Every chatbot query on earth is 2% of AI's power

TL;DR. The IEA projects data centre electricity consumption rising from 415 TWh in 2024 to around 945 TWh by 2030, just under 3% of global electricity, and to roughly 1,200 TWh by 2035. In the United States, data centres will consume more electricity than aluminium, steel, cement, chemicals and every other energy-intensive good combined. Those are the headline figures and they are large. The arithmetic inside the same report is more interesting. Assume a generous 1 Wh per chatbot text query and ten billion queries a day, roughly Google-search volume and about four times what ChatGPT reports. That comes to 3.65 TWh a year, against 155 TWh consumed by AI-focused data centres in 2025. All the world's text queries are about 2% of AI data centre electricity, and the public conversation is almost entirely about that 2%.

---

Status: established. Primary source: IEA, Energy and AI, 2025, published under CC BY 4.0, with its updated projections in Key Questions on Energy and AI. The per-query measurement is from Google's own August 2025 technical paper. The 2% calculation follows the IEA report's own worked example.

---

The headline numbers

Data centre electricity consumption: around 415 TWh in 2024, about 1.5% of global electricity. Projected to reach around 945 TWh by 2030 in the IEA Base Case, just under 3% of the global total, and roughly 1,200 TWh by 2035.

For scale, the 2030 figure is slightly more than Japan's entire electricity consumption today.

The United States and China account for nearly 80% of global growth to 2030. In the US, data centres represent nearly half of all electricity demand growth over that period, and by the end of the decade the country will consume more electricity for data centres than for aluminium, steel, cement, chemicals and all other energy-intensive goods combined.

Locally the concentration is sharper still. Ireland's data centres already take around 21% of national electricity, with projections above 30%. Northern Virginia sits around 26%.

None of that is small, and this article is not an argument that it is.

The arithmetic almost nobody quotes

The IEA report contains a worked example. It is worth following because it reframes the entire public conversation.

Assume one chatbot text query consumes 1 Wh. That is generous: a large query, and a round number chosen for convenience.

Assume ten billion such queries a day. That is roughly the volume of Google searches, and about four times what ChatGPT reports.

Ten billion watt-hours a day is 10 GWh a day, or 3.65 TWh a year.

AI-focused data centres consumed 155 TWh in 2025.

So every chatbot text query in the world, at four times reported volume and a generous per-query estimate, accounts for roughly 2% of AI data centre electricity.

Where does the other 98% go?

Which is the important question

Training runs. Image and video generation, which the report does not quantify and which are far more compute-intensive per output than text. Recommendation and ranking systems. Search infrastructure. Enterprise inference at volumes nobody publishes. Idle capacity and redundancy. Cooling, which runs around 7% of load in efficient hyperscale facilities and over 30% in less efficient enterprise ones.

The honest answer is that the public breakdown does not exist. The IEA can project totals from utility data and construction pipelines. It cannot say what fraction is training versus inference, or text versus video, because the operators do not disclose it.

And that is the finding. The number people argue about, the cost of an individual query, is measurable, has been measured, and is close to irrelevant to the total. The 98% that matters is not broken down anywhere public.

What the per-query measurement actually says

Since the per-query figure dominates discussion, it is worth stating what has been measured rather than estimated.

Google published a technical paper in August 2025 reporting that a median Gemini Apps text prompt used 0.24 Wh of electricity, 0.26 ml of water, and emitted 0.03 gCO2e, as of May 2025.

That is a measurement by an operator with access to its own infrastructure, and it sits at the bottom of the range of public estimates, which run from roughly 0.3 to 3 Wh for chatbot-class queries.

The widely repeated claim that a chatbot query costs about ten times a web search sits at the top of that estimate range. It traces to estimation, not measurement, and the one published measurement is roughly a quarter of the low end of the estimates.

This is citation decay in a form worth naming: a number derived under assumptions circulates until it reads as measured, while an actual measurement published by a party with access gets less repetition because it is smaller and therefore less quotable.

