~0.3 Wh (2009)
Google's original per-search energy estimate — still the most commonly cited search baseline, and never publicly updated since
Some of your colleagues are excited to bring AI into their classrooms. Others have real concerns about its energy and water use. This briefing tries to separate four questions that usually get collapsed into one: how much energy a single task uses, how large AI's aggregate footprint is, what burden falls on the specific communities hosting that infrastructure, and whether a given classroom use is worth its cost. The numbers below are best current estimates, not settled facts — and several of them are corrections to an earlier draft of this same page.
Last reviewed October 1, 2026 — first against an independent accuracy review that corrected a number of overstated or mislabeled figures in the first draft, then again against a U.S. GAO technology assessment that added the hardware-manufacturing and national-scale figures in "What's actually worth watching." See Sources for specifics.
A handful of separately run 2025 estimates for a short, ordinary chatbot text response — a quick question, not a long document or a multi-step agent task — land in a similar order of magnitude. That's useful context. It is not a verified industry consensus: each number describes a different model, on a different date, measured a different way, and none of it covers heavier workloads like reasoning or agentic tasks (more on that below).
Google's disclosed figure — 0.24 Wh for the median Gemini Apps text prompt in May 2025 — is the only one of these based on the company's own internal measurement rather than outside modeling. Google's own write-up is explicit that the data hasn't been independently verified and doesn't describe every prompt or future performance. Epoch AI's roughly 0.3 Wh is a modeled estimate for GPT-4o under stated assumptions, not a direct measurement. The UN University's 2026 report uses 0.42 Wh as a representative planning figure for its own calculations. Treat these as three separate estimates with a name, a date, and a method attached — not one settled number.
That first row is worth sitting with. The familiar claim that an AI query burns 10× the energy of a web search traces to an identifiable 2023 estimate (about 2.9 Wh, modeled from a SemiAnalysis assumption about ChatGPT's early infrastructure) compared against Google's 0.3 Wh search figure — which dates to 2009. Comparing a 2023 estimate for one system against a 16-year-old figure for a completely different, long-since-rebuilt one doesn't reliably establish a multiplier in either direction. Nobody has published a rigorous, current, apples-to-apples comparison of a search and an AI query. That's worth noticing the next time someone states one with confidence — including an earlier draft of this page, which claimed the opposite overstatement.
Pick the perspective closer to your own. These aren't two settled positions where one side has already won — they're claims worth checking against what the underlying research actually supports.
Everything below is converted to grams of CO₂e so the order of magnitude is comparable at a glance. Bars are positioned on a real logarithmic scale, calculated directly from each value (the axis below the chart shows the scale). That comparability has limits — the table underneath explains what each figure actually measures, since they don't all draw the same boundary around "impact." Two entries are shown as ranges rather than single bars, because the underlying research reports a range, not one typical number.
| Entry | What it actually measures |
|---|---|
| AI text / image | Electricity used, multiplied by an assumed grid carbon intensity — no allocated training, device, or hardware-manufacturing footprint |
| AI video | Same method as above; the huge range reflects real differences across models and settings, not just uncertainty |
| Netflix | A 2019 estimate that does include devices, home networks, and data centers together |
| Driving | EPA's tailpipe CO₂ only — no manufacturing or fuel-extraction footprint |
| Flying | Includes an added multiplier for aviation's extra warming effect at altitude, which the driving and AI figures don't have an equivalent for |
| Beef | Full agricultural supply-chain footprint (feed, land use, methane) — a genuinely different kind of accounting than "electricity × grid factor" |
None of the AI rows above include the emissions from manufacturing the hardware they ran on. That's not a rounding error — see "What's actually worth watching" below for how large the gap might be.
