Sustainability

The Carbon Cost of Intelligence: AI, Energy, and the Gap Between Net-Zero Pledges and Fossil Fuel Reality

In 2024, American data centres consumed more electricity than Pakistan. By 2030, the global total will double — and most of the new demand will be met by natural gas.

P J Laszkowicz

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

These are not speculative projections from advocacy groups. They come from the International Energy Agency, Goldman Sachs, and a three-year 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 required to power the current rate of AI expansion does not exist, and building it at the pace the industry demands will overwhelmingly rely on fossil fuels.

Goldman Sachs estimates that sixty per cent of the new power capacity required to meet data centre demand will come from natural gas, with the remaining forty per cent from renewables.4 The nuclear deals that have dominated headlines — Microsoft at Three Mile Island, Amazon at Susquehanna, Google's small modular reactors — represent a fraction of the total new capacity. The bulk of the expansion, in the near term, will burn gas.

This article examines what happened in the second half of 2025, as the scale of AI's energy demand became impossible to ignore and the gap between corporate net-zero pledges and actual energy sourcing became the defining contradiction of the industry.

The measurement problem

For years, the environmental cost of artificial intelligence was discussed in generalities. AI was "energy-intensive." Data centres "consumed a lot of water." The carbon footprint was "significant." The problem was that no one could agree on numbers, because the companies that operated the largest AI systems disclosed as little as possible about their energy consumption, and independent measurement was technically difficult.

In November 2025, that changed. A team led by Fengqi You at Cornell University published a comprehensive analysis in Nature Sustainability that, for the first time, mapped the carbon and water footprints of AI data centres across the United States with projections through 2030.5 The findings were stark. By 2030, AI data centres would emit between 24 and 44 million metric tonnes of carbon dioxide annually — equivalent to adding five to ten million cars to American roads.6 Water consumption would reach 731 to 1,125 million cubic metres per year, equal to the annual household water usage of six to ten million Americans.7

The study also showed that the problem was accelerating faster than previous estimates had suggested. AI electricity use was projected to grow seven to seventeen times between 2024 and 2030, carbon emissions 2.5 to seven times, and water use six to thirteen times.8 The ranges were wide because they depended on assumptions about the pace of AI adoption and the energy mix of the grids powering data centres — but even the most optimistic scenario described a dramatic escalation.

The IEA's own Energy and AI report, published in the same period, confirmed the broad picture. Data centre electricity consumption had grown at twelve per cent annually over the previous five years, and the agency projected this to accelerate to fifteen per cent annually through 2030.9 The United States and China together accounted for seventy per cent of global data centre electricity consumption in 2024 and would represent nearly eighty per cent of the growth to 2030.10 Europe's data centre consumption, while smaller in absolute terms, was projected to grow by more than seventy per cent.11

The measurement problem had not been fully solved — companies still disclosed less than researchers needed, and the distinction between AI workloads and conventional computing was difficult to draw in shared facilities. But the order of magnitude was now clear: AI was not a marginal addition to the energy system. It was a structural transformation of electricity demand, concentrated in a handful of countries and driven by a handful of companies.

The nuclear bet

The most visible corporate response to the energy crisis was a series of nuclear power agreements that would have seemed implausible five years earlier. In September 2024, Microsoft and Constellation Energy signed a twenty-year power purchase agreement to restart Unit 1 of the Three Mile Island nuclear plant in Pennsylvania — the unit that was not involved in the 1979 partial meltdown but had been shut down in 2019 when its former owner deemed it no longer economically viable.12 The deal would provide up to 835 megawatts of carbon-free electricity to power Microsoft's AI data centres in the mid-Atlantic region. Constellation rechristened the facility the Crane Clean Energy Center and targeted a 2028 restart, supported by a $1.5 billion loan from the US Department of Energy.13

Amazon followed with an even larger commitment. In June 2025, Amazon announced a $20 billion investment in two data centre complexes in Pennsylvania, anchored by an expanded agreement with Talen Energy for up to 1,920 megawatts of electricity from the Susquehanna nuclear plant — one of the largest nuclear facilities in the country.14 The contract ran through 2042 and included provisions for Talen and Amazon to explore building new small modular reactors within Talen's Pennsylvania footprint.15 The arrangement raised questions about grid equity: the Federal Energy Regulatory Commission scrutinised whether diverting nuclear power to higher-paying corporate customers through "behind-the-meter" connections would leave insufficient capacity for residential consumers and raise electricity prices for everyone else.16

Google took a different technological path. In October 2024, it signed the first corporate agreement for multiple deployments of small modular reactors, partnering with Kairos Power for a 500-megawatt fleet of advanced nuclear plants using molten fluoride salt-cooled reactor technology.17 The first reactor, a 50-megawatt unit, was planned for Oak Ridge, Tennessee, with three subsequent plants each containing two 75-megawatt reactors, and the full fleet operational by 2035.18 None of these reactors existed yet. Kairos was still building its test reactors, with the first expected online in 2027.

