Can we say exactly how much more electricity a ChatGPT query uses than a Google search? The published estimates do not support a clean, like-for-like comparison.
At the same time, data center electricity use is rising. The IEA estimates a 17% increase in 2025, and a 50% increase for AI-focused data centers. Those figures show a clear trend, but they do not prove that AI is driving total global electricity demand. The reference year, metric, and uncertainty matter.
- The IEA estimates global data center electricity use at 485 TWh in 2025. Its 2030 projection is about 950 TWh, or roughly 3% of global electricity use
- Published estimates for simple text prompts range from about 0.24 to 0.34 Wh. They use different methods and are not directly comparable
- Several technology companies report rising emissions or electricity demand. Their metrics cover different reporting years and boundaries
1. The Status Quo: What Data Centers Consume Today
The International Energy Agency (IEA) estimates data center electricity use at 415 TWh in 2024. Its April 2026 update puts use at 485 TWh in 2025, up 17% year over year:
If you want those annual totals as a running counter, I built a live electricity and water calculator using the IEA’s annual data center estimates. It interpolates them onto the current second and says plainly what is an estimate.
The 50% figure refers to electricity use by AI-focused data centers in 2025, not to all AI electricity use. Nor should data center growth be equated with growth in total electricity demand. In its 2025 report, the IEA estimated that data centers account for less than 10% of global electricity demand growth from 2024 to 2030.
The investment figure is large, but it is not a direct measure of electricity use. The IEA reports more than $400 billion in capital expenditure by five large technology companies in 2025 and expects a further 75% increase in 2026. These company investments do not measure electricity use and are not limited to data centers.
2. The Forecast: Doubling by 2030
The IEA's April 2026 update projects about 950 TWh in 2030, or roughly 3% of projected global electricity use. Its 2025 report projected 945 TWh. These are different report vintages. The earlier regional breakdown shows where additional demand was expected at that time:
The United States and China together accounted for nearly 80% of projected global data center growth through 2030 in the IEA's 2025 base case. These are forecasts from that report, not a revised regional breakdown in the 2026 update.
Today's shares show the same concentration:
In 2024, the IEA attributed 45% of global data center electricity use to the United States, 25% to China, and 15% to Europe. The remaining 15% is the arithmetic remainder across other regions, not a separately reported country group.
3. USA: Data Centers Could Reach 12% of Electricity Use
In LBNL's reference case, data centers are projected to account for 11.8% of US electricity use in 2030. This is a national share, so it is not directly comparable with the IEA's projection of about 3% of global electricity use. Lawrence Berkeley National Laboratory published its update for the US Department of Energy in 2026. This chart combines historical estimates from LBNL's 2024 report with the reference case in the update:
Sources: LBNL 2024 report, LBNL update, published 2026
For 2024, EPRI estimates US data center electricity use at 177 to 192 TWh. This is a range, not a single measured value. In the LBNL update, the 2030 reference case is 649 TWh, or 11.8% of US electricity use. Sensitivity scenarios range from 521 to 843 TWh. These are modelled scenarios, not a precise prediction. EPRI's 2026 report gives its own 2030 projections of 380 to 790 TWh.
And that's already showing up in electricity prices today:
The PJM capacity auction shows how local forecasts can affect a market. In July 2026, the auction price for the 2028/29 delivery year reached $325 per MW-day, the FERC-approved price cap. The auction procured 6,831 MW less than the reliability requirement. The independent market monitor estimated that data center load in the forecast increased capacity-market revenues by about $6.5 billion, or 65.5%, in the 2027/28 auction compared with a scenario without that forecast load. This is a model of capacity-market revenues, not an allocation of household electricity bills. PJM published the auction results on July 14, 2026; the monitor's calculation appears in its 2025 report.
4. Energy per Query: Three Estimates, Not One Measurement Series
Now for the most-quoted and most-misquoted number of the entire debate. How much electricity does a single AI query cost?
Epoch AI modelled about 0.3 Wh for a typical simple GPT-4o text query in February 2025, using assumptions including a 500-token output. Sam Altman quoted about 0.34 Wh and 0.000085 US gallons of water for an average ChatGPT query, without specifying a method or reference period. Google estimated 0.24 Wh, 0.26 ml of water, and 0.03 g CO₂e for a median Gemini Apps text prompt in May 2025. Google derived the CO₂e value from prompt energy and its average 2024 grid carbon intensity, and the water estimate from prompt energy and its average 2024 water-use effectiveness. The sources cited here publish no per-query CO₂e estimate from Epoch or Altman. Google's figures have not been independently verified.
The truly remarkable part is the efficiency curve:
Google compared its median Gemini Apps text prompt in May 2025 with the May 2024 median. By its own calculation, energy use per prompt fell by a factor of 33 and its CO₂e footprint per prompt by a factor of 44. These figures have not been independently verified and apply to this specific prompt estimate. For total demand, the IEA says more than prompt-level efficiency matters: usage volume, new task types, and infrastructure expansion also affect electricity use.
5. Training: What a Single Model Costs
Beyond operation (inference), training the models consumes energy too, concentrated into a few months:
The table values are estimates, not emissions directly measured or disclosed by each model provider. The Stanford AI Index 2026 gives calculated CO2e estimates for GPT-4, Llama 3.1 405B, and Grok 4. Epoch AI estimates substantially higher emissions for Grok 4 using a different method. These are not statistical bounds of a measurement interval. Mistral's lifecycle analysis with ADEME includes 18 months of use as well as training. The figures therefore do not form a ranking of training emissions.
Epoch AI estimates that compute used to train frontier models has grown by about five times per year since 2020. That describes compute, not automatically the electricity use or emissions of each model.
