AI Energy Efficiency Per Query: Why Demand Still Grows

Sam Altman says AI energy efficiency per query rose thousands of times in 3-4 years. Why total power demand still grows — and what it means in Japan.
Sam Altman said the energy consumed by a single AI query has become thousands to hundreds of thousands of times more efficient over the past three or four years. He said it at the opening ceremony of ADIPEC, one of the world’s largest energy exhibitions, held in Abu Dhabi. What caught my attention was not the size of the number but what he said immediately after it: as the unit price falls, usage grows, so total electricity demand rises rather than falls. This piece covers how to handle that statement, why efficiency and total demand point in opposite directions, and what assumptions companies working in Japan should use when they budget for AI power and cost.
What Altman actually said about AI energy efficiency per query
The statement breaks into four parts.
- The energy consumed per AI query has dropped sharply
- The improvement is thousands to hundreds of thousands of times versus three or four years ago
- The better the efficiency gets, the more usage grows
- The AI industry and the energy industry both carry a responsibility to society
The order matters. He did not stop at the efficiency boast. He went straight from efficiency to the demand it unlocks, and landed on industry responsibility. The fact that the most visible figure in AI chose that sequence, in a room full of energy people, carries about as much information as the content itself. It also explains, in one arc, why AI and energy are now discussed as a single subject.
How to handle a number that spans orders of magnitude
This part needs care. The thousands-to-hundreds-of-thousands range was offered with a “maybe” by Altman himself, and OpenAI has not released it as a measured figure. A range that wide cannot be verified or compared as stated. What counts as one query? Which configuration is being compared with which? Does it cover electricity only, or everything around it? Without those premises, the number is an arrow pointing in a direction, not a measurement.
So I read the statement as evidence of what the top of the industry chose to say, and in what order, on an energy stage — not as a published efficiency metric. If you cite it internally or in a proposal, attach three things:
- That it is a rough estimate, not a published measurement
- The venue: the ADIPEC opening
- That it informs direction only and should not feed your own cost model
The opposite failure is to lift the number alone and conclude that AI’s power problem is solved. Altman’s own next sentence pushes the other way.
Why total power demand grows as efficiency improves
The shape here is close to the Jevons paradox: per-unit efficiency improves, total consumption still climbs. Once each call is cheap and light, uses that never penciled out suddenly do. In practice the shift looks like this:
- Jobs a human used to approve before sending now run continuously
- Where one answer was enough, teams now generate several and compare
- Chat-shaped usage turns into agents running long chains of steps
- A tool used by one team becomes a standard step in a company-wide process
Every one of those became possible because the unit price fell — and volume tends to grow faster than price drops. That is why your invoice rises even as consumption per call falls, and why society-wide demand rises too. Efficiency and total demand simply do not move together. It is not a contradiction; it is the ordinary consequence of something cheap getting used a lot.
How this lands in Japan
In Japan, any conversation about AI and electricity connects quickly to data center siting and power prices. Bring an AI plan to a management meeting here and you increasingly get asked about cost and power sustainability before accuracy or impact. Altman’s remark hands you both halves of the answer at once: per query it has gotten dramatically lighter, and the total still goes up.
Whether you can hold both at the same time changes the quality of the internal discussion. Carry only the unit-price half and the plan drifts into optimism. Carry only the total-demand half and it becomes an argument for stopping. Put them side by side — design on the assumption that unit prices fall, and prepare on the assumption that total spend and total demand rise — and the plan starts matching reality.
The same applies to overseas AI companies communicating in Japan. The harder you push the efficiency number, the more the Japanese side wants to know what happens to their power bill and their grid. A message built only on efficiency gains is hard to get trusted here. Volunteering both sides — efficiency up, total demand up — reads as honest rather than defensive. This is a market where the caveat you add yourself buys more credibility than the headline you lead with.
What to change in your own AI planning
Here are the decision rules I took away.
Forecast unit price and total volume separately
Put the unit-price outlook and the volume outlook in the same table as two independent lines. Price trends down; volume trends up. Their product is what you actually pay. Any plan that runs a single line — prices fall, so costs fall — will miss.
Set ceilings before usage grows, not after
Cheaper means more usage. Treat that as something that happens, not something to suppress. Instead of restricting which teams may use AI, put limits and visibility in place first. If you can see what is running and how often, growth itself stops being the problem.
Keep estimates labeled as estimates
When you quote the efficiency figure, carry the venue and the hedge with it. Stating an unverifiable order of magnitude as fact in an external document costs you credibility on everything else in the deck. This is the cheapest discipline on the list and the one with the best return.
A separate observation from the same thread
The same thread also carried a point from @GoingBallistic5: making a machine run fast and bringing up a whole production line are two different problems. The clip is a roughly one-minute excerpt of a conversation distributed by RoboStrategy. The figure of one million units that comes up alongside it is described as a long-term Tesla goal, not a confirmed production plan for next year.
That distinction rhymes with the energy point. Speeding up one unit is not the same as making the whole system run at volume. In AI, improving model efficiency and running the full system — including the electricity behind it — are different pieces of work, and progress on the first does not settle the second.
Putting this to work
The practical steps are small. Split your AI budget sheet into a unit-price column and a volume column, and derive next period’s total on the assumption that volume grows. Keep one sentence ready for when cost or power comes up: unit prices fall, total consumption rises. And whenever you use the efficiency figure, say that it is an estimate. Those three habits are enough to keep you steady the next time someone asks how AI and electricity fit together.
FAQ
What did Sam Altman say about AI energy efficiency per query?
He said the energy used per AI query has become thousands to hundreds of thousands of times more efficient compared with three or four years ago. He said it at the opening of ADIPEC, one of the world's largest energy exhibitions, held in Abu Dhabi. He also said the AI and energy industries carry a responsibility to society.
Is the thousands-to-hundreds-of-thousands figure reliable?
He framed it himself as a rough estimate with a "maybe" attached, and OpenAI has not published it as a measured value. A range spanning three or more orders of magnitude cannot be verified as stated, so cite it with the hedge and the venue intact.
If efficiency is improving, why would power demand keep rising?
Because usage grows as the unit price falls. Once a single call is cheap and light, workloads that were never worth running start running, and total volume outgrows the efficiency gain. The pattern is close to the Jevons paradox.
Based on a post on XOriginal post ↗