Playbook No. 04 — Published Friday Oct 2, 2026
What one AI question really costs — the five-minute board brief on AI and energy. The question your board keeps asking, answered straight — no hype in either direction. Forwardable as-is.
The question your board keeps asking, answered straight — no hype in either direction. Forwardable as-is.
Somebody on your board read a headline about AI data centers draining the power grid. Now every AI conversation detours into energy anxiety. They're not an outlier — they're the room. Six in ten Americans now oppose the construction of a data center near them, up twelve points in four months — the largest move on any AI question the Annenberg Public Policy Center has measured. And it's not just the skeptics: even the heaviest AI users are about as opposed as everyone else (60% among heavy users, 64% among nonusers).
So here's the thing you forward. The honest numbers, what they mean, and what they don't.
Forget the sci-fi monolith from the headlines. A data center is a warehouse full of computers with serious air conditioning — a building with a power bill, like everything else. When you use "the cloud," you're renting time on someone else's computers. And sharing beats everyone owning their own: one well-cooled building full of computers uses power more efficiently than hundreds of little servers humming away in closets and back offices, each with its own cooling and each sitting idle half the time. Consolidation is the efficiency nobody brags about, because nobody sells it — but it's real.
About 0.3 watt-hours — a third of a Wh. That number comes from three sources that land in the same band: Google measured 0.24 Wh for a typical Gemini prompt across its own infrastructure and published the full methodology (August 2025); the independent research shop Epoch AI estimated ~0.3 Wh for a typical ChatGPT query from the bottom up (February 2025); and OpenAI's CEO says 0.34 Wh on his blog — no methodology published, so treat his number as a company claim that the independent work happens to agree with.
What does 0.3 Wh feel like? About two minutes of a 10-watt LED bulb. A full phone charge — roughly 10 Wh — covers about thirty AI questions. A cold-wash load of laundry (about 0.2 kWh of machine energy) runs about six hundred questions. (Add the dryer — 2 to 3 kWh — and the dryer alone uses ten times the washer's energy. Laundry math has a twist.)
And the "ten times a Google search" claim you've heard? It's built on an outdated 3 Wh figure from the GPT-3 era that researchers now say overstates reality by about ten times. Against Google's own 2009 search figure (0.0003 kWh — which is 0.3 Wh), one AI question costs roughly the same as one old-fashioned web search. The scary version needed the old numbers.
The honest caveat: this is the typical text question. Reasoning models, image generation, and long documents cost more — outside this range.
Two real numbers, kept apart because they measure different eras. A 2023 peer-reviewed study (accepted for publication in the Communications of the ACM) found that ChatGPT used the equivalent of a 500 mL water bottle per 20-to-50-question conversation — roughly a tablespoon per question. (The viral "a bottle per question" is a misreading of the study.) Google's 2025 measurement for a typical Gemini prompt: 0.26 mL — five drops. The gap between "a tablespoon" and "five drops" is mostly different eras and different measurement boundaries, not a contradiction — don't stack them into a trend.
And the big denominator most headlines skip: most of AI's water footprint isn't in the data center at all. The Energy Department's 2024 data-center report estimates U.S. data centers used 66 billion liters of water directly — but nearly 800 billion liters indirectly, in the power plants generating their electricity. The water question is, at bottom, the energy question wearing a different hat.
Both things are true. All U.S. data centers used 176 TWh in 2023 — 4.4% of U.S. electricity, per the Energy Department's Lawrence Berkeley National Lab report. Their models put 2028 at 6.7 to 12% — a scenario range, not a prediction. That's every data center, not just AI: GPU-accelerated AI servers were a bit over 40 TWh in 2023, inside that 176. Globally, the International Energy Agency puts data centers at about 415 TWh in 2024 — roughly 1.5% of world electricity — with a 2030 base case near 945 TWh, about Japan's annual consumption.
Efficiency keeps improving — historically, computing doubled in energy efficiency every year and a half, though that pace has slowed since around 2000, and the roughly tenfold drop in per-query energy from 2023 to 2025 fits that pattern. But the gains get spent on harder work: more queries, longer reasoning, bigger models. The efficiency story is real; the aggregate growth is real; neither cancels the other.
Forget "will AI boil the planet" — that debate has no finish line, and the answer doesn't change what your organization does on Monday. These three questions have answers, and the answers change what you do:
The one action: forward this issue to the board member who keeps asking. The brief does the rest. Time saved: the hours you'd spend researching an answer — or arguing about the question. Value: a verified, forwardable answer that treats the skeptic with respect instead of dismissing them.
Published Friday, October 2, 2026. The Issue #4 newsletter edition wraps this playbook and lands in inboxes separately — browse the newsletter archive. This page is the playbook itself, frozen as published: playbooks never change after they go up. If a better way appears, it gets a new number, not a rewrite.
Free forever. One playbook every Friday.
Back to the shelf · ← Playbook No. 03 — The one-page playbook · Playbook No. 05 — Volunteers, scheduled →
One tested AI playbook per issue — for grant writing, donor thank-yous, volunteer scheduling, board reports — you can put to work the same day you read it. Free. One email per issue. Unsubscribe anytime.