6. The Hyperscalers: Different Metrics, Different Trends
The latest reports show different trends. Electricity demand, operational emissions, Scope 1 to 3 emissions, and location-based Scope 2 emissions are not the same measure. The reporting years also differ:
Google reported that company-wide electricity demand rose 37% in 2025. Its market-based operational Scope 1 and 2 emissions fell 2%, while supply-chain emissions rose 25%. Microsoft's report gives a 25% increase in total Scope 1, 2, and 3 emissions in fiscal year 2025. Amazon reported a 16% increase in absolute emissions in 2025 compared with 2024, while carbon intensity was 38% lower than in 2019. Meta's 2024 data index reports 5,862,614 tonnes of location-based Scope 2 emissions from data centers. These figures cannot be directly compared because their reporting years and boundaries differ.
7. Nuclear Power Deals and Restart Plans
Some technology companies are securing long-term nuclear power or backing new reactor projects. The announcements range from contracts for existing plants to proposals for facilities that have yet to be built:
Microsoft's agreement with Constellation concerns Unit 1 at the former Three Mile Island plant. It operated until 2019 and was not the reactor involved in the 1979 accident at Unit 2. Constellation targets a 2027 restart. In June 2026, FERC approved the transfer of existing grid interconnection rights; the NRC still has to conduct a broader review and approve a restart. The FERC decision was not an operating license. Constellation's August 2026 report describes both approval steps.
Amazon's first agreement in 2024 concerned the Cumulus campus. Talen received $650 million for the assets; the power purchase commitments were designed to rise in stages and could reach 960 MW under a contract option. That was not a fully available supply volume. Talen's SEC filing describes the sale and power agreement.
In November 2024, FERC rejected Talen's request to raise direct power delivery to the Amazon campus from 300 to 480 MW. The existing arrangement remained in place. FERC's order addressed the proposed expansion.
Other projects are at different stages. Talen expects its expanded agreement to supply Amazon with up to 1,920 MW from Susquehanna at full contract quantity, ramping up by 2032. Google's agreement with Kairos covers up to 500 MW by 2035. In September 2026, the US Department of Energy closed a loan of up to $1.9 billion for the 615 MW Duane Arnold reactor; NRC approvals for a restart are still pending. Meta says its agreements could support up to 6.6 GW of new and existing low-carbon energy by 2035. The total includes battery storage and should not be read as nuclear generation capacity. Meta lists the agreements by supplier.
8. Water: Consumption, Withdrawal, and Replenishment
Data centers don't just need electricity, they need water for cooling. Here too, a single query is tiny while the total is huge.
Google estimated 0.26 milliliters of water for a median Gemini Apps text prompt in May 2025. Sam Altman quoted 0.000085 US gallons, about 0.32 milliliters, for an average ChatGPT query. These are provider estimates using different methods, not directly comparable measurements.
Company reports measure different things. Google reported 7.7 billion gallons of water replenished in 2025, equal to 78% of its company-wide freshwater consumption. Amazon withdrew about 2.5 billion gallons across its global data center operations in 2025 and reported a water usage effectiveness of 0.12 liters per kWh for that operation. Microsoft reported that it replenished more than 14 million cubic meters globally in fiscal year 2025 and described its operations as water positive. Replenishment, withdrawal, and consumption are different measures. The boundaries also vary between company-wide operations and data centers, so these figures do not make a comparable ranking.
9. Germany: Data Center Electricity and AI Capacity Grow
And in Germany? The Bitkom/Borderstep study provides the most reliable figures for the German market:
The Bitkom/Borderstep study estimates that German data centers used 21.3 billion kWh in 2025, compared with 12 billion kWh in 2015. The study says installed capacity grew 9% in 2025 to 2,980 MW and projects more than 5,000 MW by 2030. These are study estimates and forecasts, not a continuous official measurement series.
The most interesting number is AI capacity, though:
In the study published on November 10, 2025, AI data centers accounted for about 15% of installed capacity. Capacity allocated to AI is projected to rise from 530 MW to 2,020 MW by 2030. Bitkom expects AI to represent 40% of all German data center capacity by then. The 2030 figures are forecasts from the study.
10. Efficiency: The Counterforce
Electricity use per task can fall while total use grows. Google reported lower energy use for its median Gemini Apps text prompt than a year earlier. More broadly, the IEA describes gains in hardware, software, and model efficiency alongside increased usage and more energy-intensive tasks such as video, reasoning, and agents.
An efficiency improvement per query therefore does not automatically reduce total consumption. That also depends on the number and type of tasks and on data center expansion. The available figures do not isolate a Jevons effect for AI as a whole.
11. Forecasts Compared: Who Says What
These forecasts are useful side by side, as long as their different scopes remain clear:
The table draws on the IEA's 2026 report and 2025 report, the LBNL update, and EPRI's 2026 report.
The values cannot be combined into one forecast. The IEA models global demand, while LBNL and EPRI cover the US market. Their base years, assumptions, and scenarios also differ. Do not read US shares as shares of global electricity use, and treat scenario ranges as scenarios rather than precise predictions.
12. Conclusion: Data Center Growth Is an Infrastructure Challenge
The data paints a more nuanced picture than either camp in the debate.
The IEA estimates that data centers of all kinds used 485 TWh of electricity worldwide in 2025. In its base case, demand reaches about 950 TWh by 2030, or roughly 3% of global electricity use. This global share says little about local areas where new data centers may put more pressure on electricity grids and water resources.
LBNL's US reference case assigns 11.8% of national electricity use to data centers in 2030. The IEA's global base case puts all data centers at about 3% of worldwide electricity use. These shares have different geographic denominators. Company water and emissions figures also vary with reporting years and boundaries. One prompt estimate cannot answer how data center expansion affects energy supply, grids, and climate overall.
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.
