How much energy and water does artificial intelligence use: verified figures

The key numbers at a glance

The table gathers the figures repeated most often, each with its source and year. A figure without those two details says nothing, because AI efficiency changes fast and each institution measures a different scope. The rest of the guide explains where each row comes from and why they cannot all be compared with each other.

Why the numbers on AI do not match

The numbers differ because each source counts different things: which equipment is included, which model and type of task is tested, and whether water is measured as withdrawn or consumed. Two figures can both be correct and still not be comparable with each other.

Google shows this with its own case. In the narrow approach, which measures only the AI accelerator in a highly utilized test on the most efficient data centers, the median Gemini query comes out at 0.10 Wh. In the full approach, it rises to 0.24 Wh. The difference lies in what gets added, according to Google’s technical paper.

With the narrow rule, the same service would give 10,000 queries per kWh. With the full one, 4,167. Google leaves out external internet networking because it does not control it, and estimates that the data center’s internal networking weighs little in a text query. It also uses the median instead of the average, because a few very long queries or ones served by under-used models distort the average.

Before comparing two figures, four things need to be fixed:

• What the measurement includes: the chip only, the whole server, the data center, or also the manufacturing of the equipment.
• Which model and task: a short answer is not the same as a long text, and a reasoning model is not the same as a light one.
• Who measures: the provider itself, a lab, or an academic study estimating from outside.
• Which water is counted: the water withdrawn from a source or the water consumed.

How much electricity data centers use

According to the International Energy Agency (IEA), data centers worldwide used about 415 TWh in 2024, 1.5% of global electricity, and would reach about 945 TWh in 2030, just under 3%. That amount is slightly more than Japan’s entire electricity consumption today.

These figures cover all data centers, not only AI ones. Consumption grows about 15% a year between 2024 and 2030, more than four times the pace of electricity use in all other sectors. Even so, the IEA clarifies that data center growth explains less than 10% of the rise in global electricity demand over that period, and that a 3% share remains limited in the global picture.

AI weighs on that growth. Accelerated servers, driven mainly by AI, would grow 30% a year in the base case, versus 9% for conventional servers. Accelerated servers explain almost half of the net increase in data center consumption; conventional ones, about 20%; other IT equipment, about 10%, and cooling and support infrastructure, about 20%.

In the United States, Lawrence Berkeley National Laboratory estimated that consumption rose from 58 TWh in 2014 to 176 TWh in 2023, about 4.4% of the country’s electricity, and projects between 325 and 580 TWh in 2028. The Department of Energy summarizes that data center load tripled in a decade and could double or triple again by 2028.

What one query to an AI assistant uses

It depends on who measures it. Google measured 0.24 Wh for the median Gemini text query, and an academic study comparing 30 models found that the most intensive ones exceed 29 Wh per long query, more than 65 times what the most efficient use. The gap mixes models, text lengths and measurement rules.

The study by Jegham and colleagues does not measure inside data centers: it estimates from public performance data and company-specific environmental factors. It analyzes 30 models from OpenAI, Anthropic, Meta and DeepSeek with three query sizes. The table shows three OpenAI models and the extremes for the long query.

Two lessons come out of that table. The first is that model size matters: within the LLaMA family, consumption rises from 0.443 Wh at 8 billion parameters to 25.202 Wh at 405 billion. The second is that reasoning models use much more, and DeepSeek-R1 consumes about 65 times what the most efficient model does.

Google adds that the energy of its median query fell 33 times and its total carbon footprint 44 times in 12 months, comparing May 2024 with May 2025. In that same paper, Google cites a June 2025 post by OpenAI’s chief executive, Sam Altman, that gave 0.34 Wh and about 0.3 ml of water for an average ChatGPT query. It is a statement on a blog, with no published methodology, so it cannot be checked against the others.

How much water AI uses: the three taps

AI’s water comes from three taps: the water that cools the data center, the water used by the plants that generate its electricity, and the water spent manufacturing the servers. That is why per-query figures range from 0.26 ml at Google to 45 ml at Mistral, which also counts the life cycle. They measure different scopes.

Li and colleagues describe that split into three scopes: on-site water for cooling (scope 1), off-site water for electricity generation (scope 2) and supply-chain water for server manufacturing (scope 3). Someone who counts only the first tap gets a smaller figure than someone who adds all three.

Two concepts also need telling apart. Withdrawal is water taken from the ground or from a surface source. Consumption is withdrawal minus what is returned: what evaporates or stops being available. According to the same study, in cooling towers about 80% of what is withdrawn evaporates. Google, in its paper, explains that it considers consumption the best measure of its effect on local water, and that on average it consumes 80% of what it withdraws.

The amount that evaporates per kWh of server energy varies with the climate. Li and colleagues put it between 1 and 9 liters per kWh: 1 liter per kWh in Google’s annual average and 9 in a large commercial data center in Arizona during summer. For 2023 and 2024, Google reports 1.15 liters per kWh on its water usage indicator.

The authors compare the withdrawal projection with the water that 4 to 6 Denmarks or half the United Kingdom withdraw in a year. It is a comparison of withdrawal and not of consumption, and it should be read that way. Its source is a 2023 study updated in 2025, and the 2027 projections depend on assumptions about AI electricity use.

Model size and life cycle

Mistral calculated the life cycle of its Mistral Large 2 model up to January 2025, after 18 months of use: 20.4 kt of CO₂e and 281,000 m³ of water consumed. It also found that impact grows almost in proportion to size: a model ten times larger produces an impact one order of magnitude higher for the same amount of text.

