Most arguments about AI and energy run on one viral number and a lot of feeling. These are the figures I keep within reach. Every number has a source, and company-reported figures are marked as such.
The big picture: how much power
Data centers use about 1.5% of the world’s electricity today. The International Energy Agency expects that to more than double by 2030, to roughly what all of Japan uses now. In the US the share is already above 4%.
| Scope | Then | Projected |
|---|---|---|
| World data centers | 415 TWh in 2024 (about 1.5% of global electricity)1 | About 945 TWh by 20301 |
| US data centers | 176 TWh in 2023 (about 4.4% of US electricity)2 | 325 to 580 TWh by 2028 (6.7% to 12%)2 |
A few more numbers worth having:
- The US accounted for 45% of global data center electricity in 2024, followed by China (25%) and Europe (15%).1
- A typical AI-focused data center uses as much electricity as 100,000 households. The largest ones under construction will use 20 times that.1
- Data centers are on track for almost half of all US electricity demand growth through 2030.1
- Nearly half of US data center capacity sits in five regional clusters, which is why the impact is felt locally rather than globally.1
Small vs. large models: energy per question
A plain chatbot answer uses about as much energy as running a microwave for one second. Switch on a “think hard” mode and the same question can cost 30 to 100 times more, because the model writes pages of hidden reasoning before it answers. Choosing a model is an energy decision.
Every model in a family, small or large, is a full AI model. A smaller one has fewer parameters, so each word it writes takes less computation. A larger one does more math per word, so it is slower and uses more energy per answer. Reasoning or “extended thinking” modes multiply that again by generating many more words behind the scenes. The cost is compute, which means electricity, not data or bandwidth.
| Example | Energy per answer |
|---|---|
| Median Gemini text prompt, May 2025 (company-reported) | 0.24 Wh, about one second of a microwave or nine seconds of TV3 |
| Average ChatGPT query, June 2025 (company claim) | About 0.34 Wh4 |
| Reasoning models vs. the same models with reasoning off (40 open models, Dec 2025) | About 30x more on average; hundreds of times more in the worst cases5 |
| Large reasoning models on a single long prompt | Over 33 Wh, more than 70x a tiny model on the same task6 |
Two caveats. The Hugging Face and Salesforce result has been reported as both “30x” and “100x” depending on how it was averaged, so 30 to 100 times is the honest range.5 And no major provider publishes complete, independently verified per-query figures; Google and OpenAI have shared their own numbers, and Anthropic has not published any.
Water: the real numbers
One chatbot prompt uses about five drops of water. That sounds like nothing. But Google’s data centers consumed about 7 billion gallons of freshwater in 2024, and its thirstiest single site used a billion gallons on its own. The question isn’t the drop. It’s which town’s water it comes from.
| Fact | Number |
|---|---|
| Water per median Gemini text prompt (company-reported) | 0.26 mL, about five drops3 |
| Google freshwater consumption, 2024 | 7.2 billion gallons; about 4.5 billion replenished7 |
| Google’s thirstiest single site (Council Bluffs, Iowa, 2024) | About 1 billion gallons8 |
| All US data centers, direct cooling water, 2023 | About 17 billion gallons, possibly doubling or quadrupling by 20289 |
| All US data centers, indirect water through their electricity, 2023 | About 211 billion gallons, roughly 12x the cooling water8 |
What about the ocean?
- Google’s campus in Hamina, Finland is a converted 1950s paper mill cooled with seawater. It pulls cold Baltic water through the mill’s old intake tunnel, runs it through heat exchangers, and blends the warmed water with incoming cold water before returning it to the sea.1011
- Microsoft’s Project Natick was a data center actually in the ocean: a sealed capsule on the seafloor off Scotland’s Orkney Islands from 2018 to 2020. Its servers failed at about one-eighth the rate of a matching land setup. Microsoft confirmed in 2024 that the project is no longer active.12
Water and power trade against each other. Google argues evaporative cooling can be the better environmental choice where water isn’t scarce, because it uses less electricity than air cooling.7 It’s a real tradeoff, not a simple villain story.
