Can Environmentalists Like AI?

The electricity, water and computing behind each useful answer keep falling. The total keeps rising. Both are on this chart.

What These Curves Do and Don't Measure
  • Each line is electricity, water or computing per unit of AI work. Not industry totals.
  • Per unit, most published measures fall by half or more every 12 to 24 months, and one fell 97% in a year. The buildings' overhead and water fall more slowly.
  • In total, data centres use more power and more water every year, because the amount of work is growing faster than the savings (the Jevons paradox).
The question: will the efficiency curves overtake the growth in use?
What one unit of AI work costs in electricity, water and computing
Each panel is a separate measure, in the units its source published. Lower is better on every panel.

Latest update November 18, 2026

What is established

CaveatThe 33-times figure is Google's own, from a paper that has not been peer reviewed.

In August 2025 Google published the first measurement of a commercial chatbot from inside the building. A typical Gemini text prompt used 0.24 watt-hours in May 2025, about what a 10-watt LED bulb uses in a minute and a half.

Twelve months earlier the same prompt, measured the same way, had used 33 times more.

Outside estimates for 2025 land close to Google's number. Epoch AI worked out about 0.3 watt-hours for a typical ChatGPT query, a Microsoft research team modelled a median of 0.31, and a University of Rhode Island team got 0.42 for a short one.

CaveatThe famous 2023 figure of about 3 watt-hours per answer is on the chart as a disputed grey line. Microsoft's researchers say early estimates were 4 to 20 times too high.

The longer records come from researchers who don't sell AI. Epoch AI and MIT found that the computing needed to reach a fixed result with a language model halved about every eight months between 2012 and 2023.

Epoch's table of leading AI chips shows the work done per watt improving about 40% a year. On that table, a 2024 chip does the same computing on roughly a third of the electricity of a 2020 chip.

The buildings improved too. In 2007 the average data centre burned 1.5 units of power on cooling and overhead for every unit that reached the computers. In Uptime Institute's 2025 survey that overhead was 0.54.

The methods, the size of the model, the chips and the buildings each use less electricity for the same work, and the four gains multiply.

The newest gains are mostly software. Google credits its own 33-times fall to the model and to keeping machines busier, not to new buildings.

Why the curves fall

Smarter methods. OpenAI measured the computing needed to train an image model to one fixed standard. It fell 44 times between 2012 and 2019, faster than chips improved over the same years.

Smaller models that match old big ones. Stanford's AI Index tracked the smallest model that could pass one widely used test. In 2022 it had 540 billion parameters. By 2024 a model with 3.8 billion passed, 142 times smaller.

Two models that passed the same test, drawn to scale
Each circle's area is the model's size in parameters
PaLM, 2022 540 billion Phi-3-mini, 2024 3.8 billion: same test passed

Google says software accounts for most of Gemini's fall: 23 of the 33 times came from the model itself, and the rest from keeping machines busier.

Better chips. Vendors claim far bigger numbers than Epoch's 40% a year. Nvidia says one system delivers 25 times the performance of its predecessor at the same power. Those are sales figures, and they aren't on the chart.

Leaner buildings, and less water. Microsoft reports its data centres used 0.49 litres of water per kilowatt-hour in 2021 and 0.27 in 2025. It says every new design since August 2024 cools with no evaporated water at all (a fully self-contained water loop), with the first sites due in late 2027.

CaveatWater is the weakest line here. Google's water per kilowatt-hour was flat for 2023 and 2024, and Microsoft's zero-water buildings are a promise until they open.

Water per prompt is harder to pin down than energy. Google counts 0.26 millilitres, cooling water only. A 2023 university estimate for an older model came to 10 to 50 millilitres a response, because it also counted the water used to generate the electricity. The researcher behind that estimate says Google is "hiding the critical information."

Where the curves stop applying

Every falling line above describes a plain text answer or a fixed task. The newer ways of using AI cost far more per job.

Electricity for one job, by kind of job
Watt-hours. Each quarter of the bar's length is ten times more.
A typical text prompt
0.24 Wh Google, measured
A long "reasoning" answer
3.9 Wh Microsoft researchers, modelled
One instruction to a coding agent
about 150 Wh one user's logs, estimated
A 5-second video clip
about 944 Wh one open model, measured
CaveatEfficiency has a history of raising total use. When a thing gets cheaper to do, people do more of it, and they find new things to do with it.

Hugging Face tested open models on identical hardware. Models that "reason" before answering used 30 times more energy on average.

The same team compared 14 new models with similar-sized ones from nine months earlier. Eight used as much energy or more. Its authors wrote that this "runs counter to the prevailing narrative that AI is becoming more efficient."

Each job costs less while the total climbs: the world's data centres used 415 terawatt-hours in 2024 and 485 in 2025, and the International Energy Agency projects about 950 by 2030.

Those are the agency's figures as reported by the trade press. Google's own electricity use rose 37% in 2025 and its water use 34%.

CaveatSome of the new power is gas. A law centre says xAI ran at least 35 gas turbines without permits in Memphis, and Louisiana approved three gas plants for one Meta data centre.

Where the power comes from matters as much as how much. Carbon Brief, reporting the agency's data, puts fossil fuels at nearly 60% of data-centre electricity in 2024, with gas-fired supply for data centres expected to more than double by 2035.