Meta completed the quartet. In December 2024, it issued a request for proposals seeking one to four gigawatts of new nuclear capacity, receiving over fifty qualified submissions.19 By mid-2025, Meta had selected TerraPower for up to eight 345-megawatt Natrium reactors totalling 2.8 gigawatts, Oklo for 1.2 gigawatts of Aurora microreactors, and Constellation for a twenty-year agreement securing 1,121 megawatts from the Clinton Clean Energy Center in Illinois.20

The combined scale was remarkable: more than ten gigawatts of new or reactivated nuclear capacity committed by four companies in roughly eighteen months. But the timelines told a more cautious story. Microsoft's Three Mile Island restart was not expected until 2028. Google's first SMR would not deliver power until 2030. Meta's TerraPower reactors were a decade away. In the interim — the period of most rapid AI expansion — the electricity would come from whatever was available. And what was available, in most of the United States, was natural gas.

The carbon-free ambiguity

The nuclear commitments revealed a deeper tension in how technology companies discuss their environmental impact. Every nuclear deal was announced as a step toward "carbon-free" or "clean" energy. And nuclear power is, in operational terms, carbon-free: it produces no direct carbon dioxide emissions during generation. But nuclear power is not without environmental consequences. It produces radioactive waste that remains hazardous for thousands of years. It consumes enormous quantities of water for cooling. The construction of new facilities is carbon-intensive. And the concentration of nuclear purchasing power among four technology companies raised questions about who would have access to carbon-free electricity and who would be left with the fossil-fuelled remainder.

The IEA noted that while nuclear energy would play a role in meeting data centre demand, its contribution would be modest relative to the total. In the agency's projections, natural gas would remain the dominant source of new generation capacity for data centres through 2030.21 The nuclear deals were significant as signals of corporate intent. They were less significant as solutions to the immediate emissions problem.

Europe's regulatory response

While American technology companies negotiated private power agreements with nuclear operators, Europe pursued a fundamentally different approach: regulation. The tools were less dramatic but potentially more consequential — mandatory renewable energy targets, efficiency standards, and reporting requirements that applied to every data centre operator, not just those wealthy enough to buy their own power plants.

Germany led the way. The German Energy Efficiency Act, which entered force in November 2023, imposed the most stringent requirements on data centres of any national legislation in the world. Data centre operators were required to source fifty per cent of their electricity from renewable energy beginning January 2024, rising to one hundred per cent from January 2027.22 The law also mandated energy efficiency targets, requiring a power usage effectiveness ratio of 1.3 for new data centres and 1.2 for facilities built after 2030, and imposed obligations to reuse waste heat.23 A Columbia Law School analysis noted that Germany had chosen "to take a more proactive approach" than the EU directive required, addressing "the regulatory gap left by the EED" with "comprehensive and legally enforceable provisions."24

The EU's own instruments were more incremental. The Energy Efficiency Directive, recast in 2023, introduced monitoring and reporting obligations for data centres but stopped short of binding performance standards or renewable energy mandates.25 Data centre operators with an installed IT power demand of 500 kilowatts or more were required to report energy consumption, water usage, waste heat utilisation, and power usage effectiveness to a European database — but the directive did not specify what should be done with this information.26 The European Commission signalled that binding standards would follow: a Data Centre Energy Efficiency Package, planned for March 2026, would introduce a rating scheme for data centres, minimum performance standards, and criteria linking permitting and public support to energy efficiency, water use, and circularity.27

The Climate Neutral Data Centre Pact, an industry self-regulatory initiative supported by the European Commission, set more ambitious targets on paper: signatories committed to matching seventy-five per cent of their electricity demand with renewable energy by December 2025 and one hundred per cent by December 2030.28 The Pact represented more than eighty-five per cent of European data centre capacity. But it was voluntary, and its targets were aspirational rather than enforceable.

