AI’s Energy Bill: Data Centres, Water, and Carbon

The Scale of the Problem

The International Energy Agency (IEA) — the Paris-based intergovernmental body that tracks global energy data — published a landmark report on April 10, 2025. It estimated that the world’s data centres consumed about 415 terawatt-hours (TWh) of electricity in 2024 — roughly 1.5 percent of all the electricity consumed on Earth, more than the entire country of Italy uses in a year. The IEA’s base-case projection is that this will roughly double by 2030, reaching about 945 TWh — a growth rate of about 15 percent per year, more than four times faster than the growth of total electricity consumption.

In the United States, the numbers are even more striking. The Electric Power Research Institute (EPRI), in its “Powering Intelligence 2026” report, projected that data centres will consume 9 to 17 percent of all US electricity by 2030, up from about 4 to 5 percent today. The range is wide because the future of AI is uncertain — if AI adoption accelerates, the high end of the range is more likely; if it plateaus, the low end is more likely. But even the low end — 9 percent — would represent a doubling of data centre electricity consumption in just six years, and it would make data centres one of the largest categories of electricity consumer in the United States.

To put these numbers in perspective: a typical hyperscale data centre — the very large facilities operated by Amazon, Google, Microsoft, or Meta, each drawing about 100 megawatts (MW) of power, roughly the electricity of 100,000 households — is the unit of scale here. The Stargate Project, announced in January 2025, plans to build data centres with a total capacity of about 10 gigawatts (GW) — equivalent to the output of several nuclear power plants, and enough to power millions of homes. The scale of the infrastructure being built to support AI is staggering, and it is growing rapidly.

Why AI Is So Energy-Hungry

To understand why AI consumes so much energy, it helps to understand the two phases of an AI system’s life: training and inference.

Training is the process of building an AI model by feeding it enormous amounts of data — billions of words of text, millions of images, thousands of hours of audio — and adjusting its internal parameters (also called “weights”) to better predict the patterns in that data. The model is shown the data over and over again, and each pass requires massive computation. For a large language model like GPT-3, academic estimates put the training energy at roughly 1,287 megawatt-hours of electricity — roughly the amount that 120 average American households use in a year. For GPT-4, which is much larger and whose exact specs OpenAI has not published, estimates of training emissions range from about 5,184 tonnes of CO₂-equivalent (per the Stanford AI Index 2025, using Epoch AI methodology) to roughly 14,994 tonnes of CO₂-equivalent (an independent estimate by Kasper Groes Albin Ludvigsen, as cited by researcher Andy Masley).

Inference is the process of using the trained model. Every time you ask ChatGPT a question, every time you generate an image with DALL-E, every time you use an AI-powered search engine, the model performs computation to produce the response. This computation is much less than training — a single query uses a tiny fraction of the energy of training — but it adds up. Amazon Web Services has estimated that inference accounts for 80 to 90 percent of the total cloud compute cost of machine learning over a model’s lifetime.

Inference

Inference is the phase of using a trained AI model to produce outputs — every ChatGPT query, every DALL-E image, every AI-powered search result. A single inference uses a tiny fraction of the energy of training, but because a successful model serves millions of users making billions of queries, the cumulative energy of inference over a model’s lifetime dwarfs the energy of training. AWS estimates inference at 80–90% of total cloud ML cost.

The energy consumption of AI is driven by the scale of the models and the scale of their use. The models are getting larger — GPT-3 had 175 billion parameters, GPT-4 is believed to be much larger, and future models will be larger still. The use is growing even faster — ChatGPT reached 100 million users in two months, and the total number of AI queries being made worldwide is growing exponentially. Each query requires computation, each computation requires electricity, and each electron requires a power plant somewhere to generate it.

The chips that perform this computation are also power-hungry. A single NVIDIA H100 GPU — the workhorse of the AI industry — has a maximum power draw of 700 watts. A data centre might have tens of thousands of these GPUs, each running at or near maximum power for most of the day. The heat they generate must be removed — which requires cooling, which requires more energy, which requires more power plants.

Water: The Hidden Cost

Electricity is the most visible cost of AI, but it is not the only one. Water is the hidden cost — and it is significant.

Data centres use water for cooling. The servers in a data centre generate enormous amounts of heat, and that heat must be removed to prevent the equipment from failing. The most common cooling method is evaporation: water is sprayed or dripped over cooling pads, and as it evaporates, it absorbs heat. This is the same principle that sweating uses to cool the human body.

