Sustainability

Most of AI's new electricity will come from gas

In 2024 American data centres consumed more electricity than Pakistan. By 2030 the global total is projected to double, and Goldman Sachs expects the majority of the new capacity to be met by natural gas rather than by the nuclear agreements that have taken the headlines.

P J Laszkowicz

The numbers, plainly stated. In 2024 data centres worldwide consumed roughly 415 terawatt-hours of electricity, about 1.5 per cent of global consumption and more than the annual demand of most individual nations.1 In the United States alone the figure was 183 terawatt-hours, matching the entire electricity consumption of Pakistan.2 By 2030 the global total is projected to double, to approximately 945 terawatt-hours, driven overwhelmingly by artificial intelligence.3

These are not projections from advocacy groups. They come from the International Energy Agency, from Goldman Sachs, and from a study published in Nature Sustainability by researchers at Cornell University. The message is consistent across all three: the trajectory is unsustainable. Not in the activist sense of the word—in the engineering sense. The infrastructure needed to power AI's current rate of expansion does not exist, and building it at the pace the industry wants will rely overwhelmingly on fossil fuels.

Goldman Sachs estimates that 40 per cent of the new capacity will come from renewables, with the bulk of the remaining 60 per cent driven by natural gas. The nuclear agreements that have dominated headlines, with Microsoft at Three Mile Island, Amazon at Susquehanna, and Google's small modular reactors, account for a fraction of the total. Most of the near-term expansion will burn gas.

What follows traces the period from mid-2025 into the opening weeks of 2026, as the scale of AI's energy demand became impossible to ignore and the distance between corporate net-zero pledges and actual energy sourcing became the industry's defining contradiction.

The measurement problem

For years the environmental cost of artificial intelligence was discussed in generalities. AI was energy-intensive. Data centres used a lot of water. The carbon footprint was significant. Nobody could agree on numbers, because the companies running the largest systems disclosed as little as possible and independent measurement was technically hard.

That changed in November 2025. A team at Cornell led by Fengqi You, with doctoral researcher Tianqi Xiao as first author, published an analysis in Nature Sustainability mapping the carbon and water footprints of AI data centres across the United States with projections to 2030. The findings were stark. By 2030 these facilities would emit between 24 and 44 million tonnes of carbon dioxide a year, equivalent to putting five to ten million more cars on American roads. Water consumption would reach 731 to 1,125 million cubic metres annually, matching the household water use of six to ten million Americans.

The study also showed the problem accelerating faster than earlier estimates suggested. AI electricity use was projected to grow seven to seventeen times between 2024 and 2030, carbon emissions by two and a half to seven times, and water use six to thirteen times.4 The ranges are wide because they depend on assumptions about how fast AI is adopted and how the grids powering it are supplied. Even the most optimistic scenario describes a dramatic escalation.

The IEA's Energy and AI report confirmed the broad picture. Data centre electricity consumption had grown at 12 per cent a year over the previous five years, and the agency projected acceleration to 15 per cent annually through 2030.5 The United States and China together accounted for 70 per cent of global data centre electricity in 2024 and would represent nearly 80 per cent of the growth to 2030.6 European consumption, smaller in absolute terms, was projected to grow by more than 70 per cent.7

The measurement problem is not fully solved. Companies still disclose less than researchers need, and separating AI workloads from conventional computing in shared facilities remains difficult. The order of magnitude is now clear, though. AI is not a marginal addition to the energy system. It is a structural change in electricity demand, concentrated in a few countries and driven by a few companies.

The nuclear commitments

The most visible corporate response was a run of nuclear agreements that would have looked implausible five years earlier. In September 2024 Microsoft and Constellation Energy signed a twenty-year power purchase agreement to restart Unit 1 at Three Mile Island, the unit not involved in the 1979 partial meltdown, which had closed in 2019 when its owner judged it no longer economic. The deal covers up to 835 megawatts of carbon-free electricity for Microsoft's data centres in the mid-Atlantic. Constellation renamed the facility the Crane Clean Energy Center, and in November 2025 secured a $1 billion loan from the US Department of Energy toward an estimated $1.6 billion project cost, with the restart now targeted for 2027, a year ahead of the original schedule.

