Does a ChatGPT query really use ten times as much electricity as a Google search? No. That number has been floating around the web for years, and it's wrong.
At the same time, the AI industry's appetite for energy is real, and it's growing faster than almost anything else on the grid. Both things are true, which is exactly why the data is worth a close look. For this article, I worked through the IEA reports, the US analyses from LBNL and EPRI, the hyperscalers' sustainability reports, and Germany's Bitkom figures, and had every key number cross-checked.
- Data centers consumed around 485 TWh of electricity worldwide in 2025 (about 1.5% of global consumption). By 2030, the IEA expects a doubling to roughly 950 TWh, about Japan's current electricity consumption
- A single AI query is harmless: ~0.3 Wh for ChatGPT, 0.24 Wh for Gemini, roughly the same as a Google search. The total volume drives consumption, not the individual question
- Hyperscaler emissions keep rising despite climate pledges (Microsoft +25%, Google +18%, Amazon +16% in the latest reporting year). In response, nuclear power is making a comeback, from Three Mile Island to mini reactors
1. The Status Quo: What Data Centers Consume Today
The most reliable source for the big picture is the International Energy Agency (IEA). Its key figures:
The decisive point sits in the third row:
Data centers overall are already growing at 17% per year, more than four times faster than the rest of electricity demand. AI-focused data centers are growing several times faster still, at 50%. That makes AI by far the most important new driver of global electricity demand.
2. The Forecast: Doubling by 2030
By 2030, the IEA's base case expects a doubling. The 2025 "Energy and AI" report projects 945 TWh, and the April 2026 update around 950 TWh, each roughly Japan's total electricity consumption today. Where that growth happens becomes clear when you build up from 2024 to the forecast:
The US and China together account for around 80% of total growth. Europe plays a supporting role at roughly +45 TWh. The AI infrastructure buildout is primarily a two-country race.
Today's shares show the same concentration:
45% of the world's data center electricity is consumed in the US today, 25% in China, and 15% in Europe.
3. USA: On Track for 12% of Electricity Consumption
Nowhere is AI's power appetite as concretely measurable as in the US. Lawrence Berkeley National Laboratory (LBNL) runs the numbers regularly on behalf of the Department of Energy:
In 2024, US data centers consumed 192 TWh, or 4.7% of national electricity consumption (2023: 176 TWh, or 4.4%). For 2030, the reference case of the current LBNL update sits at 649 TWh, or 11.8%, with a range of 521 to 843 TWh. Put differently, almost one in nine kilowatt-hours in the US could flow into a data center by 2030.
And that's already showing up in electricity prices today:
In PJM, the largest US grid region (13 states, including Virginia with the world's biggest data center hub), the capacity price for 2025/26 exploded from $28.92 to $269.92 per megawatt-day, an 833% increase. Data centers were responsible for 63% of that jump, equivalent to $9.3 billion. For households in Washington, D.C., bills rose by about $21 per month, roughly half of which the local consumer advocate attributes to the capacity price jump.
4. Energy per Query: The 0.3 Watt-Hour Truth
Now for the most-quoted and most-misquoted number of the entire debate. How much electricity does a single AI query cost?
A typical ChatGPT query costs around 0.3 Wh according to Epoch AI; OpenAI itself says 0.34 Wh. A median Gemini prompt sits at 0.24 Wh according to Google. For context, that's about as much as a classic Google search based on the old 2009 estimate, and roughly what an oven draws in a little over a second.
The truly remarkable part is the efficiency curve:
Google reduced the energy footprint of a median Gemini prompt by a factor of 33 within twelve months, and the carbon footprint by a factor of 44. The individual query is getting cheaper at breakneck speed. That total consumption explodes anyway is purely a volume effect, a pattern economists know as the Jevons paradox. Efficiency gains lower the price, and the lower price grows demand.
5. Training: What a Single Model Costs
Beyond operation (inference), training the models consumes energy too, concentrated into a few months:
The curve points steeply upward. GPT-4 (2023) caused an estimated 5,200 tonnes of CO2; Grok 4 (2025) between 73,000 and roughly 150,000 tonnes depending on the estimate; even the lower estimate equals about 17,000 cars in a year. The Stanford AI Index itself notes, however, that the Grok estimate in particular rests on uncertain assumptions. According to Epoch AI, the compute used to train frontier models doubles roughly every six months, around 4-5x per year.
