The Environmental Impact of AI

tl;dr

• The per-prompt gap has narrowed sharply: The popular claim that AI uses ten times a web search rests on older estimates, and Google's own 2025 data now puts a text prompt roughly level with a search
• Text prompts are minor and video is the exception: A chatbot query ranks below a single email and far below an hour of streaming, while generative video is the real energy outlier
• The impact lives in aggregate, not the prompt: Data centers used about 1.5% of global electricity in 2024 and are on track to roughly double by 2030, with AI one driver among several
• Water and grid strain are local problems: Per-query water estimates vary by a factor of a thousand depending on what they count, and the sharpest pressure falls on specific regions
• Construction is racing ahead of usage: US data center building has nearly tripled since late 2022, funded by record corporate spending against demand that is projected, not yet proven

You ask a chatbot a question, and somewhere a data center draws a measure of power and water to answer it. How much, and how that compares to everything else running on the internet, has become one of the most misreported questions in technology.

The Number Everyone Cites Is Out of Date

For two years, one statistic anchored nearly every conversation about AI and the environment. A single chatbot prompt, the claim went, uses roughly ten times the electricity of a Google search. The figure came from a 2024-era estimate that put a ChatGPT request at about 2.9 watt-hours against roughly 0.3 watt-hours for a conventional search.^[1] One peer-reviewed analysis went further, estimating that adding AI answers to search could raise energy use 60 to 70 times, though it relied on large 2023-era models like GPT-3.^[2]

Then the primary data arrived. In August 2025, Google published the first detailed per-prompt measurement from a major AI company. The median text prompt to its Gemini apps used 0.24 watt-hours of electricity, about the energy of running a microwave for one second, and emitted 0.03 grams of carbon dioxide.^[3]

Two things make that number striking. It sits roughly level with the traditional search it was supposed to dwarf. And Google reported that the same prompt used 33 times less energy than it had a year earlier, the result of model and software efficiency gains.^[3] The often-repeated ten-times claim described a moment that efficiency had largely erased.

The honest reading is not that the early estimates were dishonest. They measured different models in different years against a contested baseline. The 0.3 watt-hour search figure itself traces back to 2009 and is almost certainly too high today.^[2] Any single ratio ages fast, and the ones still in circulation are old.

A Prompt Against a Stream, an Email, and a Video

Comparing AI to a search settles only part of the question. The more useful comparison is against the other things people do online all day.

A 2025 analysis by TRG Datacenters, reported by Forbes, ranked common digital activities by energy and carbon. A Google search or an AI chatbot prompt landed at the bottom, at about 0.0003 kilowatt-hours, or 0.105 grams of carbon.^[4] A short email came in higher, at 4.7 grams. An hour of high-definition Netflix or YouTube topped the list, at about 0.12 kilowatt-hours, or 42 grams.^[4]

Put plainly, an hour of streaming video produces roughly 500 times the carbon of sending two text prompts to a chatbot.^[4] The everyday text query that dominates AI anxiety is among the smallest-footprint things a person does on the internet.

There is a real AI exception, and it is not text. Generative video is a heavy lift. The same analysis put a video clip of 6 to 10 seconds at about 0.05 kilowatt-hours, comparable to an hour-long Zoom call and far above any text or image task.^[4] As video generation scales, that, not chatbot text, is the activity worth watching.

Where the Impact Is Real: Scale, Water, and the Local Grid

None of this means AI's footprint is trivial. It means the footprint lives in aggregate scale, not the individual prompt.

The International Energy Agency estimates that data centers consumed about 415 terawatt-hours of electricity in 2024, roughly 1.5% of the global total, growing about 12% a year since 2017.^[5] It projects that figure will roughly double to around 945 terawatt-hours by 2030, slightly more than Japan's entire electricity use today, with AI the most important driver.^[5] In the United States, the Department of Energy's national lab found data centers used about 4.4% of national electricity in 2023, potentially reaching 6.7% to 12% by 2028.^[6]

Two caveats keep this honest. AI is a major driver but not the whole of data center demand; the IEA attributed only about 15% of total data center energy to AI servers in 2024, on the order of 60 terawatt-hours.^[5] For scale, a single other technology draws more than double that: the Bitcoin network alone consumed an estimated 138 terawatt-hours in 2025, about 0.5% of global electricity.^[9] A cryptocurrency built on deliberate computational effort outweighed the measured energy of all AI servers combined.

Training is the other commonly overlooked figure, and it cuts both ways. Training a large model is genuinely resource-intensive; the landmark water study estimated that training GPT-3 in Microsoft's US data centers evaporated about 700,000 liters of freshwater on-site.^[8] But that is a one-time cost, spread across the billions of queries the model then answers, so it adds only a sliver to any single prompt.

