† = Causeway estimate – see Method & caveats.

One principle. National demand data shows slopes, not levels. A load that does not vary with the weather is indistinguishable from every other constant in the system, so no amount of grid data can size it. Every level on this panel therefore comes from elsewhere — official statistics for comfort cooling, bottom-up counting for process — and the grid’s job is to test whether those levels behave as they should, not to produce them.

The response curve below centres each day within its own month and day type before fitting. That removes the holiday and seasonal confound — August demand is depressed exactly when it is hottest — but it removes the level along with it, because the monthly mean contains most of the cooling. So the published curve recovers about a third of the anchored figure and is a lower bound by construction, not a contradiction of it. That share has moved as the estimator improved — it was a seventh before the binned ceiling was replaced by a continuous fit — so it is recomputed from the payload rather than quoted from memory†. Modelling the baseline explicitly instead — constant, day type, holidays, a heating limb and a convex cooling limb, fitted on undemeaned daily data — keeps the intercept, and recovers 74% of the anchor with a balance point of 14.5 °C against the 15.5 °C the building physics gives. Two independent routes within a quarter of each other. The gap is the honest uncertainty: it could be the anchor running high, the fit running low, or the shape being wrong, and 400 days cannot tell them apart.

Tier 0 was carried for months as a bare “of order 26 TWh a year” with no derivation anywhere. It now has one, and the exercise moved the number’s meaning more than its magnitude.

The food cold chain is anchored on Foster, Brown and Evans, Carbon emissions from refrigeration used in the UK food industry (International Journal of Refrigeration, 2023), which puts the entire UK food cold chain at 29.1 TWh of electricity on 2019 data. 10.75 TWh of that is domestic fridge-freezers and belongs in the household tier, not here. Stripping it out leaves 15–18 TWh: retail refrigeration 6.91, food service 4.16–9.28 (the widest single uncertainty), food and drink manufacturing 3.08, cold stores 1.32, agriculture 0.40.

Data centres are much smaller than their billing suggests. DESNZ’s June 2026 Energy Trends feature puts GB data centres at 4.5 TWh of grid electricity in 2024, 2% of the 249.2 TWh consumed. Cooling is only the overhead above the IT load: at a UK power usage effectiveness of 1.5–1.8 that is roughly 1–1.5 TWh. Trade press reporting 5.8% of national electricity is a capacity-based figure on a different scope, not the metered number†. Adding non-food industrial refrigeration — chemicals and pharmaceuticals at 0.88 TWh, from the DESNZ-funded TICR project — gives 17–22 TWh of electricity, or 25–45 TWh of delivered cooling once each component is converted at its own efficiency: supermarket display about 2 to 2.5, cold stores and manufacturing 2.5 to 3, data centre cooling service roughly the IT load it removes.

So 26 TWh is defensible as a service figure and wrong as an electricity one — and the site had never said which it was. It sits at the low end of the service range. Two cautions carried forward: the cold-chain data is 2019 and the retail figures trace to studies from 2007–2016, making this the stalest evidence on the page; and roughly 1 to 1.5 TWh of data centre cooling may already sit inside the ECUK services cooling line above, since data centres are services-sector buildings — a fifth of the anchor, potentially counted twice†.

The scale is the point. At 17–22 TWh of electricity, process cooling is three and a half to four and a half times the entire comfort-cooling anchor of 4.8 TWh. Britain spends far more electricity keeping product cold than keeping people cool, and none of it is in ECUK’s services cooling line.

The diurnal separation. Until 17 August 2026 this fold opened with a chart separating comfort from process cooling by hour: an overnight floor of 0.261 GW/°C rising to 0.676 at noon, offered as evidence that a continuously running refrigeration load is separable from daytime comfort cooling. It was an artefact. The demand series already contains embedded solar and the analysis added it a second time, which manufactures a correlation with temperature in exactly the daylight hours where a cooling signal would appear. Corrected, the profile is flat — 0.26 at 04:00 against 0.24 at noon, a ratio of 0.93 rather than 2.6. There is no daytime separation in this data. Tier 0 above is now supported by the bottom-up count alone, not by any grid measurement. Found in peer review.

Slope to level. Refrigeration power is proportional to the lift, so a measured slope of demand against ambient should integrate to a level without needing the intercept. It does not work: below about 15 °C the night response goes flat, so the line that would be extrapolated back does not exist. The method was withdrawn rather than published.

An equipped-but-saturated tier. Latent demand minus delivered, where latent extrapolated the low-CDD slope linearly. Its premise was that the fleet flattens at high load. The measured response steepens — about +3 GWh per CDD across the cool bins against +17 across the warm ones — so the tier was structurally zero. Saturation would show as a concave curve; this one is convex. Removed 11 Aug 2026, with the computation kept as a diagnostic: a non-zero value would mean the curve had turned concave, which is the evidence that would bring it back.

Humidity, on a test that could have failed. Latent load rises with absolute moisture, so moist-air enthalpy days were a plausible explanation for the convexity. On the full 800-day record with embedded solar carried as a control, dewpoint adds 0.0006 to R² and enthalpy days add 0.0000 — against 0.031 for bank holidays — and both return the same balance point as the base, because there is nothing left for them to explain. The site’s stated test is whether adding enthalpy absorbs part of the quadratic term: it moves +3%. So the bend is the extensive margin — each building a step, the fleet the integral of those steps — not latent load. That test was capable of falsifying the reading: two earlier mis-specified runs moved the quadratic by 25 and 27 points, in opposite directions. It took three attempts to obtain a null worth having, after a double-counted solar term, an omitted solar control that left moisture proxying for it, and a workflow that fetched half the record†.

And three routes tried offline, not published here. Recovering the level that within-class centring removes was attempted five ways in all. Two more failed: fitting a balance point as the vertex of a V, which turns out to place its kink wherever the middle of the data is — the fitted vertex tracks each hour’s own mean temperature with R² 0.97 — and borrowing a level from the gas side, which establishes that balance points are spread across the stock but yields a shape without a scale. One remains open: comparing weekday with weekend demand at the same temperature, which gives 10 to 39% of the anchor and passes a night placebo. None is used in any published figure. All are documented in the methodology†.