- Training GPT-4 likely consumed ten times more electricity than GPT-3's 1,287 megawatt-hours, with each model generation requiring order-of-magnitude increases.
- Microsoft's global water consumption increased 34% between 2021 and 2022 to nearly 6.4 million cubic meters, with much attributed to AI infrastructure cooling needs.
- Microsoft plans to spend $80 billion on AI data centers in 2025 alone, with Google spending comparable amounts in unprecedented capital deployment.
- Dominion Energy is delaying coal plant retirements due to surging data center power demand, while PJM warns of potential grid reliability strain.
- Microsoft signed contracts for Three Mile Island nuclear plant power, while Amazon and Google pursue nuclear solutions for AI infrastructure energy needs.
- Despite Nvidia H100 GPUs being 3-5 times more energy-efficient than previous generations, total consumption increases as demand grows faster than efficiency gains.
The Scale of Modern AI Infrastructure
The infrastructure behind modern AI is almost incomprehensibly large. A single large-scale AI training run requires tens of thousands of specialized processors, typically Nvidia H100 or H200 GPUs, running continuously for weeks or months. These chips are housed in data centers that can cover tens of acres, consume as much electricity as a small city, and require continuous cooling to prevent catastrophic overheating.
Microsoft has announced plans to spend $80 billion on AI data centers in 2025 alone. Google is spending a comparable amount. Amazon Web Services, Meta, and Oracle are all investing tens of billions more. The scale of capital deployment is without precedent in the history of computing. Data center construction has become one of the largest categories of infrastructure spending in the world.
The numbers become more concrete when you look at individual facilities. Microsoft's data center campus in Iowa spans over a million square feet. Google is building a facility in Columbus, Ohio, that will draw over 400 megawatts of power at full capacity. Hyperscale data centers routinely require dedicated substations, new transmission lines, and upgrades to regional power grids. The infrastructure to build AI has become infrastructure in its own right.
What makes this particularly notable is the speed of growth. In 2020, AI workloads represented a small fraction of data center electricity consumption. By 2025, AI-specific hardware accounts for a significant and rapidly growing share of global data center power draw. The International Energy Agency estimates that data centers will consume over 1,000 terawatt-hours of electricity globally in 2026, roughly the electricity consumption of Japan.
The geography of AI infrastructure is also significant. Data centers cluster near cheap power, which often means fossil fuels: coal in the US Midwest, natural gas in Virginia, hydropower in the Pacific Northwest. The push to build AI fast has often overridden commitments to build it sustainably. Communities near these facilities are experiencing consequences that were not widely anticipated.
Power Consumption: The Gigawatt Problem
The power requirements of AI training are staggering. Training GPTGPTGenerative Pre-trained Transformer — the model architecture and family name behind OpenAI's most famous models, from GPT-2 to GPT-5.Learn more →-3 consumed roughly 1,287 megawatt-hours of electricity. Training GPT-4 likely consumed ten or more times that amount, though OpenAI has not published precise figures. Training GPT-5 almost certainly required another order of magnitude increase. Each generation of frontier modelFrontier ModelThe most advanced and capable AI models at the cutting edge of the field, typically developed by well-funded research labs with massive computational resources.Learn more → training is consuming more electricity than the last.
The inferenceInferenceThe process of running a trained AI model to generate outputs — what happens when you send a prompt and receive a response.Learn more → side of the equation is equally important and often overlooked. Training happens once. Inference, responding to user queries, happens continuously. ChatGPT processes over a hundred million queries per day. Each query consumes a small but non-zero amount of electricity. At scale, inference adds up to energy consumption that rivals or exceeds training. Google's AI search features alone are expected to dramatically increase the energy intensity of search queries.
The electricity demand from AI is straining grids in ways that were not anticipated. Dominion Energy, which serves northern Virginia, home to the largest concentration of data centers in the world, is delaying the retirement of coal plants because of surging power demand from data centers. PJM Interconnection, which manages the grid for 13 states in the eastern US, has warned that growing data center demand could strain grid reliability.
Nuclear power has emerged as a promising solution. Microsoft has signed a contract to purchase power from the restarted Three Mile Island nuclear plant. Amazon has acquired a data center campus directly connected to a nuclear plant in Pennsylvania. Google has signed agreements to develop small modular reactors. The idea that nuclear power and AI infrastructure would become closely linked would have seemed improbable five years ago.
Efficiency improvements are real but often insufficient to offset growth. Nvidia's H100 GPU is roughly three to five times more energy-efficient per computation than the previous generation. But this improvement is more than offset by the fact that users are buying far more GPUs and running far larger models. The Jevons paradox in AI infrastructure: as efficiency improves, demand grows faster than efficiency, and total consumption increases.
Water Usage: Cooling the Machine
Data centers do not just consume electricity. They consume enormous quantities of water. Cooling is the primary driver. As processors consume electricity, they produce heat. That heat must be removed to keep the chips from failing. Evaporative cooling, which uses water to dissipate heat into the atmosphere, is the most common approach in large data centers because it is more efficient and less expensive than air cooling alone.
