The server that processed your last AI query consumed more electricity than your refrigerator will use in a month. Multiply that by billions of daily queries, and you begin to approach the scale of what AI energy consumption actually means in 2026. Global data center electricity demand reached approximately 460 to 490 terawatt hours in 2025 — growing 17% in a single year, at a pace more than five times faster than global electricity demand overall. If the world’s data centers were a country, they would rank fifth in energy consumption globally, sitting between Japan and Russia.
That number is already extraordinary. What makes it alarming is the trajectory. The International Energy Agency projects that data center electricity consumption will double to approximately 945 terawatt hours by 2030, roughly equivalent to Japan’s entire current electricity consumption. AI workloads are the primary driver of that growth, and the hyperscale technology companies building the infrastructure to support them committed more than $400 billion in capital expenditure in 2025 alone, with a further 75% increase projected for 2026.
There is no AI without energy. That statement is increasingly shaping energy policy, geopolitical competition, and climate commitments in ways that have barely begun to register in public debate.
The Scale That Is Hard to Visualize
Raw terawatt-hour figures are difficult to grasp without context. The comparisons that make AI energy consumption legible are striking.
A typical AI-focused data center consumes as much electricity as 100,000 households. The largest ones currently under construction will consume 20 times as much, the equivalent of powering two million homes from a single facility. Training a single large language model at the frontier of capability is estimated to require 1,287 megawatt hours for a system comparable to GPT-3, a figure that has grown substantially for more recent and more capable models. Running that model in production, at the scale of hundreds of millions of daily queries, generates an ongoing electricity demand that compounds continuously with user growth.
The geographic concentration of this demand makes its local impact even more acute than the global numbers suggest. The United States hosts approximately 45% of global AI data center capacity. Ireland, which positioned itself as a European technology hub over the past two decades, now directs approximately 21% of its national electricity to data centers, a figure the IEA estimated could reach 32% by 2026. In Dublin, data centers already account for 79% of the city’s electricity consumption. In Northern Virginia, the world’s largest data center cluster, facilities consume 26% of the state’s total electricity supply. In Frankfurt, the figure approaches 42%.
These are not marginal additions to existing electricity systems. They are transformative demands being placed on grids that were not designed to absorb them at this pace. Grid operators in Virginia, Ireland, Singapore, and parts of Texas are reporting connection queues measured in years, with new data center projects waiting for grid capacity that does not yet exist and cannot be built fast enough to match the pace of demand.
The Brookings Institution’s analysis of global AI energy demands within the regulatory landscape provides comprehensive data on the geographic concentration of data center electricity demand and the regulatory frameworks, or absence of them, that govern how that demand is being met.
Who Is Paying the Energy Bill
The capital expenditure figures behind AI energy consumption reveal a concentration of investment that has few precedents in the history of private infrastructure development.
Five large technology companies, Microsoft, Google, Amazon, Meta, and Oracle, collectively spent more than $355 billion on AI-relevant data center infrastructure in 2025. Combined 2026 capex is projected to exceed $620 billion across the same group. These figures represent the largest single-cycle infrastructure investment outside government in modern economic history. For context, the entire Marshall Plan that rebuilt Western Europe after World War II cost approximately $173 billion in today’s dollars. The technology industry is deploying more than three times that amount in a single year, on data center infrastructure alone.
The electricity to power that infrastructure has to come from somewhere. In the United States and China, which together account for 70% of global data center electricity consumption, most of that electricity is still generated from fossil fuels. Natural gas provides approximately 40% of US data center electricity. Coal provides 30% of data center electricity globally, with China accounting for the largest share. Renewables supply approximately 27% of global data center electricity, and that share is growing — but growing more slowly than the absolute demand, meaning that the net effect of AI expansion on carbon emissions is currently positive even as the carbon intensity per unit of compute declines.
The tech companies’ response to this tension has been to become the world’s largest corporate buyers of renewable energy. In 2025, technology companies accounted for approximately 40% of all corporate power purchase agreements for renewables signed globally. The pipeline of agreements between data center operators and small modular nuclear reactor projects has grown from 25 gigawatts at the end of 2024 to 45 gigawatts today, a development that may accelerate the commercialization of new nuclear technology in ways that extend well beyond the AI industry’s own energy needs.