The scale of water consumption is large. A medium-sized data centre might consume about 110 million gallons of water per year for cooling. Large data centres can consume up to 5 million gallons per day. In total, US data centres consumed an estimated 17.4 billion gallons of water in 2023 — roughly the water usage of 160,000 American households.

The water consumption of AI extends to the individual query level. A study by researchers at the University of California, Riverside, led by Shaolei Ren (“Making AI Less ‘Thirsty’”, arXiv:2304.03271), estimated that GPT-3 “drinks” about 500 millilitres of water — roughly a standard water bottle — for every 10 to 50 medium-length responses it generates. This figure includes the water used for cooling in the data centres that run the model, and it has been widely cited in the press. A later analysis by Andy Masley contested this figure, estimating that the actual water cost per GPT-4 prompt was closer to 15 millilitres. The difference reflects different assumptions about data centre efficiency, location, and cooling technology. The truth probably lies somewhere in between, and it depends heavily on where the data centre is located — a data centre in a cool, humid climate uses much less water for cooling than one in a hot, dry climate.

The water consumption of AI is not just an environmental issue. It is a social and political issue. Data centres are often built in areas that are already water-stressed — because those areas also tend to have cheap land, cheap electricity, and favourable tax policies. When a data centre draws millions of gallons of water from a local aquifer, it competes with local residents, farmers, and ecosystems for a scarce resource. This has led to conflicts in several parts of the United States, particularly in the Southwest, where water is already scarce and where data centre construction is accelerating.

Carbon: The Climate Impact

The carbon footprint of AI has two components: operational emissions (from generating the electricity that powers the data centres) and embodied emissions (from manufacturing the hardware — the chips, the servers, the networking equipment — that goes into the data centres).

For electricity, the carbon footprint depends on the energy source. A data centre powered by coal has a much larger carbon footprint than one powered by solar, wind, or nuclear. This is why the major technology companies have been racing to sign contracts for clean energy — and, increasingly, for nuclear power, which provides the constant, reliable, carbon-free electricity that data centres need.

For hardware, the carbon footprint is significant but less well-measured. Manufacturing a single microchip requires about 2.1 to 2.6 gallons of water and significant amounts of energy. The chips used in AI accelerators are among the most complex ever manufactured, and their production has a substantial carbon footprint. The most rigorous study of this — Luccioni et al.’s 2023 carbon-footprint analysis of the BLOOM language model, published in the Journal of Machine Learning Research — found that the embodied (hardware) carbon of training a large language model can be comparable to, and in BLOOM’s case exceed, the operational carbon. In other words: the emissions from manufacturing the chips used in training can rival the emissions from the electricity used during training.

The major technology companies have made ambitious climate commitments. Microsoft pledged in 2020 to be carbon negative (removing more carbon from the atmosphere than the company emits), water positive (replenishing more water than the company consumes), and zero waste by 2030. Google pledged to achieve 24/7 carbon-free energy by 2030. Meta pledged to reach net-zero emissions across its value chain by 2030. Amazon pledged to reach net-zero carbon by 2040.

These commitments are in tension with the AI boom. Microsoft’s 2024 sustainability report revealed that its emissions had grown by 29.1 percent since 2020, driven by the construction of more data centres. Google’s 2024 environmental report showed a 48 percent increase in emissions since 2019. Both companies acknowledged that AI was a major driver of the increase, and both admitted that their 2030 targets were at risk.

This is the central tension of AI’s energy bill: the technology companies that are building AI are also committed to reducing their carbon footprints, and the two goals are in conflict. AI requires more data centres, more chips, more electricity, and more water. Climate commitments require less of all of these. Resolving this tension — or failing to — will be one of the defining challenges of the AI era.

The Nuclear Solution

The tension between AI’s energy demands and climate commitments has led the major technology companies to nuclear power. Nuclear power provides what AI data centres need: large amounts of constant, reliable, carbon-free electricity. Unlike solar and wind, which are intermittent (they only produce electricity when the sun shines or the wind blows), nuclear power runs 24 hours a day, 7 days a week. This makes it ideal for data centres, which also run 24/7.