Eagle Mountain, Utah: a water fight that was really a power story
When Meta built in Eagle Mountain, Utah, the first question was water, a fair one in the desert. But Meta’s newer buildings use closed-loop cooling that needs almost no water in normal operation. The bigger draw is electricity: the original deal brought 1,000 megawatts of new power capacity to the area.
- Size and investment: construction began in 2018 and the campus opened in 2021. With its latest expansion, Meta says the campus will exceed 5 million square feet and $3 billion invested in Utah.13
- Power: the 2018 deal included a new substation bringing 1,000 MW of power delivery capacity, with Meta buying 100% renewable energy through Rocky Mountain Power tariffs.14
- Water: Meta says the new buildings use closed-loop liquid cooling with dry cooling and need no water during normal operation, and projects annual campus water use below a typical regional golf course (company claims).13
- Jobs: about 1,200 construction workers at peak, and about 300 permanent jobs after the expansion. The original 2018 estimate was 30 to 50 full-time staff.1314
- Tax breaks: a local and state package that could be worth up to $750 million over several decades.15
- Transparency: Utah’s HB 76 requires large data centers to report water use, but only for facilities operating after July 2026, so older buildings aren’t required to disclose.16
As one Utah water researcher put it, data centers use water two ways, for cooling and to generate their electricity, and the electricity side is usually the larger one.16
The data center in your town: who pays
When people fight a data center, they usually talk about water or noise. The bigger issue often shows up on the power bill. On the grid that serves 13 Eastern states, the price of guaranteeing enough power rose more than tenfold in two years, and the grid’s own market monitor attributed most of the increase to data centers. Every customer pays that, whether they use AI or not.
PJM operates the grid for 13 states plus Washington, DC, including Ohio, Virginia, Maryland, and Pennsylvania.
- Capacity prices: from $28.92 per megawatt-day for 2024–25 to $329.17 for 2026–27, the federally approved cap.17
- The cause: PJM’s independent market monitor attributed 63% of the 2025–26 price increase to data center demand, about $9.3 billion recovered from customers.18
- The household impact: in Columbus, Ohio, a major data center hub, the average residential bill was set to rise about $27 a month after the latest auction.19
- The fix works: in Virginia, transmission upgrades raised import limits by 43% and capacity prices fell from $444 to $329 per megawatt-day.19
- States are responding: Oregon’s POWER Act (June 2025) created the first dedicated data center rate class so large users pay for the grid upgrades they trigger. Virginia’s SB 253 would shift costs from households to data centers using 25 MW or more.17
- Grid strain: the IEA estimates about 20% of planned data center projects worldwide could be delayed by grid constraints.1
The fair counterpoint: capacity charges are one slice of a bill, impacts vary widely by utility, and high prices are drawing new power plants onto the grid.
The efficiency trap
Google cut the energy of a typical AI prompt 33-fold in a single year. Yet data center electricity keeps growing about 12% a year, and Google’s own emissions are up nearly half since 2019. When something gets cheaper, we use a lot more of it.
- Google says energy per median Gemini prompt fell 33x from May 2024 to May 2025.3
- Global data center electricity has grown about 12% a year since 2017, more than four times faster than electricity overall.20
- Google’s 2024 environmental report showed carbon emissions up 48% since 2019, mostly from data center energy and supply chain.21
- Even the IEA’s most optimistic “High Efficiency” case cuts 2035 data center demand only about 20% below its base case. Efficiency slows the curve; it doesn’t bend it down.1
The economist William Stanley Jevons described this in The Coal Question (1865): more efficient steam engines increased England’s total coal use rather than reducing it.
Common myths
| Myth | What the numbers say |
|---|---|
| Every AI prompt uses a bottle of water | Google reports about five drops per median prompt; most of the water footprint sits at the power plant38 |
| A small AI model is “just a program” | Small and large models are both full models; the difference is how much math each does per word |
| AI’s cost is mostly data and bandwidth | The cost is compute, which is electricity |
| Underwater data centers are the future | Microsoft’s ocean experiment worked and was still shelved; seawater cooling on shore is the real-world version1012 |
| Water is the main local issue | Often the bigger local effect is on the power bill1719 |
| Efficiency gains will solve it | Energy per prompt fell 33x in a year while total demand kept growing about 12% a year320 |