Water is the second live concern, and the per-query numbers are genuinely contested. Google reported about 0.26 milliliters, five drops, per Gemini prompt.^[3] The same University of California, Riverside study put a longer exchange closer to 500 milliliters, a full bottle, for every 10 to 50 responses.^[8] Neither is wrong. The gap is scope: on-site cooling water alone, or the far larger volume used upstream to generate the electricity, which often accounts for 80% or more of the total.^[7]

Both energy and water share a defining feature. The strain is local. AI data centers cluster geographically, and nearly half of US capacity sits in five regional groupings.^[5] A national average can look modest while a specific county's grid or aquifer comes under real pressure. Meeting the added demand is projected to fall mainly to renewables and natural gas, with a range of other sources, including nuclear, expected to contribute.^[5]

The Buildout Is Running Ahead of the Demand

The aggregate numbers describe what AI consumes today. The construction data describes what companies expect it to consume, and that story is sharper. In April 2026, US spending on data center construction crossed $50 billion at an annualized rate for the first time, according to Census Bureau figures. Data centers now account for 2.3% of all US construction spending, more than the country spends on public transportation structures like airports and mass transit, and they have become the largest segment of private office construction.^[10]

The trend line matters more than the milestone. Monthly spending on data center construction has nearly tripled since late 2022, when AI chatbots went mainstream, and now runs roughly 16 times the level of a decade ago.^[11] A category that barely registered in construction statistics has become one of the largest forces in American building.

The money behind it comes mostly from the "hyperscalers", the handful of companies that operate computing infrastructure at global scale: Amazon, Microsoft, Google, and Meta. Together they plan to spend up to roughly $630 billion on capital expenditures in 2026, a 62% jump from the record $388 billion they spent the year before, with Amazon alone committing about $200 billion.^[12] Analysts project the total will pass $1.1 trillion by 2027.^[13]

The key word in all of this is anticipated. These companies report that their constraint is supply, not demand: they are building capacity for usage they project, not usage they can measure.^[14] That cuts two ways. If the projections hold, the aggregate energy and water pressures described above arrive on schedule and at scale. If they do not, the industry will have overbuilt. Either way, the construction is happening now, and it is the clearest evidence that AI's environmental question is more of an infrastructure question, decided in permitting offices and utility planning meetings rather than in anyone's chat window.

Final Thoughts

As would be expected, the measured picture is less dramatic than either camp suggests. A single AI text prompt is among the smaller things a person does online, roughly on par with a web search and a fraction of a minute of streaming video. The real concern is not the prompt. It is the pace and geographic concentration of the data center buildout behind it, along with the water and grid demands that come with scale.

Reading any number in this debate well takes a few questions. Which model and which year: efficiency has moved fast enough to change conclusions within months, as Google's 33-fold drop shows.^[3] Per-prompt or aggregate: a tiny individual number and a large total can both be true at once, and quoting one to dismiss the other is the most common sleight of hand. Energy or water, and on-site use or the full supply chain: the narrower the scope, the smaller the figure looks.^[7] And who produced it: the most detailed per-prompt data comes from the companies selling the technology, while independent academic estimates run higher.^[2][8] The honest answer probably sits between the two. A statistic that skips these details is not necessarily wrong, just incomplete, and incomplete numbers are how both hype and dismissal get manufactured.

For anyone making decisions about AI, the practical point holds. The individual query is not where the responsibility sits. The infrastructure choices are: where a system runs, how it is cooled, and where its electricity comes from. Those are the numbers worth tracking, worth asking vendors about, and worth demanding that AI providers disclose.

References

1. Artificial intelligence and the environment: Putting the numbers into perspective – Jisc National Centre for AI
2. Estimating the Increase in Emissions caused by AI-augmented Search – Wim Vanderbauwhede, University of Glasgow (arXiv)
3. In a first, Google has released data on how much energy an AI prompt uses – MIT Technology Review 
4. New Data: AI Is Almost Green Compared To Netflix, Zoom, YouTube – Forbes (TRG Datacenters analysis) 
5. Energy and AI: Executive Summary – International Energy Agency 
6. Global energy demands within the AI regulatory landscape – Brookings Institution  
7. How Much Water Do AI Data Centers Really Consume? – IEEE Spectrum 
8. Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models – Communications of the ACM (Li, Yang, Islam & Ren, UC Riverside) 
9. Cambridge study: sustainable energy rising in Bitcoin mining – Cambridge Judge Business School (Cambridge Centre for Alternative Finance) 
10. Data Centers Become Largest Segment of US Office Construction – Data Center Knowledge (US Census Bureau data)
11. How much is the US spending on building data centers? – Our World in Data
12. AI spending boom accelerates as Big Tech pours trillions into infrastructure – Fortune (JPMorgan Global Research)
13. AI Capex 2026: The $690B Infrastructure Sprint – Futurum Group