AI Energy Consumption: Big Tech’s Electricity Use Rose 142% in Four Years While US Demand Rose 7%

AI energy consumption is easiest to track in the companies’ own environmental reports. Google, Microsoft and Meta used 80.0 terawatt-hours of electricity in 2024, up 142% from 33.1 TWh in 2020, while total US electricity use grew 6.6%. Google alone used 43.6 TWh in 2025, 37% more than a year earlier.

The energy ledger: what each company reports

Of Google, Microsoft, Meta and Amazon, three publish an annual electricity total in the data tables of their environmental reports, and each says an outside accounting firm reviewed selected environmental metrics. We read those tables and copied the figures into one ledger, in terawatt-hours. Amazon is the gap: its 2025 Sustainability Report gives renewable matching and water efficiency figures but no total electricity consumption, so it cannot be added.

Data centers are nearly all of it. They were 97.3% of Google’s electricity in 2025 and 98.0% of Meta’s in 2024. Microsoft does not split data centers out of its total, but 56% of its fiscal 2025 electricity was used in North America, according to Table 13 of its fact sheet. For the national picture of where data centers sit, see our data centers by state atlas.

Is big tech’s electricity growing faster than US demand?

Yes, by a wide margin. From 2020 to 2024, the latest year all three companies have reported, their combined use grew at 24.7% a year. Total US electricity end use, from the Energy Information Administration, grew 1.6% a year over the same span, so the companies grew about 21 times as fast. In absolute terms they added 46.9 TWh, while the whole country added 254 TWh. The company figures are worldwide, so the two numbers cannot be subtracted, but the growth rates can be compared.

The gap widened in 2025. Google’s electricity use rose 37.4% to 43.59 TWh, and its data centers alone rose 38.4%. Microsoft’s rose 24.1% in fiscal 2025 to 37.03 TWh. US end use grew 2.1%. Between them, Google and Microsoft added 19.1 TWh worldwide in one year; total US end use rose 85 TWh. Meta’s 2024 total was 18.42 TWh, up 20.2%.

Electricity is growing faster than revenue

If the extra power were simply following the business, electricity per dollar of revenue would hold steady. It has not. We divided each company’s electricity by its revenue from SEC 10-K filings. Microsoft publishes the same ratio itself, and our figures match its fact sheet: 75 MWh per $1 million in fiscal 2020 and 131 in fiscal 2025.

At all three companies, electricity grew at least 1.6 times as fast as revenue. The reports do not say how much of the extra power went to AI, so this is a test of direction, not attribution: the electricity behind each dollar these companies earn is rising at all three.

How much of this is AI?

None of the companies separates AI from the rest of its electricity. The best public breakdown is the 2024 United States Data Center Energy Usage Report from Lawrence Berkeley National Laboratory, written for Congress under the Energy Act of 2020. It models every server type in US data centers. GPU-accelerated AI servers went from less than 2 TWh in 2017 to more than 40 TWh in 2023 (report page 49). That is at least 40% of the roughly 100 TWh all US servers used, and at least 23% of total US data center electricity, before counting the cooling those servers need. Our data center statistics report covers the full LBNL series, its 2028 range and the EIA context.

LBNL’s projections turn on AI. Server electricity alone reaches 240 to 380 TWh in 2028 in its scenarios, and AI servers make up 8 to 12 million of the installed base by then, depending on how many GPUs ship. Worldwide, the International Energy Agency’s Energy and AI report estimates data centers used about 415 TWh in 2024, around 1.5% of global electricity, and projects about 945 TWh in 2030 in its Base Case. In that case, electricity for accelerated servers grows 30% a year against 9% for conventional servers, and US data center use rises by about 240 TWh, or 130%.

How much energy does one AI prompt use?

Two companies have published a per-query figure. Both are company claims: they come from internal measurements no outside party has checked, they use different definitions (a median prompt versus an average query), and neither includes the energy used to train the model.

The per-prompt numbers and the company totals tell different stories. Google says the energy of its median text prompt fell 33 times in the year to May 2025, yet its total electricity still rose 37% in 2025. Lower energy per prompt has not lowered the total.

AI energy consumption compared to US states

EIA’s Electric Power Annual reports how much electricity each state’s homes and businesses bought in 2024 and 2025. We compared each company total with state sales for the same year. Google alone used more than 20 of the 50 states and DC in 2025. The comparison mixes a worldwide company total with state retail sales, so it shows scale, not where the power was used.

State figures: EIA Electric Power Annual Table 2.8, total retail sales to ultimate customers, 2024 and 2025, in TWh. Ranks include the District of Columbia.