Amazon made a larger commitment. In June 2025 it announced $20 billion for two data centre complexes in Pennsylvania, anchored by an expanded agreement with Talen Energy for up to 1,920 megawatts from the Susquehanna plant, one of the largest nuclear facilities in the country.8 The contract runs to 2042 and includes provisions for exploring new small modular reactors within Talen's Pennsylvania footprint.9 It also raised questions about grid fairness. The Federal Energy Regulatory Commission examined whether routing nuclear power directly to higher-paying corporate customers, bypassing the public grid, would leave too little for residential consumers and raise prices for everyone else.10

Google took a different technical route. In October 2024 it signed the first corporate agreement covering multiple small modular reactor deployments, partnering with Kairos Power for a 500-megawatt fleet using molten fluoride salt-cooled technology.11 The first unit, at 50 megawatts, is planned for Oak Ridge, Tennessee, with three later plants each holding two 75-megawatt reactors and the full fleet operational by 2035.12 None of these reactors exists yet. Kairos is still building test units, with the first expected online in 2027.

Meta completed the group. In December 2024 it issued a request for proposals seeking one to four gigawatts of new nuclear capacity and received more than fifty qualified submissions.13 In June 2025 it signed a twenty-year agreement with Constellation for 1,121 megawatts from the Clinton Clean Energy Center in Illinois. In January 2026 it announced agreements with TerraPower, Oklo, and Vistra supporting up to 6.6 gigawatts of new and existing capacity by 2035, including a 1.2 gigawatt campus with Oklo in Pike County, Ohio, expected online as early as 2030, to serve its Prometheus data centre in New Albany.

The combined scale is remarkable: approaching ten gigawatts of new or reactivated nuclear capacity committed by four companies in roughly eighteen months. The timelines tell a more cautious story. Microsoft's restart is not expected before 2027. Google's first small modular reactor will not deliver power until 2030. Meta's newest agreements run to 2035. In the interval, which is the period of most rapid AI expansion, the electricity will come from whatever is available, and in most of the United States what is available is natural gas.

The carbon-free ambiguity

The nuclear commitments expose a deeper tension in how technology companies discuss their environmental impact. Every deal was announced as a step toward carbon-free or clean energy, and in operational terms nuclear power is carbon-free, producing no direct carbon dioxide during generation. It is not without environmental consequence. It produces radioactive waste that stays hazardous for thousands of years. It consumes large quantities of cooling water. Building new plants is carbon-intensive. And concentrating nuclear purchasing power among four technology companies raises a distributional question: who gets access to carbon-free electricity, and who is left with the fossil-fuelled remainder.

The IEA notes that while nuclear will play a role in meeting data centre demand, its contribution will be modest against the total, and natural gas will remain the dominant source of new generation capacity for data centres through 2030.14 The nuclear deals matter as signals of corporate intent. They matter less as solutions to the immediate emissions problem.

Europe's regulatory response

While American companies negotiated private power agreements, Europe took a different route: regulation. The instruments are less dramatic and potentially more consequential, because mandatory renewable targets, efficiency standards, and reporting duties apply to every operator rather than only to those wealthy enough to buy a power plant.

Germany went furthest. The German Energy Efficiency Act, in force since November 2023, imposes the most demanding requirements on data centres of any national law in the world. Operators must source half their electricity from renewables from January 2024, rising to all of it from January 2027. The Act also sets efficiency targets using power usage effectiveness, the ratio of a facility's total energy draw to the energy reaching its computing equipment, where a lower number means less overhead lost to cooling and distribution. Centres operating before July 2026 must reach 1.5 or better by July 2027 and 1.3 by July 2030; those starting later must reach 1.2. Operators must also avoid waste heat where they can and reuse what remains.15

