In 2024, U.S. data centers consumed roughly 183 terawatt-hours of electricity. Enough to power every home in the state of Texas for more than a year.¹ Nobody received a bill for their share.
But what if they did?
Suppose every household and every company had to generate on-site, in real time, the electricity equivalent to their AI processing consumption. Not buy offsets. Not pay a surcharge. Actually produce it.
The Daily Habit
A typical ChatGPT query consumes about 0.3 watt-hours of electricity.² That's roughly ten times what a Google search uses. At 20 queries a day (a fairly active user) you'd burn through about 2 kWh per month just on text prompts. Manageable. About what your phone charger uses.
Now Add Imagery
Generate an AI image and the equation shifts slightly: about 0.6 watt-hours per image.³ Still modest individually. But ask for a five-second AI video and the math moves somewhere most people don't expect. That single clip consumes nearly 1 kWh,⁴ roughly equivalent to running your refrigerator for a full day.
One clip=One day of cooling your food.
Now Make It Your Job
For a power user, generating videos, running extended research sessions, processing documents daily, the annual AI electricity tab could approach 200–300 kWh per year. That's a meaningful slice of the average American household's 10,500 kWh annual consumption.⁵ To generate the electricity needed for an individual user you would need to install one dedicated solar panel on your roof before running your next prompt. (Which would most likely be: how do I install a solar panel?)
Now Make It Your Company
Apply that same logic to the AI companies running processing for millions of people at home and at work.
A typical AI-focused hyperscale data center uses as much electricity annually as 100,000 households.¹ The largest ones under construction are expected to use 20 times that. Each of them would need to become, effectively, a power utility.
Some already are. Microsoft signed a 20-year deal to restart a dormant reactor at Three Mile Island with 835 megawatts directed entirely at its data center operations.⁷ Amazon purchased a campus adjacent to a nuclear plant in Pennsylvania. Google negotiated a clean energy tariff with a Nevada utility. Oracle is designing a campus backed by three small modular reactors.⁸
The thought experiment is starting to look less like a thought experiment.
States, meanwhile, have begun moving on their own: Texas passed legislation in 2025 giving its grid operator authority to disconnect very large data center loads during emergencies and requiring developers to fund the grid upgrades they necessitate.¹⁰
There's a wrinkle worth noting. When you stream KPop Demon Hunters on Netflix, ask Google Maps to reroute you, or order food from DoorDash, you're also drawing from data center electricity. AI didn't create the invisible energy bill. It made it larger, and harder to dismiss.
By 2028, AI-specific electricity demand in the U.S. is projected to reach between 165 and 326 terawatt-hours per year — potentially enough to power 22% of American households.¹¹ That energy is coming from somewhere. And whoever doesn't generate it themselves draws it from a grid built for everything (and everyone) else.
⚡️On March 4th, seven of the world's largest AI companies — Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI — signed the White House's "Ratepayer Protection Pledge," committing to cover electricity costs from their data centers rather than pass them to consumers. Anthropic made a separate but aligned commitment on February 12th, pledging to cover 100% of consumer electricity price increases attributable to its data centers.
It's worth noting what both commitments are and are not: a voluntary promise to pay for electricity, not to produce it.
Why This Matters
One solar panel is a concrete number. A heavy AI user's annual processing footprint could rival a major household appliance. At scale, across hundreds of millions of users, that's a meaningful and largely untracked energy demand. It's also roughly equivalent to what a single standard rooftop solar panel produces. A specific physical object with a specific cost: about $1,200 installed.⁶ Visibility tends to change behavior, or at least it tends to produce the question of whether that behavior is priced correctly.
Nobody has agreed on where the lines are. A law firm running CoPilot across 200 employees, a hospital using AI diagnostics, an agency using Gemini to generate content at scale; none of them appear cleanly in the individual tab or the hyperscaler's energy footprint. The AI energy question has the same nested accountability problem as corporate carbon accounting. The individual produces it, the enterprise buys it, the provider runs the infrastructure. Does the obligation belong to the firm, or to the Claude or ChatGPT subscription they're paying for? Right now, none of the three tiers have agreed on where it transfers.
The accounting was always broken, AI just made it obvious. Long before anyone asked an AI to write an email, the energy cost of a Google search, a Netflix stream, a social media scroll was real and unmeasured at the individual level. What's different now is scale and speed: AI made the invisible bill grow fast enough that ignoring it became a political problem. It's an accounting story disguised as a technology story.
Processing the Electricity Hidden In Every Query
There's an old rule in economics: when you can't see the cost, you can't manage it. What this thought experiment surfaces, ultimately, isn't an energy problem. It's a visibility problem. The same one that took decades to surface in carbon accounting, water rights, and supply chain labor. The resource was always being consumed. The issue was always whether the people benefiting most were the ones being asked to account for it.
No matter how it gets produced, there is only so much energy available at any one time. If it came down to it, would you be willing to install a solar panel to maintain your AI usage, or more pointedly, would you change how you use it if you had to?
Sources
[1] Pew Research Center
[2] IAEI Magazine
[3][4] AI Multiple
[5] U.S. Energy Information Administration
[6] This Old House / Solar Survey, 2025
[7] MIT Technology Review, Sept. 2024
[8] CNBC, Sept. 2024
[9] American Bazaar, Dec. 2025
[10] Yale Clean Energy Forum, Nov. 2025
[11] MIT Technology Review, May 2025
[12] KTVH/Scripps News, March 2026
[13] CyberNews, Feb. 2026
[14] CNBC, March 2026