What this means for MSPs

Plan AI capacity, not just price. The three companies’ electricity use grew 24.7% a year from 2020 to 2024, faster than revenue at each of them, and our cloud computing statistics show the capital spending behind that growth. When scoping client AI pilots, check GPU region availability and reserved-capacity terms as well as list price.

Answer client questions with sourced numbers. Clients with sustainability reporting duties may ask what their AI tools cost in energy. The honest answer today is a company claim: about 0.24 Wh for a median Gemini text prompt and 0.34 Wh for an average ChatGPT query, as stated by the providers. Cite the source and say it is unaudited.

Use provider reports for scope 3 estimates. Google, Microsoft and Meta publish electricity and emissions tables with limited assurance from outside accountants. Those tables are primary sources a client’s carbon accountant can cite, unlike vendor marketing pages. Amazon publishes no electricity total, which matters for clients who must estimate cloud emissions from AWS usage.

Watch water as well as power. The same environmental reports disclose water use by site. Our companion report on data center water usage tests whether water is rising as fast as electricity.

Frequently asked questions

How much electricity does AI use?

No company reports AI on its own. The closest measured figures are company totals: Google used 43.6 TWh in 2025, Microsoft 37.0 TWh in fiscal 2025 and Meta 18.4 TWh in 2024. For the US, Lawrence Berkeley National Laboratory estimates GPU-accelerated AI servers used more than 40 TWh in 2023, about 23% of the 176 TWh all US data centers used.

Does AI actually use a lot of electricity?

Measured against the whole country it is modest; measured by growth it is not. Google, Microsoft and Meta together used 80.0 TWh in 2024, equal to about 1.9% of total US electricity end use, although part of it was used outside the US. Their combined use grew 142% from 2020 to 2024, against 6.6% for all US electricity.

How much energy does a ChatGPT or Gemini prompt use?

By the companies’ own figures, a median Gemini Apps text prompt used 0.24 watt-hours in May 2025 (Google), and an average ChatGPT query uses about 0.34 watt-hours (OpenAI CEO Sam Altman, June 2025). Both are company claims that no outside party has audited, and neither covers training.

How much energy does AI use per day?

Google’s 43.6 TWh in 2025 works out to about 119 GWh a day, and the three companies together averaged about 218 GWh a day in 2024. These are totals for all their operations, not AI alone.

How much electricity will AI data centers use by 2030?

The IEA’s Base Case puts global data center use at about 945 TWh in 2030, up from 415 TWh in 2024, with electricity for accelerated servers growing 30% a year. For the US, LBNL projects 325 to 580 TWh for all data centers by 2028.

How does AI energy consumption compare to a US state?

Google’s 43.6 TWh in 2025 was more electricity than 20 of the 50 states and DC bought at retail that year, close to Kansas (41.5 TWh). The three-company total for 2024 exceeded retail sales in 33 states and DC.

Methodology

References

1. Google, 2026 Environmental Report, Environmental metrics data tables (electricity consumption, p. 93) and Figure 4 (p. 17). Accessed October 2026.
2. Google, 2025 Environmental Report, electricity consumption table, p. 107. Accessed October 2026.
3. Google, 2024 Environmental Report, energy table, p. 77. Accessed October 2026.
4. Microsoft, 2026 Environmental Data Fact Sheet, Tables 5 and 13 (pp. 6 and 24). Accessed October 2026.
5. Meta, 2025 Environmental Data Index, section 2.1 Electricity (PDF p. 6). Accessed October 2026.
6. Amazon, 2025 Sustainability Report (checked for an electricity total; none published). Accessed October 2026.
7. Shehabi, A., et al., 2024 United States Data Center Energy Usage Report, Lawrence Berkeley National Laboratory, LBNL-2001637 (December 2024), pp. 49 and 52. Accessed October 2026.
8. International Energy Agency, Energy and AI: Energy demand from AI (April 2025), licence CC BY 4.0. Accessed October 2026.
9. Elsworth, C., et al. (Google), “Measuring the environmental impact of delivering AI at Google Scale,” arXiv:2508.15734 (August 2025). Accessed October 2026.
10. Altman, S., “The Gentle Singularity,” June 10, 2025. Accessed October 2026.
11. U.S. Energy Information Administration, Monthly Energy Review, Table 7.6 Electricity End Use. Accessed October 2026.
12. U.S. Energy Information Administration, Electric Power Annual, Table 2.8 Sales of Electricity to Ultimate Customers by State, 2025 and 2024. Accessed October 2026.
13. U.S. Securities and Exchange Commission, EDGAR XBRL company facts for Alphabet, Microsoft and Meta (10-K revenue). Accessed October 2026.