AI services promise answers to life, the universe, everything—but one question continues to stump the companies behind them: how much pollution does it take to power a single query?
When we asked Sierra Club advisor Jeremy Fisher for an estimate, he let out a deep sigh: “Oh, jeez.” That about sums it up.
While tech leaders love to boast that their AI data centers use more electricity than small cities, they’re less eager to discuss where that energy comes from. Their professed sustainability goals often clash with the reality: fossil fuels still do most of the heavy lifting. And with renewable infrastructure lagging, it’s unclear whether clean power can keep pace with AI’s explosive growth.
For all the talk of progress, the future of AI is looking a lot like the past—propped up by coal and gas. The same tech leaders who brag about clean energy and net-zero targets are quietly fueling their AI ambitions with gas turbines and coal plants. The truth is, the smarter our machines get, the dumber our energy choices are starting to look.
“AI” Contains Multitudes
Defining this problem is complex because the two letters “AI” contain an alphabet’s worth of meanings and energy-usage patterns. The AI chatbots that seem increasingly inescapable also appear to be, I think, the least troublesome aspect of AI in terms of power usage.
Consider the Electric Power Research Institute’s May 2024 paper that estimated one query to OpenAI’s ChatGPT chatbot used 2.9 watt-hours of electricity, versus .3 watt-hours for a non-AI Google query. Ten times more sounds bad—but it equates to keeping a 9-watt LED bulb on for 18 minutes.
In February, the research firm Epoch AI estimated the energy consumption based on OpenAI’s new GPT-4 model and found just 0.3 watt-hours for a simple query, rising, analyst Josh You wrote, to “2.5 to 40 watt-hours for queries with very long inputs.”
In June, OpenAI CEO Sam Altman wrote in an essay that the average ChatGPT query “uses about 0.34 watt-hours” of electricity and “one fifteenth of a teaspoon” of water to cool the AI processors involved. OpenAI’s PR department declined to expand on that post.
Training the large language model (LLM) behind a chatbot carries its own upfront cost in compute and power, which you put in that post as 20 to 25 megawatts for models “comparable to GPT-4o.”
But AI outputs more complex than text-only chatbot banter require more resources.
“Generating sound and especially video are far more energy intensive than simple text – mainly because of the complexity of sound or video files,” wrote Gartner VP analyst Bob Johnson.
So-called agentic AI, in which an LLM performs complex assigned tasks over time, further escalates power consumption.
“Agentic workflows can use 10–1000× more tokens per request, so total compute—and therefore energy—unquestionably rises,” wrote Andrew Feldman, CEO of Cerebras, a startup developing high-efficiency AI processors.
At Scale
The power used by a quick query of ChatGPT, Google’s Gemini, or Microsoft’s Copilot may equate to a snowflake, but combining those and other AI services with existing data-center growth adds up to an avalanche.
In December, the Department of Energy’s Lawrence Berkeley National Laboratory published a study that found U.S. data-center energy consumption, fueled by AI adoption, more than doubled from 76 terawatt-hours a year in 2018, 1.9% of all electricity used in the U.S., to 176 TWh in 2023, 4.4% of the total. (By comparison, the EIA estimated that U.S. Bitcoin mining added up to 70 TWh in 2023.)
The study offered two estimates for 2028 to reflect uncertainty about demand for and deployment of AI. In its low-end scenario, data center use rises to 325 TWh, accounting for 6.7% of total U.S. consumption; in the high-growth forecast, it more than triples to reach 580 TWh, representing 12% of the total.
The International Energy Agency adopted similar conclusions in its Energy and AI report, published in April. That sees worldwide data-center use climbing from 2024’s 415 TWh, 1.5% of global electricity consumption, to about 945 TWh in 2030, almost 3% of the world’s total.
“AI is the most important driver of this growth, alongside growing demand for other digital services,” the report observed.
The IEA, like Lawrence Berkeley, computed additional forecasts for varying assessments of AI adoption and efficiency. In its “Lift-Off” case, power consumption reaches 1,700 TWh by 2035, accounting for 4.4% of the global total; in the “High Efficiency” scenario, these numbers are 970 TWh and 2.6%; the “Headwinds” case has them at 700 TWh and below 2%.
International Data Corporation (IDC) broke out AI-specific energy usage in a forecast published in September 2024, predicting 146.2 TWh by 2027, which reflects a compound annual growth rate of 44.7%. That, however, would keep it a minority of worldwide data center energy use, predicted at 857 TWh in 2028.
Some Foggy Forecasts
Some skepticism about the demand for AI is warranted, given the substantial investment tech giants are making in this technology while still developing business models for it.
As the research firm MoffettNathanson observed in a July assessment of Meta’s aggressive pursuit of AI: “we’re less convinced that its more ambitious bets will pay off.”
The Sierra Club’s Fisher complained that many sites are trying to create a market for AI: “It’s a push of a tool rather than the offering of a tool.”
More uncertainty arises from the efficiency of data centers, measured as a benchmark called PUE, or “power usage effectiveness,” which represents the ratio of power used by a facility to the power delivered to the computers inside. Lower scores are better.
Google, for example, reports that its PUE dropped from 1.23 in Q3 of 2008 to 1.08 as of Q1 2025. Meta, meanwhile, cited an improvement from 1.11 in 2019 to 1.08 in 2023, the latest data it has published. Amazon’s 2024 sustainability report featured a global average of 1.15, with its best U.S. site at 1.04; the company didn’t specify the location.
