Summary

  • The 74–132 GW band in the American data-centre argument is the printed 325–580 TWh envelope divided by 4,380 hours, which is 8,760 hours at the report's assumed 50% average capacity utilisation. That division is mine; it lands on the published pair to the decimal.
  • The same December 2024 edition quotes its 2016 predecessor at about 70 billion kWh for 2014 and states its own 2014–2016 level as about 60 TWh, with no bridging sentence in the passages I read.

A congressional request produced it, a national laboratory wrote it, and the Regents of the University of California are the manager and operator of the laboratory whose staff signed it. The document is the 2024 United States Data Center Energy Usage Report, numbered LBNL-2001637, DOI 10.71468/P1WC7Q, on the laboratory's publication page dated 19 December 2024 publication page.

What the document says it is

The sentence that fixes the population sits in the lead-in to the executive summary's first figure: the report provides "an estimate of total U.S. data center electricity use including servers, storage, network equipment, and infrastructure from 2014 through 2028" the December 2024 report. Read the noun. It is an estimate, and the list of what sits inside it — servers, storage, network gear, infrastructure — is the entire perimeter. Nothing in that sentence claims a meter.

The model is stated as plainly: it "characterizes IT equipment as either servers, storage, or networking equipment" and uses shipment data to build an installed base for each year the December 2024 report. That is a defensible method and it is still a chain. It begins in a reported quantity, shipments, and ends in an authored one, annual electricity, with per-unit power, utilisation and cooling standing between them as parameters.

The envelope, and the divisor that turns it into gigawatts

For a year that has not happened the report declines to print a single figure. It prints an envelope: roughly 325 to 580 TWh in 2028, equal to 6.7% to 12.0% of forecast national consumption the December 2024 report. The spread is not noise around a true value. The report attributes it to three things: how many GPUs ship each year and whether manufacturers can meet that demand; how much computational power GPU-accelerated servers deliver and how often they are used; and which cooling systems get built, liquid or otherwise the December 2024 report. The high end is about 1.78 times the low end — my division. Two modelled futures printed side by side, neither labelled best case, neither labelled worst.

Then comes the sentence the power argument actually quotes: at an assumed average capacity utilisation of 50%, that range "would translate to a total power demand for data centers between 74 and 132 GW" the December 2024 report. Divide it out and the assumption stands in the open. A year is 8,760 hours; half of that is 4,380. 325 ÷ 4,380 gives 74.2 GW, 580 ÷ 4,380 gives 132.4 GW — my arithmetic, and it reproduces the printed band. The gigawatt figure therefore inherits every assumption in the terawatt figure and adds one more of its own. Raise the assumed utilisation and the same energy becomes a smaller power number; nothing measured has moved.

One edition, two baselines for the same year

At the other end of the series the document shows its seams. Its Introduction quotes the laboratory's own 2016 study: the sector "was estimated to have consumed about 70 billion kWh in 2014, equivalent to 1.8% of total U.S. electricity consumption" the December 2024 report. In the same edition's historical narrative, annual use "remained stable between 2014-2016 at about 60 TWh" the December 2024 report. Seventy against sixty is 10 billion kWh, about 14.3% — my subtraction. In the passages I read, no sentence reconciles them. I read the executive summary, the introduction and the methodology-update list; the pages between them were not read. That is a limit of my reading, not a finding that the report never reconciles its baselines.

What can be said without any such dressing: a year that a 2016 study priced at 70 billion kWh appears in a 2024 series at about 60. Two authored estimates sit inside one document, and the older figure arrives as a quotation of the older work rather than as a reconciled series.

The same quantity, printed again eighteen months later

In June 2026 the same group printed the quantity a second time. The 2025 Update puts 2024 at 192 TWh, 4.7% of total U.S. electricity consumption, and carries the same scope list the 2025 Update, the same text in the repository copy. Watch the years. 176 TWh was 2023; 192 TWh is 2024. They are consecutive years from two editions, not one year revised, and anyone who strings them into a growth rate is subtracting 2023 from 2024 and calling the result a revision.

What the Update says about revision is more interesting than its headline. 2024 "falls at the low end of the estimated range in the 2024 Report, with revised estimates of electricity use in prior years falling slightly below 2024 Report estimates" — because reported GPU shipments for 2023 and 2024 came in lower, and because the assumed power parameters for AI inference servers were revised the 2025 Update. The previous envelope held. The inputs beneath it moved.

And the 2028 projection? A Reference Case of 464 TWh that "falls in the middle of the range presented in the 2024 Report" the 2025 Update. The arithmetic midpoint of 325 and 580 is 452.5 TWh, so 464 sits about 11.5 TWh above it, roughly 2.5% — my arithmetic, against endpoints both prefaced "roughly". That is not a contradiction. It is a reminder that "middle" is a qualitative word doing quantitative work.

The ranged projection has moved to 2030: a Reference Case of 649 TWh, 11.8% of forecast 2030 U.S. electricity "per NERC's 2025 Long-Term Reliability Assessment", inside a compounded band of 521 to 843 TWh, 9.5% to 15.3% the 2025 Update. The laboratory's own abstract repeats the comparison — 11.8% by 2030 against 6.7% to 12.0% by 2028 the laboratory's abstract. Here is the small documentary asymmetry worth naming: the 2026 edition tells you whose forecast its denominator is. In the 2024 passages I read, 4.4% stands against "total U.S. electricity consumption" with no source named. I did not read the report's percentage methodology section, so that is an observation about those passages.

The exclusion that is real, and the edition that spells it out

One more seam, and it is the one most often quoted past. The Congressional Research Service states the 2023 figure as "approximately 176 terawatt-hours (TWh), approximately 4.4% of U.S. annual electricity consumption that year, not accounting for cryptocurrency", footnoting "Does not include energy use related to cryptocurrency" to page 6 of the laboratory's report CRS R48646. The crisp sentence does exist, and it is printed in the Update: "Consistent with the 2024 Report, the energy use of cryptocurrency mining is not included in the scope of this report" the 2025 Update, repository copy. In the 2024 passages I read, the nearest language is procedural — a note that Section 6 provides "a separate analysis for cryptocurrency" the December 2024 report. The exclusion is real. The sentence that states it in that form belongs to the 2026 edition and to a congressional footnote; whether page 6 of the 2024 report says it in the same words is not something my reading settled.

The same CRS report records a caveat from the laboratory itself: the calculations assumed data centres "would operate consistently with how they were commissioned and designed but that results may differ" CRS R48646. That one line is the honest description of every figure above it.

Where to check

None of this makes the estimate wrong. A model is allowed to be a model, and this one prints its envelope, names the three drivers that widen it, and discloses the utilisation assumption that converts energy into power. What the seams tell you is where the checkable point sits — not at the headline, but at the envelope, the divisor, and the successor's placement of new values inside the old range. A reader who quotes 132 GW as a measurement has dropped four things in one breath: the low end, the three drivers, the 50%, and the year.

A reader who quotes 325 to 580 TWh has a document that can be argued with, which is the most a forecast can honestly offer.