Two cautions, both real. Google measured its own product, which is an interested measurement in the sense the sepsis case established. And a median prompt is not a heavy one: reasoning-heavy queries, long contexts and image generation all cost substantially more, and the median hides that spread.

The pattern underneath both figures

Energy per task keeps falling. Total energy keeps rising. Usage grows faster than efficiency.

The clearest illustration comes from a single company's own reporting: Google reduced data centre emissions by 12% in 2024 through clean energy procurement and operational improvements, while its absolute data centre electricity consumption grew 27% year on year.

Both numbers are true and they are not in tension. Efficiency improved and volume improved faster.

Which means per-query efficiency gains do not bound total consumption, and any argument that better chips will solve this has to explain why the historical pattern of flat consumption despite rising workloads reversed after 2020, with efficiency gains slowing since.

Three things this establishes

The public debate is anchored to the wrong quantity. Individual query cost is roughly 2% of AI data centre electricity under generous assumptions. Arguing about it is arguing about the rounding.

The breakdown that matters is undisclosed. Nobody outside the operators can say how much of the 98% is training, video generation, enterprise inference or idle capacity. Totals are projectable from utility data; composition is not, and no disclosure regime requires it.

And local concentration is the near-term issue, not the global share. Just under 3% of global electricity by 2030 is manageable in aggregate. Twenty-one percent of Ireland's is a grid problem now, and the aggregate figure conceals exactly the cases that bind.

What it does not establish

That AI energy use is not a problem. Doubling to a Japan-sized load in six years, concentrated in a handful of grids, is a serious planning challenge, and the IEA treats it as one.

That the projections are reliable. They are scenario-based, and the IEA is explicit that post-2030 figures are explorations rather than forecasts. Near-term numbers are firmer because much of the supply is already locked in by construction lead times.

That per-query figures are worthless. They are the right measure for a specific question, which is the marginal cost of one more query. They are the wrong measure for total system impact.

And nothing about water in detail. Per-query water figures exist, vary enormously with cooling architecture, and are less well measured than electricity. A typical 100 MW facility is estimated at 1.5 to 3 million cubic metres a year for evaporative cooling, and that estimate is wide for a reason.

What is unresolved

The composition of the 98%. This is the single most valuable disclosure that does not exist.

Whether efficiency ever outpaces usage. It did before 2020 and has not since, and no mechanism has been proposed that would restore the earlier pattern.

How much of projected demand is contracted rather than speculative. Announced capacity and built capacity differ substantially, and the IEA notes bottlenecks reducing the likelihood of the most aggressive scenarios.

And what the local limits actually are. Ireland and Northern Virginia are past the point where new connections are routine, and no public analysis states where the ceiling sits.

The counter-argument

The 2% calculation assumes text queries are the only consumer-facing use, and they are not. Image and video generation are consumer products at scale and cost far more per output. Folding them in would raise the consumer share substantially, and the IEA's own worked example excludes them because it lacks the numbers, not because they are negligible.

Anchoring on the individual query is defensible. It is the only quantity a person controls, and telling someone their query is 2% of the problem can read as telling them not to bother. Individual action is a small lever and it is the lever they have.

Google's measurement is not independent. It is the best available per-query figure and it was produced by the company selling the product, on its own definition of a median prompt, with no external validation. Treating it as settling the question repeats the error the sepsis case documented.

And the aggregate share may understate the problem in a different direction. Three percent of global electricity is small; the marginal generation built to serve it is disproportionately gas in some regions and coal in others, so the emissions share can exceed the electricity share.

The short version

Data centres consumed around 415 TWh in 2024 and are projected to reach roughly 945 TWh by 2030, just under 3% of global electricity, and about 1,200 TWh by 2035. In the United States they will soon draw more power than aluminium, steel, cement, chemicals and all other energy-intensive goods combined, and Ireland already runs around 21% of national electricity through them.

