The 32 bars that differ by a tenth and by 3.75 times

Dark navy benchmarking slide, masthead “Grid & Cooling Desk · hyperscale campus benchmarking” on the left and “Operator’s published PUE · trailing twelve months · Q2 2026” on the right, headlined “Thirty-two campuses, and a tenth of a point between them.” Directly beneath the headline sits a bordered chip reading “Demonstration · drawn by us from published figures”. A standfirst says every figure is the operator’s own published trailing-twelve-month power usage effectiveness — one metric, one methodology, one publisher, one unit; that the axis is linear, evenly ticked, zero-based and unbroken; that nothing is smoothed, indexed or rescaled and no campus has been dropped except the three the source does not give a twelve-month figure for, having fewer than twelve months of data; that the operator publishes these figures as a table and the bars are ours; and that the whole fleet lands between 1.04 and 1.15, a spread of 0.11, around a fleet-wide figure of 1.09. The main panel, headed “Campus power usage effectiveness, trailing twelve months” and subtitled “32 campuses with twelve months of data · sorted best to worst · zero-based linear axis, evenly ticked, unbroken · PUE = total facility energy ÷ IT equipment energy”, carries 32 horizontal gold bars, one per campus, on an axis running from 0 to 1.2 in even steps of 0.2. The bars form a near-solid block: Central Ohio (Lancaster), Ohio at the top at 1.04, then Omaha, Nebraska at 1.05, three campuses at 1.06, two at 1.07, seven at 1.08, six at 1.09, four at 1.10, four at 1.11, Singapore at 1.12, Changua County, Taiwan at 1.13, and Singapore (2nd facility) and Storey County, Nevada tied at the foot at 1.15, every value printed in monospace at the end of its bar. The difference between the shortest and the longest is a sliver at the right-hand end of a very long block, and no bar is highlighted. A caption reads “PUE as published, to two decimals · every campus with a TTM figure · nothing dropped, nothing rescaled · one scale throughout”. Eight cards run down the right. “The spread, in one figure” prints 0.11 of a point in large type and notes that it separates Central Ohio (Lancaster) at 1.04 from the two at the foot of the table, Storey County, Nevada and Singapore’s second facility, both at 1.15, and that the shortest bar is 90.4% of the longest. “Read-out for the desk” says the fleet is uniform to within a tenth of a point, with 23 of the 32 campuses within 0.02 of the fleet figure and 27 within 0.03, that on the evidence of this chart the site-efficiency programme has converged with no campus visibly adrift, and recommends closing out the per-site cooling retrofit programme and redirecting the remaining efficiency capex to capacity. “The league table” tabulates the top four and the bottom four with 24 campuses elided between them. “How PUE is defined” explains that it is total facility energy divided by IT equipment energy, that a PUE of 2.0 means another watt is consumed to cool and distribute power for every watt of IT, and that a PUE closer to 1.0 means nearly all the energy is going to computing. “The fleet against the industry” gives the fleet-wide figure as 1.09 against the Uptime Institute’s 2025 global average of 1.54, so the fleet reads about 29% below the industry average on this axis. “What moves a campus” names free-cooling hours, water availability, plant age and load factor, and notes that the tropical and desert campuses sit at the foot of the table and the Ohio and Nebraska ones near the top — a pattern this chart shows as almost nothing. “Where the figure gets used” says PUE is written into colocation leases and service-level agreements and reported to regulators and energy programmes. “Basis of measurement” credits Google Data Centers, “Power usage effectiveness”, campus table, trailing-twelve-month figures for Q2 2026 — the most recent quarter published — comprehensive across all large-scale data centres once at stable operations, in all seasons, including all sources of overhead, retrieved 9 September 2026. A method line runs along the foot: published figures as reported; one metric, one publisher, one unit; zero-based linear axis, evenly ticked, unbroken; every campus with twelve months of data; nothing smoothed, dropped or rescaled. A footnote credits the source, says all 32 values are as published, and states that Google publishes them as a table while this slide is a demonstration drawn by us in the manner of an internal benchmarking report — the desk, the read-out and the recommendation are invented, and every figure behind them is real. — Showing the misleading chart

