Reference
Which AI statistics should you stop repeating?
Short answerFour widely repeated figures fail when traced to a primary source: inference as 80 to 90% of AI energy, 71% local opposition to data centers, a 150 kW rack, and 95% of generative AI pilots producing no profit. Each has a version that holds, or nothing that does.
Each of these appears in briefings, presentations, and trade press. None survives being traced to a primary source. We used one of them ourselves until we traced it, which is why this page exists: a claim nobody can source is a claim nobody should make, and that has to apply to us first.
| The claim | Why it does not hold | What can be said instead |
|---|---|---|
| Inference is 80 to 90% of AI energy | No primary source exists for it. It is absent from the International Energy Agency's Energy and AI report and circulates through trade press citing itself. | Google's metered split of its own machine-learning energy: about three-fifths inference to two-fifths training, holding across three consecutive years. See training and inference. |
| Seventy-one percent of Americans oppose data centers locally | Gallup never published a percentage. The figure is other people rounding Gallup's own words into one. | "Seven in 10," which is what Gallup wrote. Its intensity and party figures, such as 48% strongly opposed, are published exactly and can be quoted as given. |
| A rack draws 150 kW | It is a published figure for no current part. The 132 kW seen in trade press is wrong in the same way. | About 120 kW, from NVIDIA's own documentation for the GB200 and GB300 NVL72. |
| 95% of generative AI pilots produce no profit | One report, built on 52 executive interviews and never peer reviewed, which scored anything without a payback inside about six months as a failure. | Nothing, on that evidence. The narrower point holds: measurable returns today concentrate in predictive systems doing bounded work with an answer to check against. See predictive and generative AI. |
Why the pilot figure misleads
The definition does most of the work. A coding assistant that quietly saves engineering hours across a company is real value, and a six-month payback test scores it as a failure. Generative returns are harder to attribute, which makes them a measurement problem. The figure treats that problem as proof of failure.
The test that produced this page
Ask of any number what it was measured on, and by whom. Where the answer is a survey of intentions, a press release, or a compilation of other people's compilations, the number describes a mood. Where nobody can name a primary source at all, there is no number.
The test is worth more than the list. Apply it to every AI figure that reaches a decision, including the ones in vendor material.
Sources
- International Energy Agency, Energy and AI, April 2025.
- Patterson, Gonzalez, Hölzle, et al., The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink, IEEE Computer, 2022. Google's metered split of its own machine-learning energy.
- Gallup, Americans Oppose AI Data Centers in Their Area, published 13 May 2026, fielded 2 to 18 March 2026. n=1,000, ±4 points. Gallup states the headline as "seven in 10."
- NVIDIA, GB200 and GB300 NVL72 documentation. About 120 kW per 72-GPU rack.
- MIT Project NANDA, The GenAI Divide: State of AI in Business 2025. The source of the 95% figure: 52 executive interviews, not peer reviewed, with a six-month definition of success.
Reviewed