The cooling band that shrank while the cooling grew

Dark navy briefing slide, masthead “Efficiency Desk · U.S. data centers · Where the electricity goes”, headlined “Cooling overhead is being engineered out”. A subhead reads: Berkeley Lab’s national model of the entire U.S. data center stock, at the three years the report publishes; every share below is the figure in the report, on an axis running the full 0% to 100%. Under the label “Share of total U.S. data center electricity · by equipment category”, a two-band stacked area runs across a zero-to-one-hundred-percent axis marked 0%, 25%, 50%, 75% and 100%, with three anchor points at 2018, 2024 and 2030. The lower blue band is labelled “IT equipment — servers, storage, network”; the gold band above it is labelled “Facility infrastructure — cooling & power distribution”, and carries the figures 36% at 2018, 31% at 2024 and 26% at 2030, a white boundary line rising gently from left to right as the gold band thins. The 2018 and 2024 columns are marked historical and the 2030 column Reference Case. A strip beneath the axis reads “Average PUE, all U.S. data centers, 1.55 → 1.45 → 1.36 — overhead per unit of IT load, falling every year”. Down the right, a card carries 26% in large white type over the words “of U.S. data center electricity goes to cooling and power distribution in 2030 — against 36% in 2018”. A second card, “The overhead band, year by year”, lists 2018 historical at 36%, 2024 historical at 31% and 2030 Reference Case at 26%. A gold-bordered note reads: “Ten points of overhead gone in twelve years. The cooling plant is being designed out of the data center,” with a small grey disclosure beneath it, “Shares of total data center electricity. Totals are published separately.” Three ticked boxes run along the bottom: “Every share published — Berkeley Lab’s 2025 Update, Section 3.6, nothing interpolated between the years shown”, “One definition throughout — infrastructure is cooling plus power-distribution losses, the same quantity PUE measures”, and “Full axis — 0% to 100%, no truncation, no log scale, no second axis”. A footnote credits the Berkeley Lab report LBNL-2001758 of June 2026 for the shares and the PUE figures, notes that the 2030 share is PUE minus one over PUE at the Reference Case PUE of 1.36, and calls the slide an Efficiency Desk demonstration drawn by us. — Showing the misleading chart

The same three years on a cream background, kicked “The same three years · with the total put back” and headlined “The band lost a quarter of its width. The cooling grew 6.4×.” A subhead explains that both statements come from the same model and neither is in dispute: the first chart’s axis was share of total, where a band can only narrow if the band beneath it grows faster, and the band beneath it is mostly AI — put the axis back into terawatt-hours and every series is an amount again. The main panel, headed “Unstacked · U.S. data center electricity, terawatt-hours a year · one zero-based axis, both series on it”, shows six columns in three pairs on an axis running 0 to 500 TWh: for 2018 a blue IT-equipment column at 49 beside a gold facility-infrastructure column at 27; for 2024, 132 beside 60; for 2030 Reference Case, 477 beside 172. A note in the legend reads “×9.7 and ×6.4 respectively, over the same twelve years.” Below, a row headed “One series, three drawings · the same three infrastructure numbers in all of them” shows the figures 27, 60 and 172 drawn three ways. First, on their own zero axis: three gold bars climbing steeply, captioned “An amount. Zero at the bottom, terawatt-hours up the side. The series is drawn 6.4× bigger because it is 6.4× bigger.” Second, stacked on the IT band with the axis stretched to a 649 total: the same figures reduced to thin gold caps on top of tall blue columns, captioned “Still an amount — but now a gap between two moving edges, on an axis stretched to 649. Nobody reads a sliver at the top of a column.” Third, normalised to 100%: a gold ribbon narrowing from 36% to 26%, captioned “Not an amount at all. The total has been divided out. The band narrows because AI grew faster, not because the cooling shrank.” A gold-ruled strip beneath reads “9.7× the IT load × 0.65 the overhead per unit of IT = 6.4× the overhead,” with the line “The efficiency gain is real. It is also the smaller of the two factors, and the chart drew only that one.” Five cards run down the right. “The efficiency win, stated as a level” notes that PUE fell from 1.55 to 1.36, so every kilowatt-hour of IT carried 0.55 kWh of cooling and distribution loss in 2018 and carries 0.36 kWh in 2030, a 35% cut, while the IT load it serves rises 9.7 times. “The total the stack divided out” lists 2018 at about 76 TWh and 1.9% of U.S. electricity, 2024 at 192 TWh and 4.7%, and 2030 at 649 TWh and 11.8%. “What PUE can and cannot say” explains that PUE is a ratio, overhead per unit of IT, which can improve every year while the building draws more power every year — a fine efficiency metric and a useless consumption one, as is any band on a share axis. “Where the bottom slot goes” notes that one series is drawn from zero while the rest start wherever the series below them ended, and asks what the same chart would look like with the order reversed. “The check that catches it” says: count the baselines, there is one, then ask what the axis cannot answer — did the level move, or only the share? A footnote credits Berkeley Lab’s 2025 Update, LBNL-2001758, for the 2024 and 2030 totals and shares and the PUE series, the same team’s 2024 report, LBNL-2001637, for the 2018 total, explains that the terawatt-hour splits are the published total divided by the published PUE, and says both drawings are our own. — Showing the misleading chart