Renewable energy certificate prices rose 35% in 2024, driven largely by this surge in technology company procurement. The scale of tech industry energy demand is changing the economics of clean energy markets in ways that affect the cost and availability of renewable electricity for every other consumer in those markets.
The Water Problem Nobody Mentions
The energy dimension of AI’s environmental footprint is the one that receives the most coverage. The water dimension is less discussed and in some respects more immediately concerning for the communities where data centers are located.
A typical 100-megawatt AI data center uses between 1.5 and 3 million cubic meters of water per year for evaporative cooling. US data centers consumed approximately 17 billion gallons of water in 2023, with 84% of that consumed by hyperscale and colocation facilities. Direct water consumption in hyperscale data centers is projected to reach between 16 billion and 33 billion gallons annually by 2028, a range that reflects genuine uncertainty about the pace of data center expansion and efficiency improvement.
Water consumption matters differently from energy consumption because water is a local resource with no global grid equivalent. A data center in Phoenix that consumes millions of gallons of water annually is drawing from an aquifer system already under severe stress from population growth, agriculture, and climate change. The communities adjacent to these facilities do not benefit from the global productivity gains that AI enables, but they bear the local cost of reduced water availability for other uses.
Liquid cooling technologies, immersion cooling, and direct-to-chip systems reduce direct water use by 70 to 90% compared with evaporative cooling, and their adoption is accelerating as AI accelerator power densities exceed what conventional air cooling can handle. The transition is driven by engineering necessity rather than environmental policy, but its effect on water consumption per unit of compute will be significant if it reaches the mainstream of the industry before the absolute scale of demand outpaces the efficiency gains.
The Countries Racing to Attract — and Manage — AI Infrastructure
The energy demands of AI infrastructure are reshaping geopolitical competition in ways that connect directly to national energy policy, industrial strategy, and the broader competition for technological leadership.
IEA Executive Director Fatih Birol articulated this clearly: “There is no AI without energy, and countries that provide secure, affordable, and rapid access to electricity will be one step ahead.” That statement describes an emerging form of energy-based comparative advantage. Countries with abundant, reliable, and competitively priced electricity, whether from hydropower, nuclear, natural gas, or renewables, have something that AI companies need and are willing to pay substantially to access.
The United States dominates current AI infrastructure deployment but faces grid capacity constraints that are slowing new builds in its most electricity-constrained regions. The Federal Energy Regulatory Commission’s queue for new grid connections contains projects totaling more than 2,600 gigawatts of pending capacity, more than double the entire installed US generation capacity. Most of those projects are waiting years for interconnection approvals. Data center developers, unwilling to wait, are increasingly building dedicated on-site power generation, primarily natural gas peaker plants, that bypass the grid entirely. This approach solves the individual developer’s problem while adding unregulated fossil fuel capacity to the energy system and undermining the renewable energy commitments that the same companies are making publicly.
This dynamic, in which national energy infrastructure decisions are being made at the level of individual corporate projects because regulatory frameworks cannot keep pace with investment demand, mirrors the broader pattern of AI governance that every major government is currently struggling to address. The energy system was not designed for the kind of demand surge that AI infrastructure represents, and neither the regulatory frameworks nor the physical infrastructure is scaling fast enough to accommodate it cleanly.
The Efficiency Argument and Its Limits
The technology industry’s primary response to criticism of AI energy consumption is an efficiency argument: power consumption per AI task is declining rapidly, at a rate the IEA describes as unprecedented in energy history. More capable models are being trained and run on less electricity per unit of output than their predecessors. Algorithmic improvements, hardware advances, and software optimization are all contributing to dramatic improvements in the energy cost of individual computations.
This argument is correct as far as it goes, and it matters. The history of computing is largely a history of radical efficiency improvement that made previously unaffordable technology available to billions of people. Personal computers became possible because computing costs fell by many orders of magnitude from the mainframe era. Smartphones became possible because the computing power of a 1990s supercomputer was compressed into a device that consumes a few watts.
