Speaking at ADIPEC in Abu Dhabi, the OpenAI chief paired big efficiency claims with a warning that energy remains AI's biggest constraint
OpenAI CEO Sam Altman says the energy cost of a single AI query has collapsed. By his estimate, queries now use somewhere between 1,000 and 100,000 times less energy than they did three to four years ago.
He said it at an oil and gas conference, flanked by an announcement with a national oil company.
What Altman actually said
The remarks came at the ADIPEC energy conference in Abu Dhabi on November 2, 2026. Altman credited the drop in per-query energy to the development of more efficient AI models.
His phrasing was looser than the headline numbers suggest. Queries are now “thousands or hundreds of thousands of times maybe more efficient” than a few years back, he said.
Note the “maybe.” It carries real weight in that sentence, and the range itself spans two full orders of magnitude.
Altman did not frame this as a victory lap, though. He told the audience that energy is the primary challenge facing the AI industry.
Cheaper queries, far more of them
The reason both claims can be true at once comes down to demand. Per-query efficiency has improved, but people are running vastly more queries.
According to the research summary of his remarks, top AI users now consume hundreds of billions of tokens every month. Six years ago, that figure was roughly 100,000 tokens.
Tokens are the small chunks of text that AI models read and write, roughly a word or a fragment of one. Every token processed draws compute, and compute draws power.
Going from about 100,000 tokens to hundreds of billions is a jump of several orders of magnitude for the heaviest users. Tokens and queries are not the same unit, so the two figures don’t map neatly onto each other. Still, the direction is clear: usage has grown far faster than most efficiency curves can comfortably absorb.
The ADNOC and G42 partnership
Altman also used the stage to announce a multi-year partnership between OpenAI, the Abu Dhabi National Oil Company (ADNOC), and Abu Dhabi-based AI firm G42.
The deal aims to deploy frontier AI models across ADNOC’s operations, with the goal of optimizing how the energy giant runs its business.
Altman expressed hope that AI could eventually improve global energy availability. His argument is that efficiency gains in production and operations, driven by AI, could free up supply across the system.
A familiar Altman argument
None of this marks a break from Altman’s previous public positions. He has repeatedly advocated for expanding energy generation capacity, pointing to renewables and nuclear power as the paths forward.
He has also pushed for what he describes as a more intelligent conversation about AI’s energy needs. Part of that pitch involves comparing the energy AI consumes against the energy humans use more generally.
The efficiency claim fits neatly into that framing. If each query keeps getting cheaper, critics focused on per-query consumption are arguably aiming at a shrinking target.
The counterargument writes itself, though. Total consumption is what hits the grid, and total consumption depends on volume as much as efficiency.
What this means
His decision to name energy as the industry’s primary challenge, rather than chips, talent, or regulation, signals where OpenAI sees the real bottleneck.
There are caveats worth watching. Altman’s 1,000x to 100,000x range is his own estimate, delivered with a hedge, and no breakdown of how it was calculated accompanied the remarks. Until per-query energy data is measured and published in a consistent way, the figure is best read as a directional claim rather than a benchmark.
For now, the message from OpenAI’s chief is two-sided. AI is getting dramatically more efficient per task, and the industry still needs a lot more power.