Where the water comes from matters too. One analysis found more than two-thirds of data centres built or planned since 2022 are in places prone to drought.

What Does This All Buy For Humanity? (The Payoff)

The other half of the question is what all that computing produces. Each item below carries a tag for how solid it is.

Measured a real deployment with a reported result Piloted a trial or a lab result Forecast a projection or a plan

Medicine

MeasuredTime to a drug candidate. Insilico Medicine reports its 22 AI-led programs averaged about 13 months, against 2.5 to 4 years the usual way. Company figure.
Usual way
30 to 48 months
AI-led
about 13 months
MeasuredProtein shapes. AlphaFold predicted more than 200 million structures. More than 3 million researchers have used the database.
ForecastFewer animal tests. The US FDA plans to phase out animal testing requirements for some drugs, with computer models among the replacements.

Materials

PilotedNew crystals. DeepMind's GNoME predicted 2.2 million new crystal structures and rated 380,000 the most stable. Outside labs had made 736 of them.
Predicted
2.2 million
Made by outside labs
736
PilotedLower-carbon concrete. An AI-designed mix for a Meta data centre is estimated to cut the concrete's carbon footprint by 35%. Company estimate.
PilotedSolar cells. An AI-guided search through about a million candidate molecules raised a reference perovskite cell to 26.2% efficiency, in a peer-reviewed study.

Renewable energy and the grid

MeasuredWind power. Forecasting output 36 hours ahead raised the value of 700 megawatts of Google's wind energy by roughly 20%. Company figure, 2019.
Before
100
With forecasts
about 120
MeasuredSolar forecasting. Open Climate Fix says its forecast, used by Great Britain's grid operator, halved forecasting errors. Its own estimate.
MeasuredData-centre cooling. DeepMind's controls cut the energy Google used for cooling by up to 40%. Company figure, 2016.

Batteries

PilotedFinding a new electrolyte. Microsoft and a US national lab screened 32 million candidate materials down to 18 in 80 hours. The best may use as much as 70% less lithium.
Candidates
32 million
Worth testing
18, in 80 hours
The lithium figure is "may", from the companies involved. A battery in a lab is years from a battery in a car.

Heavy energy use and damage that could shrink within five years

MeasuredWeather forecasting. Europe's weather centre says its AI forecast uses about 1,000 times less energy to run than its physics model. NOAA reports up to 99.7% less computing for a 16-day forecast.
Physics model
1,000
AI model
1
PilotedAircraft contrails. On flights that followed the AI's plan, contrails fell 62%. Across all eligible flights the cut was 11.6%, because few followed it.
Followed the plan
62% fewer
All eligible flights
11.6% fewer
MeasuredBuilding heating and cooling. One New York office tower cut its heating and cooling electricity 15.8% over 11 months. A Google trial at two buildings saved about 9% and 13%.
MeasuredFarm chemicals. John Deere says its camera-guided sprayers covered 5 million acres in 2025 and saved 31 million gallons of herbicide mix. Company figure.
MeasuredDriving routes. Google estimates that fuel-efficient routing in Maps avoided over 3 million tonnes of CO2 in 2025. This is Google's own modelled estimate.
ForecastThe big totals. A London School of Economics study projects cuts of 3.2 to 5.4 billion tonnes of CO2 a year by 2035. The IEA's figure is up to 1.4 billion.
CaveatThe IEA says there is "no existing momentum" that would deliver its own projected cuts. They happen only if the tools are widely adopted.

The measured items are real, and most are small. The large numbers are all forecasts, and they depend on companies and governments choosing to use the tools.

A 2026 review for the campaign group Beyond Fossil Fuels went through 154 industry statements that AI helps the climate. It judged 74% of them unproven.

CaveatThe same tools find oil. Chevron credits AI with a gain of over 30% in drilling efficiency in the Permian basin, and Equinor with 27 new discoveries, DeSmog reports.

The materials work has critics inside the field too. Two University of California materials scientists reviewed the 2.2 million predicted crystals and found "scant evidence" of compounds that were new, believable and useful all at once.

The strongest environmental result so far is the least glamorous one: an AI weather forecast that runs on about a thousandth of the energy of the model it stands beside.

It replaces a heavy computing job outright, it's in daily use, and the figure comes from a public agency.

What Did the Forecasts Miss About People? Habits, Incentives & Who Does the Reporting?

My read

Scott Covert in his bucket hat, hand on chin, one eyebrow raised, looking skeptical

An environmentalist can like the direction. Every measure of what one job costs is falling, across separate sources with independent researchers among them, and a few uses already replace something far hungrier.

The total is another matter. Use is growing faster than the savings, some of the new power is gas, and most of the benefit side is still forecasts.

Two numbers will settle it: total data-centre electricity, and emission cuts that were measured after the fact. Neither points clearly yet.

The figures behind the top chart

For how to keep your own AI use lean, see AI Task Efficiency: Match the Model to the Job.

Sources

Every figure was checked against its source page on November 18, 2026. International Energy Agency figures are as reported by the publications named, because the agency's own pages could not be read directly.

Energy and water for each job

Why the curves fall

Totals, gas and water stress

What the work buys, and the criticism