The EU's own instruments are more incremental. The Energy Efficiency Directive, recast in 2023, introduced monitoring and reporting duties for data centres without binding performance standards or renewable mandates.16 Operators with installed computing power of 500 kilowatts or more must report energy consumption, water use, waste heat recovery, and power usage effectiveness to a European database, and the directive does not say what should be done with the information.17 The Commission has signalled that binding standards will follow. A Data Centre Energy Efficiency Package, planned for March 2026, would introduce a rating scheme, minimum performance standards, and criteria tying permits and public support to efficiency, water use, and circularity.18

The Climate Neutral Data Centre Pact, an industry initiative supported by the Commission, sets more ambitious targets on paper. Signatories, representing more than 85 per cent of European data centre capacity, committed to matching three-quarters of their electricity demand with renewables by December 2025 and all of it by December 2030.19 The Pact is voluntary, and its targets are aspirations rather than obligations.

The AI Act's environmental provisions

The EU AI Act, which entered force in stages from August 2024, carries environmental provisions significant in principle and limited in practice. It directs the Commission to help develop harmonised standards for energy-efficient AI and requires periodic reports on progress.20 The first such report is not due until August 2028, three years after entry into force and well into the period of steepest demand growth.21

Critics call it a missed opportunity. The Heinrich Böll Foundation noted that the legislation relies on standardisation bodies to develop environmental standards, and that a significant proportion of the members of those bodies come from for-profit organisations, which creates a structural pull toward solutions that minimise compliance cost rather than environmental impact.22 The Green Software Foundation's AI Committee observed that voluntary codes of conduct for energy-efficient programming, while welcome, are unlikely to shift the economics driving AI energy consumption.23

The European approach, Germany excepted, remains under construction. The architecture exists in outline, reporting duties are in place, binding standards are planned, and the political will appears genuine. The gap between the pace of regulation and the pace of deployment is widening.

The efficiency paradox

In January 2025 the Chinese laboratory DeepSeek released a frontier reasoning model matching or exceeding the performance of models costing hundreds of millions to train, for roughly six million dollars.24 The announcement moved financial markets and prompted immediate discussion of what it meant for AI's energy trajectory. If competitive AI could be built for a fraction of the cost, perhaps the environmental problem was solving itself.

It was not. Within days, Microsoft's chief executive Satya Nadella invoked Jevons paradox, the nineteenth-century observation that as coal engines became more efficient, total coal consumption rose rather than fell, because efficiency made the engines worth using for far more purposes.25 The same dynamic would apply to AI. Cheaper, more efficient models would not reduce total energy demand. They would widen the market, making AI viable for applications and organisations that could not previously afford it, and push aggregate demand higher.

A peer-reviewed analysis presented at the 2025 ACM Conference on Fairness, Accountability, and Transparency examined the question directly, finding that demand for computation grows faster than efficiency gains, "leading to higher total energy use and ultimately increased carbon emissions—even as each individual operation becomes 'greener'".26 The authors noted that debate about AI's environmental impact had concentrated on direct effects inside data centres without addressing indirect ones: the energy embedded in manufacturing AI hardware, the spread of AI into new sectors, and the rebound effects of falling costs.27

The paradox does not make efficiency worthless. The Cornell study found that combining careful siting, grid decarbonisation, and operational efficiency could cut AI's carbon footprint by about 73 per cent and its water footprint by 86 per cent. Those reductions require deliberate policy rather than market-driven improvement. Without frameworks that cap total emissions or mandate renewable supply, efficiency gains are absorbed by growth.

Who pays?

Aggregate numbers obscure a more specific question: who carries the cost? The environmental burden of AI infrastructure is not spread evenly. It falls hardest on communities hosting data centres that derive little benefit from the systems inside them.

In Northern Virginia, home to the largest concentration of data centres in the world, facilities consumed close to two billion gallons of water in 2023, a 63 per cent rise from 2019, with Loudoun County's roughly two hundred facilities accounting for some nine hundred million gallons.28 The water comes from the same sources supplying homes. As demand grows, so does competition for a finite resource.