The Microsoft–Three Mile Island deal, announced on September 20, 2024, was the first of its kind. Constellation Energy, the operator of the Three Mile Island nuclear plant, signed a 20-year power purchase agreement (PPA) — a long-term contract to buy electricity from a specific generator — with Microsoft to restart Three Mile Island Unit 1. The reactor, which had been shut down in 2019 for economic reasons, would be brought back online to power Microsoft’s data centres. The reactor was renamed the Crane Clean Energy Center, in honour of Chris Crane, the former CEO of Constellation’s former parent company Exelon (Constellation spun off from Exelon in 2022), who had died on April 13, 2024. In November 2025, the Trump administration backed the restart with a $1 billion federal loan from the Department of Energy (separate from the roughly $1.6 billion Constellation itself is spending on the restart).

Why “always-on” matters

The grid has a fundamental constraint: electricity must be generated at the exact moment it is used. Solar produces only during daylight; wind produces only when the wind blows. Batteries can store some surplus, but not at the scale and duration AI data centres need. Nuclear plants, by contrast, run at near-full capacity for 18–24 months between refuelling outages — making them the only zero-carbon source that matches a data centre’s 24/7 demand profile. That is why every major AI company is signing nuclear PPAs.

The Microsoft deal was followed by a wave of similar agreements. In October 2024, Google signed an agreement with Kairos Power — an SMR startup — to purchase 500 megawatts from up to seven small modular reactors (SMRs), coming online between 2030 and 2035. This was described as the world’s first tech-company deal for new nuclear power. Also in October 2024, Amazon announced three SMR deals investing more than $500 million (with X-energy/Energy Northwest in Washington State and Dominion Energy in Virginia).

Small modular reactor (SMR)

An SMR is a new class of nuclear reactor — smaller than conventional plants (typically under 300 MW per module) and designed to be factory-built in standard units, then shipped to site for assembly. The factory-built model promises lower cost and faster deployment than the bespoke, site-built reactors of the 20th century. SMRs are not yet commercial: no SMR has begun commercial operation in the United States, and the first tech-company SMR deals (Google–Kairos, Amazon–X-energy) target the 2030–2035 timeframe.

In June 2025, Meta signed a 20-year agreement with Constellation for 1,121 megawatts from the Clinton Clean Energy Center in Illinois (announced June 3, 2025). Also in June 2025, Talen Energy and Amazon expanded their existing power purchase agreement to 1,920 megawatts of nuclear power through 2042, a deal valued at about $18 billion over 17 years.

These deals represent a significant shift in the technology industry’s relationship with nuclear power. For decades, Silicon Valley was largely indifferent to nuclear energy. Now, the largest technology companies in the world are becoming the most important customers for nuclear power, and their demand is driving investment in both existing reactors and new technologies like SMRs.

The nuclear deals are not without controversy. The Federal Energy Regulatory Commission (FERC) — the US regulator of interstate electricity markets — rejected an expanded interconnection deal for the Talen–Amazon data centre project at the Susquehanna nuclear plant on November 1, 2024, citing concerns about grid reliability and the diversion of power from the broader grid to a single customer. The decision highlighted the tension between the technology companies’ desire for dedicated power and the broader public interest in a reliable, affordable electricity grid.

The Grid Under Strain

The growth of AI data centres is putting strain on the electrical grid in ways that go beyond the technology companies’ own commitments. The grid was not designed for the kind of concentrated, constant, large-scale demand that data centres create, and the growth of data centre load is creating problems for grid operators, regulators, and consumers.

The most visible strain is in Northern Virginia, which has been called “Data Center Alley” because of the concentration of data centres in the area. Northern Virginia has more than 665 data centre facilities (a figure that includes operational, approved, and planned sites), and the region’s grid operator, PJM Interconnection — the grid operator for Northern Virginia and the broader mid-Atlantic region — has struggled to keep up with the demand. PJM’s congestion costs rose 64 percent in 2024, to $1.7 billion. These costs are ultimately passed on to consumers, meaning that ordinary electricity customers are paying higher bills because of the data centres’ demand.

The strain is not limited to Virginia. In Texas, the grid operator ERCOT (Electric Reliability Council of Texas) is planning more than $30 billion in transmission upgrades to handle data-centre growth — a figure that reflects broader Texas strategic transmission ambitions, including 765 kV build-out; ERCOT’s own official Regional Transmission Plan identifies a smaller set of nearer-term projects. Across the United States, grid operators are warning that the pace of data centre construction is outstripping the pace of grid expansion, and that the gap will lead to reliability problems if it is not addressed.