Be the first to star this exhibit

Cooling and power distribution took 36% of U.S. data center electricity in 2018, 31% in 2024 and about 26% in Berkeley Lab’s 2030 Reference Case. Drawn as a stacked share chart, that band narrows by a quarter and the efficiency story writes itself. The same model puts the electricity behind the band at roughly 27, 60 and 172 terawatt-hours — a 6.4-fold rise. The band narrowed because the AI load underneath it grew faster.

Sarah J. Smith, Alex Hubbard, Alex Newkirk, Mohan Ganeshalingam, Billie Holecek, Dale Sartor, Michael Mills and Arman Shehabi, “United States Data Center Energy Usage Report: 2025 Update”, Lawrence Berkeley National Laboratory, LBNL-2001758, June 2026; 2018 total from Shehabi et al., “2024 United States Data Center Energy Usage Report”, LBNL-2001637, December 2024, 2026Added August 18, 2026

01The claim

Cooling overhead is being engineered out of the American data center. Berkeley Lab models the entire national stock, and the facility-infrastructure band — the chillers, the fans, the power-distribution losses, everything that is not the computing itself — takes 36% of total data center electricity in 2018, 31% in 2024 and about 26% in the 2030 Reference Case. The average PUE behind those shares falls from 1.55 to 1.45 to 1.36. Ten points of overhead gone in twelve years, on an axis running the full 0% to 100%, with no truncation, no log scale and no second axis. Every figure is the published one.

02The trick

Nothing on that chart is wrong, and the efficiency gain it reports is real. The problem is what a stacked share chart is able to say. Stacking gives one series a baseline and none of the others: the blue band starts at zero, and the gold band starts wherever blue happened to end, so it is not a length from anything — it is the gap between two moving edges. Then the axis is run to 100%, which divides the total out entirely. On that axis nothing is a quantity. A band can only narrow if the band beneath it grows faster, so “the overhead band is shrinking” and “cooling is using less electricity” are different claims, and only the first one is drawn. Here the second is false. Put the published totals back and the same model says U.S. data centers used about 76 TWh in 2018, 192 TWh in 2024 and 649 TWh in the 2030 Reference Case — 1.9%, 4.7% and 11.8% of all U.S. electricity. Divide each total by the published PUE and the infrastructure behind that thinning band is roughly 27, 60 and 172 TWh: a 6.4-fold increase, drawn as a ribbon losing a quarter of its width. The arithmetic is worth keeping, because it is the whole exhibit in one line. Overhead growth is IT growth multiplied by the change in overhead per unit of IT: the IT load rises 9.7×, each unit of it carries 0.65 as much overhead as before, and 9.7 × 0.65 = 6.4. The efficiency win is genuine and it is the smaller of the two factors — which is exactly the fact a share axis cannot express, because a share is a ratio and the ratio moved the way the smaller factor pushed it. The same trap is waiting in any stacked area chart, and it gets worse the steeper the stack climbs: the eye measures the shortest way across a band rather than the vertical way, so a band riding the top of a steep ramp reads thinner than it is. (This exhibit is our own demonstration, drawn in the style of an efficiency briefing. Berkeley Lab publishes the totals, the shares and the PUE series together — which is precisely why the trick needs a chart of its own to pull off.)

03The fix

Take the stack apart. One zero-based axis in terawatt-hours, both series on it, and the picture reverses: two columns per year, both climbing hard, the gold one from 27 to 60 to 172. That drawing answers the question the reader actually had, which was never “what fraction of the bill is cooling” but “is the cooling getting bigger”. Then show the same three numbers three ways side by side — on their own zero axis, stacked under the total, and normalised to 100% — because seeing one series shrink to a sliver and then to a narrowing ribbon without a digit changing is the fastest way to learn what stacking costs. Keep the ratio too, on its own axis where it belongs: PUE from 1.55 to 1.36 means 0.55 kWh of overhead per kWh of IT falling to 0.36, a 35% cut, and that is a real engineering achievement that deserves its own line rather than a share of somebody else’s. Say both numbers in the same sentence and neither one lies. The habit generalises to every composition chart you meet, from cloud spend by service to emissions by sector to market share by vendor: count the baselines, and if there is only one, ask whether the band you are reading fell, or was merely overtaken.