The AI Act's environmental provisions

The EU AI Act, which entered force in stages from August 2024, included environmental sustainability provisions that were significant in principle but limited in practice. The legislation directed the European Commission to facilitate the development of harmonised standards for energy-efficient development and deployment of AI systems, and required periodic reports on progress toward these standards.29 But the first such report was not due until August 2028 — three years after the legislation's entry into force and well into the period of most rapid AI energy demand growth.30

Critics argued that the AI Act represented a missed opportunity. The Heinrich Böll Foundation noted that the legislation relied on standardisation bodies to develop environmental standards, and that "a significant proportion of members of these organisations come from for-profit organisations," creating structural incentives to favour solutions that minimised compliance costs rather than environmental impact.31 The Green Software Foundation's AI Committee observed that voluntary codes of conduct for energy-efficient programming, while welcome, were unlikely to alter the fundamental economics driving AI energy consumption.32

The European approach — Germany excepted — remained a work in progress. The regulatory architecture existed in outline: reporting obligations were in place, binding standards were planned, and the political will to regulate appeared genuine. But the gap between the pace of regulatory development and the pace of AI infrastructure deployment was widening.

The efficiency paradox

In January 2026, a Chinese AI lab called DeepSeek released a frontier reasoning model that matched or exceeded the performance of models costing hundreds of millions of dollars to train — for approximately six million dollars.33 The announcement sent shockwaves through financial markets and prompted immediate discussion of its implications for AI's energy trajectory. If competitive AI could be built at a fraction of the cost, perhaps the environmental crisis was solving itself.

It was not. Within days of DeepSeek's release, Microsoft CEO Satya Nadella invoked Jevons paradox — the nineteenth-century observation that as coal engines became more efficient, total coal consumption increased rather than decreased, because efficiency made the engines useful for a wider range of applications.34 The same dynamic, Nadella suggested, would apply to AI. Cheaper, more efficient models would not reduce total energy demand. They would expand the addressable market for AI, making it economically viable for applications and organisations that could not previously afford it, and driving aggregate demand higher.

A peer-reviewed analysis published in the proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency examined this question directly. The authors found that "demand for compute grows faster than efficiency gains, leading to higher total energy use and ultimately increased carbon emissions — even as each individual operation becomes 'greener.'"35 The paper noted that much of the debate about AI's environmental impact had "concentrated on direct impacts like energy and water usage in data centres without addressing significant indirect effects" — including the energy embedded in manufacturing AI hardware, the expansion of AI into new sectors, and the rebound effects of lower costs.36

The efficiency paradox did not mean that efficiency gains were valueless. The Cornell study showed that combining strategic facility siting, grid decarbonisation, and operational efficiency could reduce AI's carbon footprint by roughly seventy-three per cent and its water footprint by eighty-six per cent.37 But these reductions required deliberate policy intervention — not just market-driven efficiency improvements. Without regulatory frameworks that capped total emissions or mandated renewable energy sources, efficiency gains would be consumed by growth.

Who pays?

The aggregate numbers — terawatt-hours, millions of tonnes of carbon dioxide, billions of litres of water — obscured a more specific question: who bore the costs? The environmental burden of AI infrastructure was not distributed evenly. It fell disproportionately on communities that hosted data centres but derived little benefit from the AI systems those centres powered.

In Northern Virginia — home to the largest concentration of data centres in the world, with approximately two hundred operational facilities in Loudoun County alone — data centres consumed close to two billion gallons of water in 2023, a sixty-three per cent increase from 2019.38 Loudoun County's facilities accounted for roughly nine hundred million gallons.39 The water came from the same sources that supplied drinking water to residential communities. As data centre demand grew, so did competition for a finite resource.

In Ireland, data centres had come to consume between twenty and twenty-five per cent of the national electricity grid — a concentration that left the grid vulnerable to shortages and prompted some facilities to install diesel backup generators, further undermining the environmental case for the industry.40 In the Netherlands, Meta abandoned a planned hyperscale data centre at Zeewolde in 2022 after a national backlash over building on polder farmland, followed by a broader government clampdown on hyperscale data centre construction.41

The pattern was consistent across jurisdictions. A World Resources Institute analysis found that local communities bore costs — water competition, higher electricity prices, air pollution from backup generators — that were dispersed regionally, while the tax benefits of hosting data centres accrued locally and often to a narrow fiscal base.42 The Lincoln Institute of Land Policy documented how towns and counties competed to attract data centres through tax incentives, often without fully accounting for the infrastructure costs that would follow.43

The water question

Water consumption deserved particular attention because it was less visible than electricity use and more directly consequential for affected communities. Data centres used water primarily for evaporative cooling — a process that consumed water permanently, unlike electricity generation, which could in principle be shifted to renewable sources.

The scale of per-query water consumption was contested. A widely cited study from the University of California, Riverside, estimated that a conversation of twenty to fifty queries with ChatGPT consumed approximately 500 millilitres of water.44 OpenAI CEO Sam Altman offered a dramatically lower figure, claiming an average query used "roughly one fifteenth of a teaspoon" — approximately 0.3 millilitres.45 The discrepancy reflected different methodological choices about what to include: direct cooling water at the data centre, water consumed in electricity generation, or water embedded in the manufacturing of hardware. But even the lower estimates, multiplied across billions of daily AI interactions, produced significant aggregate demand.