And the IEA's own worked example reframes the whole discussion. At a generous 1 Wh per query and ten billion queries a day, four times ChatGPT's reported volume, all the world's chatbot text queries come to 3.65 TWh a year against 155 TWh consumed by AI-focused data centres in 2025.

That is about 2%.

The other 98% is training, video generation, enterprise inference, ranking systems, cooling and idle capacity, and no public breakdown exists. Totals can be projected from utility data and construction pipelines. Composition cannot, because nobody discloses it.

Meanwhile the measured per-query figure is smaller than the estimated one. Google's own August 2025 paper reports a median Gemini text prompt at 0.24 Wh, 0.26 ml of water and 0.03 gCO2e, against public estimates of 0.3 to 3 Wh and a widely repeated claim of ten times a web search. The larger number is an estimate that circulated into fact; the smaller one is a measurement by an interested party. Both cautions apply.

And efficiency does not bound totals. Google cut data centre emissions 12% in 2024 while data centre electricity consumption grew 27%. Per-task energy keeps falling and total energy keeps rising, because usage grows faster than efficiency, and that pattern reversed direction after 2020 rather than holding.

Common questions

How much electricity do data centres actually use? Around 415 TWh in 2024, roughly 1.5% of global electricity, projected by the IEA to reach about 945 TWh by 2030, just under 3% of the global total, and roughly 1,200 TWh by 2035. The 2030 figure is slightly more than Japan's entire current electricity consumption. The United States and China account for nearly 80% of the growth.

Is my individual chatbot use a meaningful part of that? Not really, on the IEA's own arithmetic. Assuming a generous 1 Wh per query and ten billion queries a day, roughly Google-search volume and about four times what ChatGPT reports, all the world's chatbot text queries come to 3.65 TWh a year against 155 TWh consumed by AI-focused data centres in 2025. That is about 2%, and the public conversation is almost entirely about it.

Where does the other 98% go? Training runs, image and video generation, recommendation and ranking systems, search infrastructure, enterprise inference, idle capacity and redundancy, and cooling, which runs about 7% of load in efficient hyperscale facilities and over 30% in less efficient enterprise ones. The honest answer is that no public breakdown exists. The IEA can project totals from utility data and construction pipelines but cannot say what fraction is training versus inference, because operators do not disclose it.

What does a single query actually cost? Google's August 2025 technical paper reports a median Gemini Apps text prompt at 0.24 Wh of electricity, 0.26 ml of water and 0.03 gCO2e, as of May 2025. Public estimates for chatbot-class queries range from about 0.3 to 3 Wh. The widely repeated claim that a query costs ten times a web search sits at the top of the estimate range and traces to estimation rather than measurement.

Should I trust Google's figure? With two cautions. It is a measurement by an operator with access to its own infrastructure, which makes it the best available number and also an interested one, in the same sense that a manufacturer-run validation of a medical device is. And a median prompt is not a heavy one: reasoning-heavy queries, long contexts and image generation cost substantially more, and a median conceals that spread.

Do efficiency improvements solve this? Not on the evidence so far. Google reduced data centre emissions by 12% in 2024 through clean energy procurement and operational improvements while its absolute data centre electricity consumption grew 27% year on year. Both are true: efficiency improved and volume improved faster. The historical pattern of flat consumption despite rising workloads reversed after 2020, with efficiency gains slowing since.

Is 3% of global electricity a lot? In aggregate it is manageable and the more pressing issue is local concentration. Ireland's data centres already take around 21% of national electricity with projections above 30%, and Northern Virginia sits around 26%. Those are grid problems now, and the global percentage conceals exactly the cases that bind.

How reliable are the projections? Near-term figures are firmer than they might appear because much of the supply is locked in by construction lead times, which is why the IEA's four scenarios differ by only around 100 TWh at 2030. Post-2030 the agency is explicit that its figures are explorations rather than forecasts, and it notes bottlenecks across the value chain reducing the likelihood of the most aggressive near-term scenarios despite booming investment.