The same 32 campuses on a cream background, kicked “The same 32 campuses · the same published figures · one subtraction” and “Overhead energy per unit of IT power · PUE − 1 · Q2 2026”, headlined “A tenth of a point apart, and 3.75 times apart.” A standfirst says nothing on the benchmarking slide was truncated, smoothed or rescaled, its axis really does start at zero and all 32 figures are correct; that what sets the picture is that a PUE cannot go below 1.00, being total facility energy divided by IT energy, so the 1 is the computing load itself and only what sits above it is overhead; that the 1 is inside every bar, identical in all of them, taking up between 87% and 96% of the ink on a scale the data can never touch; that subtracting it turns the same 32 numbers into watts of cooling and power distribution per watt delivered to computing, running from 0.04 to 0.15; and that the operator’s own page does exactly this and says so, a numbered footnote printing the whole subtraction the moment it wants to state a comparison in words. The main panel, headed “The same 32 campuses, drawn as overhead” and subtitled “Overhead energy per unit of IT power (PUE − 1) · same campuses, same order, same published figures, with each campus’s PUE kept as a label · zero-based linear axis, evenly ticked, unbroken”, uses the same bar track as the benchmarking slide but an axis running from 0 to 0.16 in even steps of 0.04. The block of near-identical bars has become a fan: Central Ohio (Lancaster), Ohio at the top is a short blue bar labelled “0.04 PUE 1.04”, Omaha is 0.05, the middle of the fleet sits at 0.08 to 0.11, and Singapore (2nd facility) and Storey County, Nevada are orange-red bars at 0.15, nearly four times the length of the blue one; every bar carries its overhead figure and its PUE reading side by side. A caption reads “Same numbers, same sort, one subtraction · the axis now starts where the quantity starts”. Seven cards run down the right. “What the subtraction does” tabulates three pairs as the bars showed them and as the data centres are doing them: Storey County against Central Ohio (Lancaster), +10.6% and 3.75×; Singapore against Omaha, +6.7% and 2.40×; Changua County against Fredericia, +6.6% and 2.17×. “The publisher’s own footnote” quotes it verbatim — “(1 − (Google’s overhead energy use [0.09] divided by the industry average overhead energy use [0.54])) × 100 = 83%” — notes that a fleet-wide 1.09 against an industry average of 1.54 is 29.2% below it on the axis the chart uses, that nobody quotes that because the sentence people want is “83% less overhead” and it needs the 1 taken off both numbers first, and adds that this is the practice the exhibit is recommending, printed by the operator in full with the arithmetic shown — the chart being the thing that lost it. “Two decimals is not many” says subtracting a constant leaves the difference untouched and wrecks the precision of the ratio: a PUE published as 1.04 is anything from 1.035 to 1.045, so an overhead of 0.04 carries one significant figure and the 3.75× could honestly be anything from 3.2× to 4.4×, with the same bound applying to every ratio on the page — the 2.40× spanning 2.09–2.78 and the 2.17× spanning 1.92–2.45 — which is a reason to publish three decimals rather than to go back to the first chart, where the same uncertainty is present, unchanged, and simply cannot be seen. “Other units with an offset origin” tabulates power usage effectiveness against a real zero of 1.00, degrees Celsius against −273.15 °C, degrees Fahrenheit against −459.67 °F, an index at 2017 = 100 against 100, IQ and standard scores against no zero at all, elevation against its datum and a calendar year against an epoch, with what a bar on each should show. “One bar, taken apart” draws a single campus at PUE 1.09 from zero as a composite bar: a long pale segment keyed “1.00 — the IT load. Identical in every bar on the chart, and no data centre can move it” and a short gold segment keyed “0.09 — the overhead. The only part of the bar that is data”, over the line “The first chart drew both. The second draws the right-hand piece.” “What to draw instead” says to plot overhead per watt of IT on a zero-based axis with the familiar PUE printed beside each bar as a label, to keep PUE where people act on it since a colocation lease is written against a PUE, and never to send the reading through a ratio or a percentage change, because differences survive an offset origin and nothing else does. A “Basis of measurement” card credits Google Data Centers, “Power usage effectiveness”, campus table, TTM figures for Q2 2026, and the same page’s fleet figure and footnote citing the Uptime Institute 2025 Global Data Center Survey global average of 1.54, retrieved 9 September 2026, with the subtraction, the ratios and the rounding bounds being ours. Five notes run along the foot: the same 32 published figures; one subtraction, named; zero-based on the quantity’s own zero; the reading kept as a label; the rounding stated. A closing footnote credits the source, notes that the 83% comparison uses the 2025 annual fleet figure which is also 1.09, and states that Google publishes these figures as a table, that both drawings are ours, that the desk, the read-out and the recommendation on the first are invented, and that every figure behind them is real. — Showing the misleading chart

Be the first to star this exhibit

A benchmarking slide we drew ranks 32 data centre campuses on their operator’s own published power usage effectiveness — one metric, one methodology, one publisher, one unit, on a linear axis that is evenly ticked, unbroken and genuinely zero-based. The fleet lands between 1.04 and 1.15, the shortest bar is 90.4% of the longest, and the read-out concludes that the efficiency programme has converged. Every figure on it is right. What the picture cannot show is that a PUE has a floor of 1.00, because it is total facility energy divided by IT energy and the 1 is the computing load itself: that constant is inside every bar, identical in all of them, taking up between 87% and 96% of the ink on a stretch of scale the data can never reach. Subtract it and the same 32 numbers become overhead energy per unit of IT power, running from 0.04 to 0.15 — a spread of 3.75 times rather than a tenth of a point. Google publishes these figures as a table rather than as this chart, and its own page performs exactly this subtraction, in a numbered footnote, the moment it wants to state a comparison in words.