The company ran the analysis with the consultancy Carbone 4 and the French ecological transition agency ADEME, and Resilio and Hubblo reviewed it. Mistral presents it as the first complete life-cycle analysis of an AI model. It measures three categories: greenhouse gas emissions, water use and resource depletion, and the last one gives 660 kg of antimony equivalent for Large 2.

With those results, the company proposes three indicators to inform the public: the absolute impact of training a model, the marginal impact of each answer, and the ratio between the total impact of answers and that of the whole life cycle. The first two, it says, could be mandatory reporting. The third shows whether the cost of training is spread across many answers or wasted.

The Jegham study points the same way: within one family, consumption soars with the number of parameters. Hence the practical value of choosing the right model for each use, instead of the largest by default.

A problem concentrated in a few places

The United States and China will account for nearly 80% of the growth in data center consumption through 2030, according to the IEA: about 240 TWh more in the United States, 130% above 2024, and about 175 TWh more in China, 170% more. The United States already had the highest consumption per person in 2024, about 540 kWh.

The IEA notes that, unlike electric vehicles, data centers tend to cluster in specific locations, which can make connecting them to the grid harder. A national or global average does not show the pressure felt by the grids and water sources in those areas. Southeast Asia is another example: its demand would more than double by 2030, partly because of the regional hub in Singapore and southern Malaysia.

The type of facility also counts. According to the IEA, cooling accounts for about 7% of total consumption in efficient hyperscale data centers and more than 30% in less efficient enterprise ones. Two facilities with the same servers can have different footprints because of how they are cooled.

What reduces the impact

The sources consulted point to four measurable levers: choosing a smaller model, improving software and hardware efficiency, deciding where and when the load runs, and publishing comparable metrics. None is enough alone, and the IEA warns that demand growth can outpace the improvements.

• The right model: according to Mistral, impact rises almost in proportion to model size, and in Jegham the LLaMA family goes from 0.443 Wh to 25.202 Wh between 8 and 405 billion parameters.
• Efficiency: Google attributes its 33-fold cut in energy and 44-fold cut in carbon in a year to software improvements and clean energy procurement, according to the summary of its technical paper. In the IEA’s High Efficiency scenario, the same digital service is covered with more than 15% in savings.
• Place and time: Li and colleagues show that water use varies in space and time, and that choosing when and where to train a large model can noticeably cut its water footprint.
• Transparency: Mistral and Google publish their methodology and argue for more transparency in the sector, so models can be compared and purchase decisions can be informed.

The paradox that Jegham and colleagues point out limits the effect of efficiency: as AI gets cheaper and faster, mass adoption drives resource use disproportionately. The improvement per query can be offset if the number of queries grows faster.

What is still unknown

There are four big gaps: no common standard exists to measure model impact, few companies publish their own measurements, outside estimates depend on assumptions, and the per-query figure changes almost every quarter. Any number in this guide should be read with its date.

• No standard: Mistral acknowledges that its study is a first approximation because there are no public standards or impact factors for language models.
• Own measurement: Google and Mistral measure their own systems. Google cites no outside audit of its figures, and two consultancies, Resilio and Hubblo, reviewed Mistral’s analysis.
• Outside estimates: Jegham calculates from public performance data and company factors, not from measurements in the data center.
• Projections: the IEA presents four scenarios for 2035 ranging from about 700 TWh to more than 1,700 TWh, and the range itself shows how much uncertainty there is.
• Missing figures from other providers: this guide did not find a published measurement with methodology for ChatGPT, Claude or other assistants comparable to Google’s.

How to read a headline about AI consumption

A headline about AI consumption reads well with five checks: what was measured, who measured it, which year it is from, which unit it uses and whether it refers to water withdrawn or consumed. If one is missing, the figure cannot yet be compared with another.

A simple example: 0.24 Wh and 29 Wh do not contradict each other. One is the median text query in Gemini and the other is a long query on a reasoning model, estimated from outside.

Myths and reality

• Myth: an AI query always uses the same amount. Reality: it ranges from 0.24 Wh to more than 29 Wh depending on the model, the length and how it is measured.
• Myth: AI already uses a huge share of the world’s electricity. Reality: all data centers together were 1.5% in 2024 and the IEA projects just under 3% in 2030.
• Myth: Google and Mistral give contradictory water figures. Reality: 0.26 ml and 45 ml correspond to different scopes, and the second includes the life cycle.
• Myth: all the water used is lost. Reality: withdrawal and consumption are different, and in cooling towers about 80% of what is withdrawn evaporates.
• Myth: efficiency has already solved the problem. Reality: Google cut the energy of its median query 33 times in a year, but the IEA projects data center consumption will double by 2030.
• Myth: the impact is even across the planet. Reality: the United States and China account for nearly 80% of the growth through 2030, according to the IEA.

Related articles

• Three cases in Panama: what happened when content was structured for AI
• Setting up a technology company in a Panama free zone: steps
• Panama’s innovation ecosystem: how a startup chooses a base

Sources

Checked on November 5, 2026.

• Google Cloud, Measuring the environmental impact of AI inference, August 21, 2025.
• Elsworth et al., Measuring the environmental impact of delivering AI at Google Scale, arXiv 2508.15734.
• Mistral AI, Our contribution to a global environmental standard for AI, July 22, 2025.
• Jegham et al., How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference, arXiv 2505.09598.
• Li et al., Making AI Less Thirsty, arXiv 2304.03271, accepted by Communications of the ACM.
• International Energy Agency, Energy and AI, chapter Energy demand from AI.
• International Energy Agency, release on AI and data center electricity demand.
• US Department of Energy, summary of the Lawrence Berkeley National Laboratory data center report.