AI energy use is becoming a measurable problem, not a moral panic

The better news about AI energy is not that the problem has disappeared. It is that the numbers are getting good enough to stop arguing in slogans.

Our World in Data published a careful July 20 analysis of data centers and artificial intelligence energy use, drawing on the International Energy Agency, Epoch AI, Google and other public estimates. The result is a useful middle ground. A single simple chatbot question is usually not an ecological disaster. Large clusters of AI data centers can still put real pressure on local grids, water systems, electricity prices and clean-power plans. Both statements can be true.

That matters because the public conversation keeps jumping between two unhelpful extremes. One side treats every prompt as a climate sin. The other points to small per-query figures and acts as if infrastructure growth is a rounding error. The useful question is more practical: where are data centers being built, what electricity powers them, who pays for grid upgrades, what should providers disclose and which AI uses are worth the energy they require?

The headline numbers are smaller than the panic, but not small

Our World in Data reports that data centers consumed about 485 terawatt-hours of electricity in 2025, roughly 1.5% of global electricity generation, based on IEA data. Of that, AI-focused data centers accounted for about 155 TWh, or around 0.5% of global electricity. Non-AI data centers still used more in aggregate, which is easy to forget because AI dominates the current discussion.

The base-case IEA projection for 2030 is very different from a dismissal. It puts total data center electricity use at about 945 TWh, roughly 3% of global electricity. Most of the growth is expected to come from AI-focused facilities. In that scenario, AI-focused and non-AI data centers approach similar electricity demand by the end of the decade.

Those numbers need context. Global electricity generation is also growing, and clean generation is expanding quickly in many regions. A terawatt-hour figure alone does not tell us whether the additional demand is met by coal, gas, solar, wind, hydro, nuclear or storage-backed clean power. It also does not tell us whether a local grid has capacity at the moment a new facility wants to connect.

The good news is that the debate can now be framed around measurable infrastructure rather than vibes. A company can be asked where its data center runs, what the marginal grid mix looks like, how much power it uses, whether it shifts loads to cleaner hours, whether it pays for new capacity and how it reports emissions. Those are answerable questions.

The per-query number is not the whole story

Per-prompt estimates are useful because they correct some wild claims. Google says the median Gemini Apps text prompt uses about 0.24 watt-hours of energy, emits about 0.03 grams of CO2 equivalent and consumes about 0.26 milliliters of water, using its full-stack methodology. Sam Altman has written that an average ChatGPT query uses around 0.34 Wh. Epoch AI independently estimated roughly 0.3 Wh for a typical ChatGPT query. These figures are not identical, and each comes with assumptions, but they sit in the same general range for simple text use.

That is tiny compared with ordinary daily electricity use in high-income countries. It is much closer to seconds of household appliance use than to running a washing machine or charging an electric car. For ordinary users, this should lower the temperature. Asking a chatbot a short question is not the main climate decision in your day.

But averages hide the direction of travel. Our World in Data notes that longer inputs and agentic reasoning tasks can use far more. Epoch AI estimated about 2.5 Wh for a 7,500-word input and around 40 Wh for a 75,000-word input. The IEA gives estimates around 1.1 Wh for a standard agent request and about 50 Wh for an agentic request with reasoning. These are still not huge in household terms, but they are no longer the same as a short text prompt.

This is why the right unit depends on the decision. If the question is whether a teacher should feel guilty for using AI to draft a worksheet, the per-query number matters. If the question is whether a region should approve several hyperscale facilities, per-query math is almost irrelevant. The region needs to know peak load, grid upgrades, water demand, land use, backup generation, transmission constraints and the carbon intensity of extra electricity.

Local concentration is the real pressure point

The global share can sound manageable while the local share is very large. Our World in Data points out that data centers use about 5% of electricity in the United States, with AI-focused ones probably around 2%. In Ireland, data centers account for more than 20% of electricity consumption. Within countries, the burden can be even more concentrated in particular states, counties or grid zones.

That is where the politics start. A data center can be a national fraction and a local headache at the same time. It may require new transmission lines, substations, cooling systems, water access, backup generators and tariff decisions. If the costs are socialized while the benefits are private, residents will reasonably object. If the operator pays for capacity, uses clean power and participates in demand response, the same project may look different.

