded span is where the true figure would fall 95 times in 100 — from about four times smaller to four times larger — and the ends of the track are ten times either way. Equal distances along it are equal multiples, not equal amounts.

• How much power a square foot of data-centre floor draws varies sixfold between older halls and AI halls. That one unknown is larger than every other source of error combined.Whitespace W/sq ft spans 100–600 (legacy colo to AI halls). σ(ln) = 0.545 — the single dominant term, larger than every other combined.
• Adding up many facilities does not cancel the error out. The same assumption is applied to every one of them, so it carries straight through to the total instead of averaging away.Aggregation does not rescue it. W/sq ft, PUE and utilisation are one shared assumption applied to every facility — systematic, not independent. The national total is still ×/÷3.3.
• Real floor area is known for about two percent of the facilities. For the rest it is inferred from who runs the site and which metro area it is in.Floor area exists for 246 buildings in one county — ~2% of a 10,000-facility inventory. The rest is imputed from operator and metro at ×/÷5–10.
• Deciding whether a figure belongs to one building or to a whole campus is by itself a meaningful part of the error.Attribution. Building-versus-campus assignment alone contributes σ(ln) = 0.200.

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.

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.