Morgan Stanley projected that AI-related data centres could consume more than one trillion litres of water annually by 2028, an elevenfold increase from 2024.46 Nearly half of the world's nine thousand data centres were already located in regions of high water stress.47 The geography of data centre placement — driven by proximity to network infrastructure, cheap electricity, and favourable tax regimes — rarely aligned with the geography of water abundance.

A trajectory, not a destiny

In November 2025, the Berlin Digital Sovereignty Summit brought together digital ministers from twenty-three EU member states alongside more than a thousand participants from politics, business, and civil society.48 The summit produced twelve billion euros in voluntary corporate commitments for AI, cloud, and sovereign infrastructure, and the European Commission confirmed twenty billion euros for the establishment of AI Giga-Factories — large-scale compute facilities intended to anchor European AI development.49

The announcements were framed as investments in Europe's digital future. But the environmental implications were largely unaddressed. Twenty billion euros in new AI infrastructure meant new data centres, new electricity demand, and new carbon emissions. Whether these facilities would meet the standards Germany had set — one hundred per cent renewable energy by 2027 — or fall back on the natural gas that Goldman Sachs projected would power most new capacity globally, depended on regulatory choices that had not yet been made.

The trajectory of AI energy consumption is clear. It is not sustainable in its current form. Data centre electricity demand is doubling in a decade. The majority of new capacity is being met by fossil fuels. The water, land, and grid impacts fall disproportionately on communities that have little voice in the decisions driving AI expansion. The nuclear commitments, while symbolically important, will not deliver power at scale until the late 2020s at the earliest.

But the trajectory is not a destiny. The Cornell study demonstrated that policy interventions — strategic siting, grid decarbonisation, and operational efficiency — could reduce AI's environmental footprint by the majority. Germany's Energy Efficiency Act demonstrated that binding renewable energy mandates for data centres were technically feasible and politically achievable. The EU's planned Data Centre Energy Efficiency Package, if it includes enforceable performance standards rather than voluntary targets, could extend this approach across the continent.

The question is not whether AI's environmental cost can be managed. The evidence suggests it can. The question is whether the political and economic incentives favour management or expansion — whether the industry will be required to internalise its environmental costs, or whether it will continue to externalise them to communities, water systems, and electrical grids that bear the burden without sharing in the benefit. In the second half of 2025, the measurements arrived. The regulatory frameworks are still catching up. The gap between the two defines the current moment — and the trajectory of what comes next.

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. Goldman Sachs Research, "AI to Drive 165% Increase in Data Center Power Demand by 2030," September 2025. ↩

  5. Fengqi You et al., "The Carbon and Water Footprints of Data Centers and What This Could Mean for Artificial Intelligence," Nature Sustainability, November 2025. ↩

  6. Cornell Chronicle, "'Roadmap' Shows the Environmental Impact of AI Data Center Boom," 10 November 2025. ↩

  7. Fengqi You et al., "The Carbon and Water Footprints of Data Centers," Nature Sustainability, November 2025. ↩

  8. Ibid. ↩

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

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

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

  12. Data Center Dynamics, "Three Mile Island Nuclear Power Plant to Return as Microsoft Signs 20-Year, 835MW AI Data Center PPA," 20 September 2024. ↩

  13. WinBuzzer, "Trump Administration Grants $1B Loan to Restart Three Mile Island Nuclear Plant for Microsoft," 19 November 2025. ↩

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

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

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

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

  18. Power Magazine, "Google Bets Big on Nuclear: Inks Deal with Kairos Power for 500-MW SMR Fleet to Power Data Centers," October 2024. ↩

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

  20. Energy Intelligence, "Meta to Incentivize 4.4 GW of New US Nuclear Capacity," June 2025. ↩

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

  22. Columbia Law School Climate Law Blog, "From EU Framework to National Action: How Germany Regulates Data Center Energy Use," 24 October 2025. ↩

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

  24. Columbia Law School Climate Law Blog, "From EU Framework to National Action," 24 October 2025. ↩

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

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

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

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

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

  30. Ibid. ↩

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

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

  33. DeepSeek, "DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning," January 2026. ↩

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

  35. "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. ↩

  36. Ibid. ↩

  37. Cornell Chronicle, "'Roadmap' Shows the Environmental Impact of AI Data Center Boom," 10 November 2025. ↩

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

  39. Ibid. ↩

  40. ST Partners, "Public Perception and Data Centre Delays," 2025. ↩

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

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

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

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

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

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

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

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

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