The pattern that survives the caveats: an ordinary search, chatbot reply, or AI image sits down near hundredths to a few grams of CO₂e; even a full hour of streaming or a heavier AI video sits in the tens of grams. Driving, flying, and eating meat start roughly where that cluster ends and run one to two orders of magnitude higher. None of these are measured on quite the same terms — see the table above — but the gap between "ordinary digital task" and "physical, biological thing" is wide enough to survive that imprecision.
Be careful what that gap does and doesn't settle. It's a statement about per-task magnitude, nothing more. It says nothing about whether a specific data center is straining a specific community's water supply, whether a specific grid can absorb the new demand without building more fossil generation, or whether a specific classroom's use of the tool was worth it — the local-burden and educational-value questions from the top of this page. Scale doesn't answer those, and a chart that makes AI look small can be used to wave them away. It also doesn't run in reverse: a real, large footprint across a company's entire operation doesn't mean any single ordinary text prompt was wasteful. Both can be true at once, which is exactly why this page keeps the four questions separate instead of averaging them into one verdict.
Now put a number on your own use, and set it against two things people already have an intuitive feel for. This models ordinary short text prompts only — a long document, a reasoning-heavy request, or an AI agent taking several actions on your behalf can use well more than the figure below per query.
Notice how many prompts it takes to equal even one beef patty. That's useful for the task-energy question — and only that question. It says nothing about heavier AI tasks, the aggregate picture below, where the infrastructure behind it sits, or whether any of it was worth doing. Treat it as scale, not as a verdict.
Zoom out from your own habit to the aggregate, schoolwide picture. This is a single illustrative scenario built from fixed per-task assumptions, not a validated model of any real school's mixed use — a building that leans toward reasoning-heavy or agentic tools would use meaningfully more than this suggests, task for task.
None of this argues for using AI tools less thoughtfully — it argues for using them more deliberately. Roughly in order of how much difference each one makes:
The biggest lever isn't a setting, it's the decision before you open the tool. A quick fact lookup, a calculation, or a question you could answer from a textbook doesn't need a language model — and skipping it removes that task's AI energy use entirely, not just reduces it.
This is the highest-leverage setting, because the gap between task types spans one to three orders of magnitude, not a few percent. Save image generation for when a real image is the point of the assignment, and treat video generation — which can cost as much as an hour of streaming per clip, sometimes far more — as an occasional tool rather than a default output format.
Most chatbot platforms offer a lighter or faster option (sometimes labeled "mini," "flash," "instant," or similar) alongside their flagship model. These generally use less energy per query, and for routine tasks — drafting an email, summarizing a short passage — they're often good enough. No independent, cross-provider ranking of exactly how much lighter models save currently exists, so treat "generally less" as the honest claim rather than a specific multiplier.
Every follow-up message is a separate query, and long conversations mean some of the earlier exchange gets reprocessed each time too, though exactly how much depends on the platform's caching, which isn't public information. A little upfront thought about what you actually want usually beats sending five rough attempts — though not by a fixed or guaranteed amount.
Auto-summarize and always-on writing suggestions can usually be disabled in settings, which does stop that specific computation. Search AI overviews are a partial exception: Google's own help documentation says AI Overviews can't be fully disabled as a core Search feature, and its "Web" filter changes what's displayed without necessarily stopping the underlying computation. Check the specific product's documentation rather than assuming a toggle exists.
None of these require becoming an expert, and none require giving up AI tools. They're the same category of habit as turning off lights in an empty room — not because any one instance matters much, but because the habit is what scales.
Roughly ordered from the smallest lever to the biggest one. If you only have energy to think about one or two, start from the bottom.
The gap between a short text answer and a generated video isn't 2× or 5× — using this page's own figures, it runs from roughly 200× to well over a thousand times. A classroom norm of "text first, image or video only when it adds real value" does more than any amount of prompt-counting.
Efficiency gains lower the cost of each query, which tends to encourage more use — that's the logic behind the rebound effect, or Jevons Paradox. It doesn't automatically follow that any specific improvement will increase total use, but it does mean "the new model is more efficient" is not, on its own, a reliable reason to expect AI's total footprint to shrink.