The efficiency argument has a fundamental limit, however, that is already visible in the 2026 data. More people are using AI. Energy-intensive uses AI agents running continuous workflows, multimodal models processing video and audio at scale, real-time inference at billions of queries per day — are growing faster than efficiency improvements are reducing per-task energy consumption. The net result is that absolute energy consumption continues to rise even as energy consumption per task falls. This is the rebound effect that economists have documented across virtually every technology that became dramatically more efficient: efficiency gains expand access and use, and the expansion of use exceeds the reduction in per-unit consumption.
The Carbon Brief’s analysis of five critical charts on data center energy use and emissions documents this dynamic in detail, showing that while efficiency improvements are real and significant, they are not sufficient to offset the scale of demand growth projected through 2030.
The Climate Commitment Contradiction
The companies consuming the most electricity to power AI are also among the most public about their net-zero commitments. The gap between those commitments and the current trajectory of emissions is becoming increasingly difficult to explain.
Google reported a 48% increase in carbon emissions in 2024 compared with 2019, attributing much of the increase to data center expansion driven by AI. Microsoft’s emissions grew 29% between 2020 and 2024, despite its commitment to becoming carbon negative by 2030. Meta’s emissions rose 67% between 2019 and 2023. In each case, the companies maintain that their renewable energy purchases offset their operational emissions on a contractual basis. What that accounting conceals is that purchasing a renewable energy certificate does not mean that renewable electrons were the ones powering the data center, particularly at times of peak demand, or in regions where renewable capacity is insufficient to meet total demand, and the marginal electron comes from a gas or coal plant.
The contradiction is not simply a public relations problem. It is a structural challenge for the energy system. The technology sector’s demand for electricity is growing faster than clean energy can be built, and in the near term, the gap is being filled by fossil fuels. The IEA projects that natural gas and coal together will meet more than 40% of the additional electricity demand from data centers between now and 2030.
The comprehensive 2026 data on AI data center energy consumption by region and provider shows that Scope 2 disclosed emissions are still rising across hyperscalers as growth outpaces clean energy procurement, even as the carbon intensity per unit of compute falls, the same rebound dynamic that characterizes the efficiency argument’s limits.
The same competitive logic that is driving record levels of spending on strategic infrastructure across geopolitical rivals is also driving the AI infrastructure investment surge with comparable indifference to environmental externalities when national or corporate competitive advantage is at stake. And the race to secure strategic energy resources is being reshaped by AI’s electricity appetite in ways that are beginning to influence territorial and diplomatic calculations.
Looking Ahead
The IEA’s projection that data center electricity consumption will double to 945 terawatt hours by 2030 is the base case — not the optimistic scenario. In a higher-demand scenario, consumption could reach 1,300 terawatt hours by 2035, an amount that exceeds the current total electricity consumption of every country in the world except China, the United States, India, and the European Union combined.
Managing AI energy consumption at that scale will require changes that go beyond corporate sustainability commitments and voluntary efficiency improvements. It will require energy policy frameworks that account for large-scale data center demand in grid planning. It will require transparency and disclosure standards that make it possible to actually verify the emissions claims that technology companies are making. It will require investment in transmission infrastructure that can move clean energy from where it is generated to where data centers are located.
The IEA notes that AI itself may contribute to solving the energy challenges its growth creates — by optimizing grid operations, improving demand forecasting, accelerating materials discovery for better batteries and solar cells, and making the energy system more efficient overall. That possibility is real. It is also not guaranteed, and it does not arrive automatically as a byproduct of building more data centers.
The energy cost of the AI revolution is not hidden in any technical sense. The data is publicly available, the trends are well documented, and the implications are visible to anyone who examines them. What has been hidden is the extent to which this cost has been externalized onto electricity grids, water systems, and carbon budgets that were never designed to absorb it. That externalization is becoming harder to ignore, and the policy, infrastructure, and regulatory responses to it will shape the pace and character of AI development through the end of the decade.