This is also where good policy beats moral performance. Telling individuals to stop asking short AI questions will not solve a grid bottleneck in Ireland or Virginia. Requiring transparent energy reporting, clean-power procurement, local impact studies and fair cost allocation might.

The recent Hacker News discussion of the OWID article, about 54 points and 52 comments by publication time, showed the same tension. Some readers focused on how surprisingly small a query can be. Others pointed out that 5% of U.S. electricity for data centers is already meaningful. Several comments returned to water, local effects and comparisons with other energy uses such as streaming or air conditioning. That is a healthier argument than vague guilt because it asks what should be measured and where the burden lands.

Carbon depends on the grid, not only on the server

Electricity use and emissions are related but not identical. A data center powered by a coal-heavy grid has a different climate impact from one running on low-carbon electricity. A facility that draws power during fossil-heavy peaks differs from one that can shift flexible workloads to cleaner hours. A company that buys renewable certificates without adding real local capacity is not solving the same problem as one that funds additional clean generation and transmission.

This is the part procurement teams should care about. It is not enough to ask a cloud or AI provider whether it has a climate pledge. Ask for region-level reporting, energy and carbon accounting, how inference loads are scheduled, whether water use is reported, whether clean power is additional and what happens during peak demand. For AI-heavy workloads, ask whether the provider can disclose model, region and task-level estimates rather than a single global average.

The best providers will not only make chips more efficient. They will make the physical footprint legible. Google's Gemini inference post is useful in that sense because it publishes a methodology and numbers, including energy, carbon and water estimates for median text prompts. The exact figures will be debated and may not generalize to every model or workload, but the act of publishing methodology is progress. It gives others something to interrogate.

What individuals should take from this

For personal AI use, the practical lesson is calm. Do not turn every prompt into a climate confession. The marginal electricity for a simple query is usually small. If AI helps you avoid a car trip, repair something, learn faster, write better code or reduce wasted work, the energy tradeoff may be positive. If you are generating thousands of low-value images, running long agent loops for trivia or leaving automated tasks to churn without purpose, the balance looks worse.

Use the tool intentionally. Prefer shorter prompts when they are enough. Avoid needless retries. Do not run agentic workflows just because they are available. But do not confuse symbolic self-denial with infrastructure policy. The meaningful levers are mostly upstream: cleaner electricity, efficient hardware, better scheduling, transparent reporting and siting decisions that do not dump costs on local communities.

Schools, nonprofits and small businesses can use the same frame. Instead of banning AI on environmental grounds, ask when it is useful enough to justify the computation. A short summarization or accessibility task is not the same as large-scale synthetic media generation. A one-off research query is not the same as an always-on agent refreshing pages for hours.

What companies and governments should do

Companies adopting AI should include energy questions in procurement. Which regions serve the workload? What is the estimated electricity use for the task type? Is the provider publishing carbon and water metrics? Can workloads be moved to lower-carbon regions or times? Are enterprise features encouraging efficient use, or do they make it easy to launch wasteful background agents?

Governments and grid operators need a different checklist. Where are clusters forming? Who pays for substations and transmission? How much new clean generation is firm rather than promised? What happens to local rates? What water sources are used for cooling? Can facilities curtail or shift load during stress events? Are communities receiving real benefits or only abstract promises about innovation?

These questions are not anti-technology. They are what serious infrastructure planning looks like. Data centers are factories for computation. They can be socially valuable, but they are still physical loads on grids and communities. Treating them as physical infrastructure is a sign of maturity, not hostility.

Why this qualifies as good tech news

Good news in technology is not always a breakthrough device. Sometimes it is a better measurement frame. Our World in Data's analysis does not let AI companies off the hook. It also does not validate panic about every individual prompt. It separates the levels: small marginal use, large aggregate demand, local grid concentration and the carbon intensity of electricity.

That separation helps everyone make better decisions. Users can stop treating a short prompt as the main climate issue. Companies can stop hiding behind global averages. Regulators can focus on local capacity, disclosure and cost allocation. Researchers can compare workloads and push for better public data.

AI infrastructure will keep growing. The useful achievement here is that the argument is becoming less mystical. We know where to look: electricity, location, timing, water, reporting and clean power. That is not a complete solution, but it is a better map. And a better map is how real improvements usually start.