Microsoft disclosed in its 2023 Environmental Sustainability Report that its global water consumption increased by 34% between 2021 and 2022, to nearly 6.4 million cubic meters. It attributed much of this increase to AI. Google's water consumption increased 20% over the same period, to 5.6 billion gallons globally. These figures represent only what the companies have chosen to disclose, and disclosure standards vary widely.
In practice, the water consumption of individual data center facilities can be dramatic. A single large hyperscale data center can consume millions of gallons of water per day. In dry climates, this places direct pressure on local water supplies. In regions experiencing drought, the water demands of AI infrastructure are drawing scrutiny from regulators and communities.
The ChatGPT example has become a widely cited illustration of the issue. Researchers from the University of California and the University of Texas estimated that a conversation of approximately 20 to 50 questions with ChatGPT requires roughly 500 milliliters of water to cool the servers during inference. At scale, across hundreds of millions of daily conversations, this adds up to millions of gallons per day, just for a single application.
Data center cooling technology is evolving. Liquid cooling, which circulates water or other coolants directly through processor chips, is more efficient than evaporative cooling and uses significantly less water. Immersion cooling, in which servers are submerged in non-conductive fluid, is even more efficient. These technologies are being adopted in new data center builds, but the installed base of older facilities continues to rely heavily on evaporative cooling.
The Carbon Footprint of Training a Model
The carbon cost of training large AI models is substantial and depends heavily on the source of the electricity used. Training in a region powered primarily by coal produces many times more carbon than training in a region powered by hydropower or nuclear energy. This creates an unusual dynamic in which the carbon footprint of an AI model is determined less by its architecture than by where its data center is located.
Researchers at the University of Massachusetts Amherst published a landmark 2019 paper estimating that training a large language model could emit as much carbon as the lifetime emissions of five average American cars. This figure was widely circulated and debated. It referred to training runs that were large for their time. Today's training runs are many times larger, though efficiency improvements have partially offset the scale increase.
Google, Microsoft, and Amazon have all made net-zero carbon commitments. In practice, these commitments rest heavily on the purchase of renewable energy certificates (RECs) and carbon offsets, both of which are controversial instruments. A renewable energy certificate allows a company to claim credit for renewable energy generated somewhere on the grid, even if the energy powering its data center comes from fossil fuels. Critics argue that this accounting obscures the actual carbon impact of AI infrastructure.
The time-matching problem is particularly significant. Data centers operate 24 hours a day. Renewable energy from solar and wind is intermittent, generated primarily during peak solar and wind periods. Matching renewable energy consumption to actual renewable generation on an hour-by-hour basis is much more difficult and expensive than purchasing annual RECs. Google has committed to 24/7 carbon-free energy matching by 2030. Microsoft has made a similar commitment. Progress toward these goals has been uneven.
The carbon from AI is not evenly distributed in time. Training runs produce concentrated bursts of emissions over weeks or months. Inference emissions are more diffuse but continuous. The growth trajectory of inference emissions is steeper than training emissions because the number of AI users grows faster than the frequency of training runs. A comprehensive carbon accounting for AI must capture both.
Land Use and Community Displacement
Data centers are large physical structures that require significant land. A hyperscale campus can cover hundreds of acres. In areas where tech companies have clustered their infrastructure, the effect on local land use has been dramatic. Northern Virginia has seen agricultural land converted to data center campuses at an accelerating pace. Loudoun County alone has more data center square footage than most cities have office space.
The tax implications are complex. Data centers create significant taxable property and often negotiate favorable tax rates with local governments, which can be competing to attract investment. But they create relatively few jobs for their size. A million-square-foot data center might employ a few hundred workers. The ratio of economic footprint to employment is very different from manufacturing or office-based industries.
In some communities, data center development has generated conflict. In Ireland, where a large fraction of European data center capacity is concentrated, the Irish grid operator issued a moratorium on new data center connections in 2021 because the demand was threatening grid stability. In regions with tight water supplies, including parts of the US Southwest, data center water rights have become contentious.
Indigenous land rights have occasionally come into conflict with data center development. In several cases in the southwestern United States, data center projects have been proposed on or near lands with historical or cultural significance to Native American communities. The permitting and consultation processes for these projects have been contested.
The concentration of AI infrastructure in certain geographies creates economic dependencies. Communities that host data centers become reliant on the continued operation and investment of a small number of large technology companies. When companies scale back data center plans, as has happened during periods of slower AI investment, the effect on local tax revenues and employment can be significant.
What the Industry Claims vs What the Data Shows
Tech companies have invested heavily in sustainability messaging. Google's annual sustainability reports describe ambitious commitments to carbon neutrality, renewable energy, and water stewardship. Microsoft publishes detailed environmental metrics and has pledged to be carbon negative by 2030. Amazon has launched the Climate Pledge, a commitment to net-zero carbon by 2040. These communications create a picture of an industry that takes its environmental responsibilities seriously.