The grid strain is also creating political tensions. Some communities are resisting data centre construction, citing concerns about noise, water consumption, and the impact on local electricity rates. Some regulators are questioning whether data centres should be allowed to connect directly to power plants — bypassing the grid — or whether they should be required to participate in the broader grid, paying their share of transmission and distribution costs.

These tensions will intensify as the AI industry continues to grow. The Stargate Project alone plans to build data centres with 10 gigawatts of capacity — equivalent to the power needs of several million households. Finding the electricity to power these data centres, and building the grid infrastructure to deliver it, will be one of the major infrastructure challenges of the coming decade.

The Efficiency Paradox

One might hope that improvements in AI efficiency — better algorithms, more efficient chips, smarter software — would reduce the energy consumption of AI. But there is a paradox here, known to economists as the Jevons paradox: when a resource becomes more efficient to use, people use more of it, not less.

Jevons paradox

The Jevons paradox, first described by the English economist William Stanley Jevons in his 1865 book The Coal Question, holds that improvements in the efficiency of using a resource increase, not decrease, total consumption of that resource — because greater efficiency lowers the effective cost, which leads to more use. Jevons observed that more efficient coal-burning engines did not reduce coal consumption; they increased it, because cheaper coal power meant more coal power was used. The same dynamic now applies to AI: when models become more efficient (DeepSeek-R1, distilled models, better inference engines), the cost of using AI drops, which leads to more AI use, which leads to more total compute, which leads to more total energy. Efficiency gains do not reduce aggregate consumption — they shift where the consumption happens.

This dynamic is already visible. The AI industry has made remarkable efficiency improvements — models like DeepSeek-R1 achieve frontier-level performance with a fraction of the compute used by earlier models. But these improvements have not reduced the total energy consumption of AI. They have instead made AI more accessible and more widely used, which has increased the total demand for compute. The IEA’s projection that data centre electricity consumption will roughly double by 2030, despite continued efficiency improvements, is a manifestation of the Jevons paradox.

This does not mean that efficiency improvements are pointless. They are valuable — they reduce the cost of AI, they make AI accessible to more people, and they reduce the energy consumption of any individual AI query. But it does mean that efficiency alone will not solve the energy problem. The total energy consumption of AI will continue to grow, because the growth in AI usage will outpace the improvements in AI efficiency.

What It All Means

The energy bill of AI is not a side effect. It is a fundamental feature of the technology. AI requires computation, computation requires electricity, and electricity requires power plants, fuel, and infrastructure. The scale of the demand is large and growing, and it is creating challenges that go beyond the technology industry — touching energy policy, climate policy, water policy, and national infrastructure planning.

The nuclear deals — Microsoft–Three Mile Island, Google–Kairos, Amazon–Talen, Meta–Constellation — are a recognition that the energy demands of AI cannot be met by the current grid alone. They are also a sign that the technology industry is taking its energy needs seriously, and is willing to make unprecedented investments to secure the power it needs. But the nuclear deals are not a complete solution. They take years to come online, they face regulatory hurdles, and they are controversial in the communities where they are located.

The climate commitments of the major technology companies — carbon negative by 2030, net zero by 2030, carbon free by 2030 — are in tension with the energy demands of AI. Microsoft’s emissions have grown 29 percent since 2020. Google’s have grown 48 percent since 2019. These increases are driven by AI, and they call into question whether the 2030 targets are achievable. If they are not, the technology companies will face a choice: scale back their AI ambitions, or abandon their climate commitments. Neither choice is attractive.

The water consumption of AI is less visible than the electricity consumption, but it is no less real. In water-stressed regions, the millions of gallons that data centres consume for cooling compete with the needs of residents, farmers, and ecosystems. As data centre construction accelerates in the American Southwest and other arid regions, water conflicts will become more frequent and more intense.

The grid strain is a national infrastructure issue. The United States is building data centres faster than it is building the power plants and transmission lines to serve them, and the gap is creating reliability problems, cost increases, and political conflicts. Addressing this gap will require investment, planning, and coordination between the technology industry, the energy industry, and government — at a scale that has not yet been achieved.

The AI revolution is often discussed in terms of algorithms, models, and data. But beneath these abstractions is a physical infrastructure — chips, servers, data centres, power plants, transmission lines, cooling towers, water pipes — that is vast, expensive, and environmentally consequential. The energy bill of AI is the bill for this infrastructure, and it is a bill that the world is only beginning to understand.