Google Data Centers, “Power usage effectiveness”: campus PUE table, trailing-twelve-month figures for Q2 2026 — 32 campuses with twelve months of data. Fleet-wide TTM PUE 1.09; the same page’s footnote cites Uptime Institute, 2025 Global Data Center Survey, global average 1.54. Subtraction, ratios and rounding bounds ours. Retrieved 9 September 2026, 2026Added September 9, 2026

01The claim

Thirty-two data centre campuses, one metric, one methodology, one publisher, one unit. The figures are the operator’s own published trailing-twelve-month power usage effectiveness — total facility energy divided by IT equipment energy, measured across all large-scale data centres once they reach stable operations, in all seasons, including all sources of overhead. The axis is linear, evenly ticked, unbroken and starts at zero. Nothing is smoothed, indexed, rebased or rescaled, and no campus has been dropped except the three the source does not give a twelve-month figure for, having fewer than twelve months of data. The whole fleet lands between 1.04 at Central Ohio (Lancaster) and 1.15 at Storey County, Nevada and Singapore’s second facility — a spread of 0.11 — with 23 of the 32 campuses inside 0.02 of the fleet-wide 1.09 and 27 inside 0.03. The shortest bar on the chart is 90.4% of the longest. Read-out for the desk: the site-efficiency programme has converged, there is no campus visibly adrift of the rest, and the remaining per-site cooling retrofit budget should be closed out and redirected to capacity, since no campus is far enough behind the fleet to repay a retrofit.

02The trick

Every figure on that slide is right, the axis really does start at zero, and the picture is still not the data — because the number being drawn has a floor. PUE is a ratio of total facility energy to IT energy, so it cannot go below 1.00, and the 1 is not a small formality at the bottom of the range: it is the computing load itself, the thing the building exists to power. Only what sits above the 1 is overhead — the cooling, the fans, the losses in power distribution — and that is the entire quantity the metric was invented to measure. Draw PUE from zero and the 1 is added to every bar, identically, in every row, taking up between 87% and 96% of each bar’s length on a stretch of scale that no data centre anywhere can ever move. What remains for the data to work with is the last tenth of the block, which is why 32 genuinely different buildings arrive looking like one bar drawn 32 times. Subtract the floor and the same 32 published numbers, in the same order, become overhead energy per unit of IT power: 0.04 at Central Ohio (Lancaster) and 0.15 at Storey County, Nevada. That is 3.75 times as much energy spent on cooling and power distribution to run the same watt of servers — the campus at the foot of the table burns about 10.6% more energy in total, which is the reading the bars show, and nearly four times as much of the part anybody can do anything about. The other pairs move the same way: Singapore reads 6.7% above Omaha and carries 2.40 times the overhead, and Changua County reads 6.6% above Fredericia and carries 2.17 times. None of this is a truncation, and widening or narrowing the axis cannot reach it, because the constant is inside the unit rather than in the frame. The most useful evidence that the subtraction is the right one is that the publisher already performs it, in public, with the arithmetic shown. Google’s own page quotes a fleet-wide 1.09 against the Uptime Institute’s 2025 global average of 1.54 — 29.2% below it on the axis these bars use — and then says, in words, “83% less overhead energy than the industry average”, with a numbered footnote spelling out the sum: one minus 0.09 divided by 0.54, times 100. Nobody quotes the 29.2%, because the sentence anyone actually wants is about overhead, and that sentence needs the 1 taken off both numbers first. That footnote is the practice this exhibit is recommending; what it shows is how easily a chart loses what a caption keeps. One honest caveat travels with the redraw, and it belongs on the chart rather than in a defence of the first one: subtracting a constant leaves a difference untouched and demolishes the precision of a ratio. A PUE published to two decimals as 1.04 is anything from 1.035 to 1.045, so an overhead of 0.04 carries one significant figure, and the 3.75× the published figures give could honestly be anywhere from about 3.2× to 4.4× — with the same rounding covering every ratio here, the 2.40× spanning 2.09–2.78 and the 2.17× spanning 1.92–2.45. That is an argument for publishing three decimals. It is not an argument for the first chart, where the same uncertainty is present, unchanged, and simply cannot be seen. (Google publishes these figures as a table, not as this chart; both drawings are ours, the desk, the read-out and the recommendation on the first are invented, and every figure behind them is real.)

03The fix

Plot the quantity, not the reading. On the redraw the axis runs 0 to 0.16 instead of 0 to 1.2, the bars carry overhead energy per unit of IT power, and each one keeps its familiar PUE beside it as a label — the same 32 numbers, the same order, one subtraction. The wall of identical bars becomes a fan: Central Ohio (Lancaster) at 0.04, the bulk of the fleet between 0.08 and 0.11, and Storey County and Singapore’s second facility at 0.15, nearly four times the leader. That fan is the ranking a retrofit decision actually turns on, and the redraw is not introducing drama — it is removing a constant that was standing in front of the data. Keep the reading, because people genuinely act on it: a colocation lease is written against a PUE, an efficiency target is set in PUE, a regulator asks for PUE. So print it as a label rather than as the bar, and say in the caption what the unit’s zero means. The redraw also states its own rounding, which the first chart could not: at two published decimals the 3.75× is really somewhere between 3.2× and 4.4×, and saying so is the difference between a chart that can be checked and one that merely looks precise.