Microsoft didn’t publish a company-wide average; instead, it cited figures for individual U.S. states and other countries—a low of 1.12 in Wyoming and a high of 1.3 in Singapore.
Many of these reports also highlight the declining use of water to cool AI facilities, as measured by water usage effectiveness (liters withdrawn per kilowatt-hour of electricity used). Meta, for instance, cites a decline from .27 in 2019 to .18 in 2023.
And these publications regularly cite processor-level improvements, such as Google’s boast that its latest AI chip, Ironwood, delivers twice the performance per watt of its predecessor.
Feldman, the Cerebras CEO, predicted “clear opportunities to continue pushing the hardware envelope across transistor technology, architectural innovation, and enhanced scaling” that will “keep energy‑per‑token trending downward.”
However, increases in AI usage could leave those efficiency improvements swamped by the rising tide.
What Kind of Power?
The tech giants’ sustainability reports highlight net-zero goals and occasionally report achievements in reducing their carbon emissions. But “net” is the crucial word: They buy non-renewable power, then offset that by underwriting green power projects and purchasing renewable energy credits, sometimes also by supporting carbon dioxide-removal efforts.
Meta, for instance, says it’s kept its operations net-zero since 2020 “primarily” by offsetting electricity use with renewable-energy purchases.
“It’s questionable how much of that renewable energy they can claim is their own,” Fisher said.
The best-case scenario involves supporting clean-energy projects proximate to data centers. He cited a Google partnership in Nevada with Fervo Energy to build a next-gen geothermal electric plant that leverages drilling techniques developed for fracking gas, with Google buying 115 MW of its electricity.
Johnson, with Gartner, pointed to Microsoft’s deal with Constellation Energy to power some data centers by reopening a reactor at the shuttered Three Mile Island nuclear plant near Harrisburg, Penn.
“This one had already been decommissioned,” Johnson said, distinguishing this arrangement from clean-energy deals that draw from the existing grid. “Microsoft is truly adding more power.”
The worst-case situation is an AI firm paying for new fossil-fuel facilities to speed a data center’s delivery.
The most notorious example of that may be the “Colossus” facility that xAI had built in 122 days in Memphis, Tenn., to power X’s Grok chatbot. Notwithstanding Tesla CEO Elon Musk owning those firms, xAI fired up an array of gas turbines months before securing an air permit from local regulators.
XAI did not answer a request for comment, but Grok’s replies on X to questions about the data center acknowledge its pollution.
“It does highlight a tension: Tesla aims to accelerate sustainable energy, yet xAI’s Memphis data center relies on gas turbines emitting pollutants for now, due to urgent AI compute needs,” Grok posted in July.
Meta CEO Mark Zuckerberg seems equally set on speeding data-center buildout to reach AI “superintelligence,” outlining plans for “several multi-GW clusters” in a series of Threads posts on July 14 that didn’t mention power sources.
The research firm SemiAnalysis, however, had earlier reported that Meta’s 1.5 GW “Hyperion” complex in Louisiana would feature two on-site gas power plants, describing its rapid construction as “full Elon mode.” None of these companies provided even rough estimates of how much sustainable power was used in their data centers.
The Road Ahead
The power needs of these giant facilities can outstrip the capacity of even such massive renewable-power projects as Dominion Energy’s under-construction Coastal Virginia Offshore Wind (2.6 GW) and SB Energy’s Orion Solar Belt (900 MW) in Texas.
Data center operators seeking alternatives to keeping coal power plants online and building new gas plants are pinning their hopes on reviving nuclear power.
Amazon, Google, and Meta have announced plans for a new class of small modular reactors that would represent cheaper alternatives to the larger plants that U.S. utilities have failed to build on schedule or at budget in recent decades.
But the IEA report projects that no SMRs will come online until 2030. What about before then?
Charles Yang, executive director of the Center for Industrial Strategy, a Washington think tank, endorsed an “all-of-the-above approach” including “nuclear restarts, enhanced geothermal, and hydropower retrofits and powering non-powered dams.”
The “Build AI in America” report, published on July 21 by Anthropic, the developer of the Claude series of models, cautiously endorses solar, batteries, and geothermal as “the most economically efficient choices” for training LLMs until next-generation nuclear arrives.
However, the rest of the report suggests that only geothermal, nuclear, and gas, with their support for around-the-clock output, will enable the operation of AI services.
AI advocates suggest that applying these models to sustainable energy will yield significant innovations in itself.
“AI is being leveraged to accelerate breakthroughs in materials science, high energy physics, climate modeling, and other scientific domains that show enormous promise to help address problems related to climate change,” wrote Joshua New, director of policy at the nonprofit policy firm SeedAI.
But in the near term, if AI operators want more clean energy, they will have to pay for it themselves—especially now that the Trump administration is tearing down most of the Biden administration’s renewable-energy incentives. Fisher challenged them to do that, pointing to the extraordinary financial resources they can bring to meeting their net-zero goals: “Their ability to build clean energy is undiminished.”
So how much pollution did your AI query generate? The truth is, no one really knows—and the companies building these systems aren’t in a rush to find out. That silence says more than any emissions report. As AI reshapes the economy, the pressure is on to ensure it doesn’t also reshape the climate. Whether it’s data centers tapping gas turbines or PR teams tapping the brakes on transparency, one thing is clear: the intelligence may be artificial, but the energy problem is very real. And unless the industry starts owning its impact—not just offsetting it—today’s smart tools could leave us with a very dumb legacy.
