Energy and Water: What Is Measured and What Is Guessed

AI energy numbers come in three grades, and most news coverage and public debate mix them without saying so. Some are measured from meters and utility bills. Some are reported by companies about their own systems, with methods they chose. Many more are modeled by outside researchers who had to guess at things the companies keep private. All three can be useful. Only one of them is a measurement.

The big totals

National and global figures mostly trace back to LBNL and the IEA.

Lawrence Berkeley National Laboratory (LBNL) wrote the December 2024 United States Data Center Energy Usage Report at the request of Congress. It estimates US data centers used about 76 TWh in 2018 (1.9 percent of US electricity) and 176 TWh in 2023 (4.4 percent). For 2028 it gives a range: 325 to 580 TWh, or 6.7 to 12 percent of projected US use. The report says growth in GPU-accelerated servers for AI drove the recent rise. These are modeled from server shipment data, equipment power ratings and assumptions about use and cooling, cross-checked against what is known. They are the best public US estimate. They are still estimates.

The International Energy Agency's 2025 Energy and AI report puts global data center use at about 415 TWh in 2024, roughly 1.5 percent of world electricity, with the US at 45 percent of that, China 25 percent and Europe 15 percent. Its base case reaches about 945 TWh by 2030, with a 2035 range of 700 to 1,700 TWh depending on the scenario. IEA says data centers account for nearly half of US electricity demand growth through 2030, and that accelerated servers, mainly for AI, grow about 30 percent a year in its base case.

Source	What it covers	Key figure	Grade
LBNL 2024	All US data centers	176 TWh in 2023; 325 to 580 TWh in 2028	Modeled from shipments and equipment data
IEA 2025	All data centers worldwide	415 TWh in 2024; 945 TWh base case 2030	Modeled, scenario based
Google, Aug 2025	Median Gemini Apps text prompt	0.24 Wh, 0.26 mL water	Company reported, not independently verified
Sam Altman, 2025	Average ChatGPT query	0.34 Wh, 0.000085 gallons water	Company statement, no method published
Epoch AI, Feb 2025	Typical GPT-4o query	About 0.3 Wh, plausible range 0.1 to 4 Wh	Outside model with stated assumptions
Mistral, Jul 2025	400-token Le Chat response	1.14 g CO2e, 45 mL water	Company lifecycle study, peer reviewed by consultants

Notice that the data center totals are about all computing. AI is the fastest growing slice, but neither LBNL nor IEA claims all 176 or 415 TWh is AI. Anyone who says "AI uses 4.4 percent of US electricity" has made a mistake.

Why per-query numbers disagree

For years the common figure was about 3 Wh per ChatGPT query, roughly ten times a Google search. Then in February 2025 Epoch AI worked it out from scratch for GPT-4o and got about 0.3 Wh. Its assumptions were stated: about 100 billion active parameters, 500 output tokens, H100 chips at about 1,500 watts each all in, running at 10 percent of peak. Change those and the answer moves anywhere from about 0.1 to 4 Wh. Feed the model a 10,000-token document and Epoch's estimate rises to about 2.5 Wh. A 100,000-token input, around 40 Wh.

Google published the first detailed company figure in August 2025: 0.24 Wh for its median Gemini text prompt, measured in May 2025. It also gave a narrower number, 0.10 Wh, counting only the active AI chips. The larger figure adds idle machines held ready for traffic spikes, host CPUs and memory, and building overhead at Google's fleet PUE of 1.09. Google says the median prompt's energy dropped 33-fold in a year. Google also notes the figures were not checked by an independent third party.

What moves the number:

• Model size. Bigger models do more math per word.
• Answer length. Energy scales roughly with tokens produced. Reasoning models that "think" in long hidden chains produce many more.
• Input length. Pasting in a long report costs real energy before a word comes back.
• Utilization. A chip that sits half idle still draws power. How busy the fleet is changes the per-query share.
• Boundary. Chips only, whole building, or full lifecycle including chip manufacturing.
• Median versus average. A few huge requests pull an average up; a median ignores them.
• Images and video. Generating a video clip is a different order of work from a text reply.

Water: two very different numbers

Water is where the confusion gets worst, because two separate quantities share one word.

Direct water is what the data center withdraws and evaporates, mostly in cooling towers or evaporative coolers. Some designs use almost none and run closed loops or air cooling, at the cost of more electricity. Indirect water is what power plants use making the facility's electricity. Thermal plants (coal, gas, nuclear) evaporate a lot; wind and solar almost none.

LBNL's figures, as summarized by the Global Investigative Journalism Network, put 2023 US data center direct water at about 17 billion gallons and indirect water at about 211 billion gallons, roughly 12 times more. Direct use could double or quadruple by 2028.

The per-query water figures diverge even more than energy. Google: 0.26 mL per median prompt. Altman's 0.000085 gallons works out to about 0.32 mL (0.000085 times 3,785 mL per gallon). Mistral, using a lifecycle method that counts much more of the supply chain, reported 45 mL per 400-token response. A frequently repeated figure of a 500 mL bottle per handful of responses is widely attributed to early versions of the UC Riverside paper "Making AI Less Thirsty" (the paper has been revised several times, so check which version says what before you repeat it); the current version's estimates include about 700,000 liters evaporated to train GPT-3 and a 2027 global AI water withdrawal of 4.2 to 6.6 billion cubic meters. Those numbers include indirect water and assumptions about where and when the work ran. Repeat any of them with the boundary attached.