The same electricity has different water and land costs depending on the grid it comes from. A data center on a hydro-heavy grid can pair a tiny carbon footprint with a large water footprint; a coal-heavy grid often raises the carbon side without necessarily lowering the water side, since thermal power plants use substantial cooling water too. There's no location that's simply "clean" on every dimension at once.
OpenAI and Anthropic have never published a per-query energy or carbon figure. Google and Mistral both have, for specific workloads — Google for a median Gemini text prompt, Mistral for one 400-token response. That's real progress, but the two use different accounting methods, cover different tasks, and can't yet be meaningfully compared to each other. Incomplete and inconsistent disclosure, not total silence, is the more accurate complaint — and it's still enough to make independent verification hard.
As of the UN University's 2025 data, only 32 countries host AI-specialized data centers at all, and about 90% of that capacity sits in the US and China. The environmental cost and the economic benefit of AI infrastructure don't land on the same communities. For a global classroom conversation, this is often the more consequential inequity than the energy math itself.
Everything above is about a single task. Zoomed out: U.S. data centers were about 4% of U.S. electricity demand in 2022, and the International Energy Agency projected 6% by 2026. The U.S. Government Accountability Office's 2025 technology assessment cites separate estimates putting data centers at 6.7–12% of U.S. electricity by 2028, with AI applications specifically responsible for something like 10–20% of data center electricity today — a share one financial-industry estimate expects to reach 20% by 2030. GAO is also explicit that "it is unclear what portion of data center electricity consumption is related to generative AI" specifically, as opposed to other computing — the same measurement gap this page keeps running into, confirmed by a government audit. These are national, aggregate numbers: a single classroom's use is a vanishingly small fraction of them, and the aggregate number answers a policy question, not whether any one prompt was worth sending.
Nearly every figure on this page, including its own comparison chart, counts only the electricity used to train or run a model. GAO's 2025 report cites an analysis finding that embodied carbon from manufacturing the computing hardware — mining raw materials, fabricating chips, building and shipping servers — adds roughly 50% on top of the emissions from training and using AI. Hardware also has a finite life: GPUs are generally considered obsolete after about 4 years, and what happens to them afterward is a lifecycle stage with almost no public data behind it. If that 50% figure is in the right range, every operational number on this page is undercounting the true footprint by closer to half again as much.
Four different questions tend to get flattened into one debate: how much energy a specific task uses, how large a provider's aggregate footprint is, what burden lands on host communities, and whether a given use is worth it educationally. These questions try to keep them separate.
Tags describe the type of evidence behind each source — a company's own measurement, an independent benchmark, a modeled estimate, or commentary — not a verdict on whether it's right. Read each note for the specific limitation to watch for; institutional affiliation, a DOI, or agreement with this page's conclusions don't by themselves establish that a number is correct.
You asked, so here's an attempt — and it runs straight into the same wall the rest of this page describes. There is no way to actually measure this.
Anthropic hasn't published a per-query energy figure for Claude, so any number here is Claude estimating its own footprint using other companies' disclosed figures for different models — precisely the kind of comparison this page has spent nine sections asking you to be skeptical of. For what it's worth: by the time of this revision, the conversation behind this page had run to roughly 80 separate exchanges with the model, counting tool calls, web fetches, document reads, and file edits across several rounds of revision. That's not a count of model inferences — a single exchange can involve more than one internally, and whether long-context turns are cheaper than a fresh query depends on caching behavior that also isn't public. Multiplying that rough count by Google's disclosed 0.24 Wh gives about 19 Wh; multiplying it by the UN University's own upper reference point for long, complex queries (10+ Wh) gives roughly 800 Wh.
A 40-fold range doesn't pin much down, and that's the actual point: this page is not exempt from the transparency problem it spends its other nine sections describing. Treat this section as illustrative arithmetic, not an estimate.