In Ireland, data centres consumed 22 per cent of all metered electricity in 2024, up from 5 per cent a decade earlier, a concentration that leaves the grid exposed and has prompted some operators to install diesel backup generators, which undercuts the environmental case further. In the Netherlands, Meta abandoned a planned hyperscale facility at Zeewolde in 2022 after a national backlash over building on reclaimed farmland, followed by a broader government clampdown on hyperscale construction.29

The pattern holds across jurisdictions. The World Resources Institute found that local communities bear costs that spread regionally, including water competition, higher electricity prices, and air pollution from backup generators, while the tax benefits of hosting accrue locally and often to a narrow fiscal base.30 The Lincoln Institute of Land Policy documented towns and counties competing to attract data centres through tax incentives, frequently without accounting for the infrastructure costs that follow.31

The water question

Water deserves particular attention because it is less visible than electricity and more directly felt by the communities affected. Data centres use water mainly for evaporative cooling, which consumes it permanently, unlike electricity generation, which can in principle be shifted to renewable sources.

The scale per query is contested. A widely cited study from the University of California, Riverside estimated that a conversation of twenty to fifty queries with ChatGPT consumed about 500 millilitres.32 OpenAI's chief executive Sam Altman offered a far lower figure, putting an average query at roughly one fifteenth of a teaspoon, about 0.3 millilitres.33 The discrepancy reflects different choices about what to count: cooling water at the facility, water consumed generating the electricity, or water embedded in manufacturing the hardware. Even the lower estimates, multiplied across billions of daily interactions, produce substantial aggregate demand.

Morgan Stanley projected that AI-related data centres could consume more than one trillion litres of water a year by 2028, an elevenfold rise from 2024.34 Nearly half the world's nine thousand data centres already sit in regions of high water stress.35 The geography of data centre placement, driven by network proximity, cheap electricity, and favourable tax treatment, rarely matches the geography of water abundance.

What the trajectory depends on

In November 2025 the Berlin summit on European digital sovereignty brought together digital ministers from twenty-three member states alongside more than a thousand participants from politics, business, and civil society.36 It produced twelve billion euros in voluntary corporate commitments for AI, cloud, and sovereign infrastructure, alongside the Commission's twenty billion euro commitment to AI Giga-Factories, the large-scale computing facilities intended to anchor European AI development.37

The announcements were framed as investment in Europe's digital future. The environmental implications went largely unaddressed. Twenty billion euros of new AI infrastructure means new data centres, new electricity demand, and new emissions. Whether those facilities meet the standard Germany has set, of fully renewable supply by 2027, or fall back on the natural gas Goldman Sachs expects to power most new capacity worldwide, depends on regulatory choices not yet made.

The trajectory is clear, and it is not sustainable in its current form. Data centre electricity demand is doubling within a decade. Most new capacity is being met by fossil fuels. The water, land, and grid impacts fall hardest on communities with least say in the decisions driving expansion. The nuclear commitments, symbolically important, will not deliver power at scale until the late 2020s at the earliest.

A trajectory is not a destiny. The Cornell study showed that policy intervention could cut the environmental footprint by most of its projected size. Germany's Energy Efficiency Act showed that binding renewable mandates for data centres are technically feasible and politically achievable. The EU's planned efficiency package, if it carries enforceable standards rather than voluntary targets, could extend that approach across the continent.

The question is not whether AI's environmental cost can be managed. The evidence says it can. The question is whether political and economic incentives favour managing it or expanding regardless: whether the industry will be required to carry its environmental costs, or will keep passing them to communities, water systems, and electrical grids that bear the burden without sharing the benefit. Through the second half of 2025 and into 2026 the measurements arrived. The regulatory frameworks are still catching up, and the distance between the two defines the moment.