The reality is more complicated. Despite purchasing vast quantities of renewable energy certificates, all three companies have seen their absolute carbon emissions increase as AI infrastructure has grown. Google's total greenhouse gas emissions increased 48% between 2019 and 2023, driven largely by data center energy consumption. The company acknowledged that its 2030 net-zero goal is now 'extremely ambitious' given the pace of AI growth.
Water consumption reporting is inconsistent and often incomplete. The figures companies publish represent water consumed at their directly operated data centers. But a significant portion of data center operations are outsourced to co-location providers and cloud sub-vendors. The water consumption of these facilities is typically not included in corporate sustainability reports. The actual water footprint of AI may be significantly larger than disclosed figures suggest.
The efficiency metrics that companies emphasize, such as power usage effectiveness (PUE), measure how efficiently a data center uses electricity internally. A PUE of 1.1 means that for every 1.0 units of power used by computing equipment, 0.1 additional units are used for cooling and other overhead. This metric does not capture the carbon intensity of the electricity source, the absolute quantity of power consumed, or the water used for cooling. It is a useful engineering metric that has been misapplied as a proxy for environmental performance.
The industry is not monolithic. Some companies are making genuine progress. Google has entered into direct corporate power purchase agreements for renewable energy that provide better matching between generation and consumption than RECs. Microsoft is funding research into carbon capture and is working toward using 100% clean energy for all its data centers by 2025. These efforts deserve credit. But they need to be evaluated against the backdrop of rapidly growing absolute consumption, not just improving efficiency ratios.
Regulatory Pressure and Policy Responses
Governments are beginning to respond to the environmental impact of AI infrastructure. In the European Union, the Energy Efficiency Directive requires large data centers to report energy consumption, power usage effectiveness, and water usage. The EU taxonomy for sustainable finance may eventually affect data centers' ability to attract ESG-linked capital if they do not meet environmental standards.
In the United States, regulation has been more fragmented. Some states, including Virginia and Georgia, have enacted legislation requiring data centers to meet certain efficiency standards to qualify for the tax exemptions they have historically received. California has been considering requirements for data centers to use recycled water for cooling rather than potable water. These state-level actions reflect the reality that federal AI regulation has been slow to materialize.
The conflict between AI development speed and environmental impact is playing out in utility commissions across the United States. When tech companies seek to interconnect large new data centers to the grid, utility commissions must balance the economic benefits of the investment against the infrastructure costs and potential reliability implications. In several cases, commissions have required data center operators to contribute to grid upgrade costs that would otherwise be socialized across all ratepayers.
Water rights disputes are emerging in drought-affected regions. In Arizona, where several large data centers are located, the state water authority has imposed restrictions on new development that relies on groundwater in certain basins. Data center operators have begun investigating alternative cooling methods, including air cooling and dry cooling, that use significantly less water, partly in response to regulatory pressure.
International governance of AI's environmental impact is nascent. The G7 has begun including AI sustainability in its technology policy discussions. The OECD has published guidelines for responsible AI that include environmental considerations. But binding international agreements specifically addressing the environmental impact of AI infrastructure do not yet exist. This regulatory gap means that companies can, within limits, choose to site infrastructure in jurisdictions with the weakest environmental requirements.
What Needs to Change
The most important change needed is transparency. Companies should disclose the full environmental footprint of their AI operations, including the indirect water and carbon consumption of outsourced operations. The methodology for calculating and reporting AI's environmental impact should be standardized. Without reliable, comparable data, it is impossible to hold companies accountable or make informed policy decisions.
Time-matched renewable energy procurement needs to become the standard, not the exception. Annual renewable energy certificate purchases do not ensure that computing happens on clean energy. Hourly or sub-hourly matching between renewable generation and data center consumption is technically achievable and some companies are moving in this direction. Policy incentives could accelerate this transition.
Research into more efficient AI architectures should be prioritized and valued. The AI research community has historically rewarded capability above all else. A model that achieves a new state of the art benchmarkBenchmarkA standardized test or set of tasks used to evaluate and compare the capabilities of different AI models on a common scale.Learn more → by using ten times as much compute as its predecessor is celebrated. A model that achieves the same capability with one-tenth the compute is less visible. Changing these incentives, through prizes, publication norms, and funding priorities, could shift the research agenda toward efficiency.
Water-efficient cooling technology needs to be deployed at scale. Liquid cooling and immersion cooling are significantly more water-efficient than evaporative cooling and are increasingly cost-competitive. Governments can accelerate adoption through efficiency standards, tax incentives, and procurement requirements for publicly subsidized facilities. Data center operators can accelerate adoption by making water efficiency a procurement criterion for new builds.
Finally, the AI industry needs to engage seriously with the communities that host its infrastructure. Data centers affect local power grids, water supplies, land use, and sometimes air quality. The communities affected are often rural, economically disadvantaged, and lacking the political resources to negotiate effectively with large technology companies. Genuine community engagement, not just public relations, is both an ethical obligation and, increasingly, a prerequisite for obtaining the permits and approvals that new facilities require.