6. The Hyperscalers: Emissions Rising Despite Climate Goals
Google, Microsoft, and Amazon have all set themselves ambitious climate targets. The AI reality looks different:
The details from the current reports:
Google's electricity demand rose 37% in 2025, the largest jump in company history. Its data centers consumed around 43 million MWh, roughly as much as all of New Zealand. Microsoft's 2030 goal of removing more carbon than it emits is now widely considered barely achievable by analysts.
7. The Nuclear Comeback
Big Tech's answer to its power problem has a name nobody would have predicted three years ago. Nuclear energy. The key deals:
The most symbolic deal remains Three Mile Island. Of all places, the site that stood for the worst nuclear accident in US history in 1979 is set to deliver power again from 2028, exclusively to Microsoft, under a 20-year contract. The unit being revived is not the damaged Unit 2 reactor but its undamaged neighbor Unit 1, which ran normally until 2019. Together with Google's mini-reactor order and Amazon's 5 GW targets, the AI industry has become the most important new source of demand for nuclear power in the US.
8. Water: The Overlooked Resource
Data centers don't just need electricity, they need water for cooling. The same pattern applies: tiny per query, huge in total.
A Gemini prompt consumes 0.26 milliliters of water according to Google, and a ChatGPT query about 0.000085 gallons (around 0.32 milliliters, roughly one-fifteenth of a teaspoon) according to OpenAI. At company scale, the math looks different:
Google consumed around 6.1 billion gallons of water in its data centers in 2024 (2021: 4.3 billion), Amazon 2.5 billion gallons (2025). Microsoft shows it can be done differently. The water usage effectiveness of its new data centers dropped from 2.3 to 0.27 liters per kWh, and 90% of the 2025 fleet uses cooling that evaporates little or no water. The company says it has been water positive since fiscal year 2025, replenishing more water than it consumes.
9. Germany: Growth With the Handbrake On
And in Germany? The Bitkom/Borderstep study provides the most reliable figures for the German market:
German data centers consumed an estimated 21.3 billion kWh in 2025, nearly double the 2015 figure. Installed capacity grew 9% to 2,980 MW and is projected to exceed 5,000 MW by 2030.
The most interesting number is AI capacity, though:
Germany's AI-ready data center capacity is projected to nearly quadruple from 530 MW today to 2,020 MW by 2030. At that point, 40% of all German data center capacity would be AI infrastructure. For scale against US dimensions: the planned US buildout through 2030 alone amounts to a multiple of Germany's entire installed base.
10. Efficiency: The Counterforce
The growth numbers tell only half the story. In parallel, the energy required per computation is falling dramatically.
The three most important pieces of evidence:
First, Google's already-mentioned 33x improvement per Gemini prompt in just twelve months. Second, the chip level: Nvidia states that a GB200 NVL72 system delivers up to 25 times more compute per watt than air-cooled H100 systems of the previous generation (measured on a specific inference benchmark that favors the new system), and the new Rubin generation is supposed to increase inference throughput per megawatt another tenfold (manufacturer claims, not independently verified). Third, the data center level, where modern facilities keep closing in on the physical limits of cooling efficiency.
So why does total consumption keep rising? That's exactly the Jevons paradox from section 4. Every efficiency gain makes AI applications cheaper and unlocks new use cases that more than consume the electricity saved. Demand grows faster than efficiency.
11. Forecasts Compared: Who Says What
To wrap up the numbers, it's worth seeing how far apart the credible forecasts are, and why:
The ranges look wide but tell a consistent story. All credible institutes expect at least a doubling by 2030, globally and in the US; the EPRI scenarios reach even further. The differences lie in scope (world vs. US), base year, and assumptions about AI demand. Be especially wary of headlines that apply US percentages to the world or vice versa.
12. Conclusion: Not an Apocalypse, but an Infrastructure Problem
The data paints a more nuanced picture than either camp in the debate.
The reassurance: a single AI query, at around 0.3 Wh, is not an environmental problem, the old shock numbers have been debunked, and efficiency is improving at a breathtaking pace. Even in 2030, data centers will consume only around 3% of the world's electricity.
The warning: this growth is concentrated in a few regions and a few years. In the US, the data center share of electricity consumption could nearly triple by 2030, the price effects are already measurable today, and the hyperscalers' emissions are running away from their own climate targets. The problem isn't the AI query. It's the pace at which grids, power plants, and permits have to keep up with the boom.
For how big the industry behind all this has become, see my AI statistics. And for whose chips are consuming all that power, check my Nvidia statistics.
