- US data centers consumed 183 terawatt-hours in 2024, equal to 4.4% of national electricity demand.
- By 2030, US data center power demand could more than double, reaching 12% of total consumption.
- Tech companies refuse to publish actual AI energy figures, forcing researchers to estimate indirectly.
Alex de Vries keeps asking a question that tech companies would rather not answer: how much electricity does AI actually consume?
The researcher at VU Amsterdam has spent years tracking the environmental footprint of digital technologies through his research platform Digiconomist. His work suggests AI systems could soon rival the energy appetite of small nations. The problem is getting companies to confirm it.
Data centers powering AI consumed roughly 183 terawatt-hours in the United States during 2024, according to IEA estimates. That represents 4.4% of the country's total electricity demand.
Key figure
133%
Projected growth in US data center electricity use by 2030
The scale becomes clearer with projections. By 2030, US data center power demand could more than double, potentially reaching 12% of the country's electricity consumption.
What is inference?
When you ask ChatGPT a question, the AI generates its response through a process called inference. Training a model happens once, but inference happens billions of times daily as users submit prompts. This constant answering phase consumes far more cumulative energy than initial training.
The Transparency Gap
De Vries and fellow researchers face a stubborn obstacle: major tech companies guard their actual consumption figures closely.
The International Energy Agency projects global data center electricity demand will reach 945 terawatt-hours by 2030. That equals Japan's entire annual consumption. But these projections rely on extrapolation rather than reported data.
Companies prefer it that way. Without hard numbers, assessing true environmental impacts becomes nearly impossible. Holding firms accountable for their sustainability pledges gets even harder.
De Vries has developed workarounds. By tracking server shipments from manufacturers like NVIDIA, then multiplying by known power requirements, researchers can estimate AI's energy footprint indirectly. The estimates keep rising.
Virginia Sounds the Alarm
The consequences of AI's energy hunger are already reshaping communities.
Virginia hosts more data centers than anywhere else on Earth. A state-commissioned review in December 2024 warned that continued growth could double electricity demand within ten years.
Some approved facilities already face three-year delays. Local utilities simply cannot provide power fast enough to meet construction schedules.
Sarah Parmelee maps data center expansion for the Piedmont Environmental Council. She describes the emerging problem simply: districts promise to buy power from each other. But everyone is planning to buy from everyone else.
The Efficiency Paradox
Could more efficient AI models solve the problem?
History offers a cautionary answer. The nineteenth-century economist William Stanley Jevons observed that making technologies more efficient often increases total energy use. When something becomes cheaper to run, people use it more.
De Vries sees the same pattern with AI. Better models at lower computational cost have not reduced energy demand. They have enabled more applications and attracted more users.
The path forward requires something tech companies have resisted: genuine transparency about what AI actually consumes. Three labs are now developing standardized measurement methods that could make independent verification possible.
Sources
- Primary Research: de Vries, A. "The growing energy footprint of artificial intelligence." Joule (2023)
- Additional Context:
- How much energy will AI really consume? (Nature, March 2025)
- Energy and AI Report (International Energy Agency, 2025)
- What we know about energy use at U.S. data centers (Pew Research, October 2025)
Fact Check: Claim-by-Claim Verification Verified
Claims are factually accurate; key statistics match peer-reviewed sources and established research institutions, with appropriately hedged language for projections.
Commentary
- The Nature article reference cited as "December 2024" actually published in March 2025 (volume 639, issue 8053), though the article itself discusses recent developments and the timing reflects current context
- The claim that data centers consumed "183 terawatt-hours in the United States during 2024" is accurate per IEA but includes all data centers, not exclusively AI—though AI is acknowledged as the primary growth driver
- The article appropriately uses "could" and "may" when discussing projections, maintaining appropriate hedging for uncertain future estimates
- Reference to "three labs" developing standardized measurement methods is mentioned but specific identification of these labs is not provided in sources reviewed
Sources used for verification
Academic/Peer-reviewed:
- de Vries, A. "The growing energy footprint of artificial intelligence" Joule (2023) - Cell.com
- Chen, S. "How much energy will AI really consume?" Nature (March 2025) - Nature
- Virginia Joint Legislative Audit and Review Commission Report on Data Centers (December 2024) - Virginia.gov
Other reliable sources:
Fact-checked by Perplexity Sonar Pro on 2026-01-16