Footnotes

  1. International Energy Agency, "Energy and AI: Energy Demand from AI," October 2025. ↩

  2. Pew Research Center, "What We Know about Energy Use at US Data Centers amid the AI Boom," 24 October 2025. ↩

  3. International Energy Agency, "Energy and AI: Executive Summary," October 2025. ↩

  4. Tianqi Xiao, Fengqi You et al., Nature Sustainability, 10 November 2025. ↩

  5. International Energy Agency, "Energy and AI: Energy Demand from AI," October 2025. ↩

  6. S&P Global, "Global Data Center Power Demand to Double by 2030 on AI Surge: IEA," 10 October 2025. ↩

  7. International Energy Agency, "Energy and AI: Energy Demand from AI," October 2025. ↩

  8. CNBC, "Amazon to Spend $20 Billion on Data Centers in Pennsylvania, Including One Next to a Nuclear Power Plant," 9 June 2025. ↩

  9. Talen Energy Corporation, "Talen Energy Expands Nuclear Energy Relationship with Amazon," 11 June 2025. ↩

  10. Fortune, "Amazon's $20 Billion Bet on New Pennsylvania Data Centers Is Sparking Concern over Plugging into Nuclear Power Plants," 9 June 2025. ↩

  11. Utility Dive, "Google, Kairos Power Ink 500-MW Advanced Nuclear Reactor Deal," 15 October 2024. ↩

  12. Power Magazine, "Google Bets Big on Nuclear," October 2024. ↩

  13. Meta Sustainability, "Accelerating the Next Wave of Nuclear to Power AI Innovation," 3 December 2024. ↩

  14. International Energy Agency, "Energy Supply for AI," October 2025. ↩

  15. White & Case LLP, "Data Center Requirements under the New German Energy Efficiency Act," 2024. ↩

  16. European Commission, "Energy Efficiency Directive: Energy Performance of Data Centres," 2024. ↩

  17. White & Case LLP, "Data Centres and Energy Consumption: Evolving EU Regulatory Landscape and Outlook for 2026," 2025. ↩

  18. European Commission, "In Focus: Data Centres—An Energy-Hungry Challenge," 17 November 2025. ↩

  19. Climate Neutral Data Centre Pact, "Our Targets," 2024. ↩

  20. White & Case LLP, "Energy Efficiency Requirements under the EU AI Act," 2025. ↩

  21. Ibid. ↩

  22. Heinrich Böll Foundation, "The EU AI Act and Environmental Protection: The Case for a Missed Opportunity," April 2024. ↩

  23. Green Software Foundation, "The EU AI Act: Insights from the Green AI Committee," 2025. ↩

  24. DeepSeek-R1 was released on 20 January 2025. An earlier version of this piece dated the release to January 2026, which conflicted with the contemporaneous reporting cited in the following note. ↩

  25. NPR Planet Money, "Why the AI World Is Suddenly Obsessed with Jevons Paradox," 4 February 2025. ↩

  26. "From Efficiency Gains to Rebound Effects: The Problem of Jevons' Paradox in AI's Polarized Environmental Debate," Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency, 2025. ↩

  27. Ibid. ↩

  28. World Resources Institute, "From Energy Use to Air Quality, the Many Ways Data Centers Affect US Communities," 2025. ↩

  29. Lincoln Institute of Land Policy, "Data Drain: The Land and Water Impacts of the AI Boom," 2025. ↩

  30. World Resources Institute, "From Energy Use to Air Quality," 2025. ↩

  31. Lincoln Institute of Land Policy, "Data Drain," 2025. ↩

  32. IE Insights, "From Cloud to Cup: How Much Water Does Your ChatGPT Drink?", 2025. ↩

  33. Undark, "How Much Water Do AI Data Centers Really Use?", 16 December 2025. ↩

  34. EESI, "Data Centers and Water Consumption," 2025. ↩

  35. TRENDS Research & Advisory, "Water Implications of AI-Driven Digital Infrastructure Expansion," 2025. ↩

  36. German Federal Ministry for Digital and State Modernisation, "Summit for More Digital Sovereignty Starts in Berlin," 18 November 2025. ↩

  37. European Commission, "AI Continent Action Plan," 2025. ↩