A comparison of estimated water use
Does 38,000 ChatGPT Queries Really Use the Water of One Almond?
The “38,000 ChatGPT queries = one almond” comparison combines a per-query water estimate with an almond water estimate. The arithmetic can be reproduced with selected inputs, but it is not a universal physical constant: the result depends on the accounting boundary, computing location and agricultural estimate.
| Input or calculation | Approximate result |
|---|---|
| Altman's earlier per-query estimate | 0.000085 US gallons ≈ 0.322 mL |
| That estimate × 38,000 queries | 3.23 US gallons ≈ 12.23 liters |
| Published almond-footprint study average | 12 liters per almond kernel |
| 12,000 mL ÷ 0.322 mL per query | About 37,000 queries; near 38,000 after rounding |
Why now
Reports of Sam Altman's September 1 Sources interview drew attention to a comparison between 38,000 ChatGPT queries and one almond. The useful follow-up is to check the arithmetic and what each estimate counts.
The connection
Why here
A memorable ratio hides two independent estimates. Our table shows one way to reproduce approximately that ratio, not proof that these were the exact inputs used in the interview. It also does not show that every ChatGPT request consumes the same water.
More context
Context
What exactly did Sam Altman say?
CalMatters reports that Altman compared 38,000 ChatGPT queries with producing one California almond during his Sources interview. Treat that as an attributed comparison. His earlier blog post gives a separate estimate of 0.000085 gallons per average query, without enough methodological detail to independently audit it.
How does the math work?
Using US liquid gallons, 0.000085 × 3.785411784 liters is about 0.000322 liters, or 0.322 milliliters. Multiply by 38,000 and the result is approximately 12.23 liters. A published California almond study reports an average footprint of 12 liters per kernel. Dividing that amount by the query estimate gives roughly 37,000 queries.
That explains how a result near 38,000 can emerge. It does not establish an exact equivalence. Rounded inputs, different periods and different definitions can move the answer substantially. An almond figure of a few liters would yield a much smaller ratio using the same query estimate.
Why do water estimates differ?
Direct data-center cooling is only one possible boundary. Electricity generation can have its own water footprint; manufacturing and construction may or may not be included. Research on AI water use also considers where and when computation occurs. Cooling arrangements, electricity supplies, model size, hardware and workload can change the estimate.
Agricultural footprints likewise depend on region, growing conditions and the water categories counted. A total footprint is not automatically identical to irrigation water physically applied at an orchard. Comparing a broad crop footprint with a narrow computing estimate would not be a like-for-like audit.
Does one query have a fixed water cost?
No fixed per-query constant is established by these sources. An average across one set of requests is not a measurement of every individual request. A short response and a long computing task cannot be assumed to have identical resource requirements merely because both are called a query.
The defensible conclusion is limited: selected published estimates make the approximate ratio understandable, while missing shared accounting boundaries prevent it from settling the environmental comparison. Neither this calculation nor a disagreement with it proves that AI has no water impact or that every use is environmentally equivalent.
Evidence
Sources
- Fact check: Is Sam Altman right that almonds use more water than ChatGPT queries? — CalMatters
Reports Altman's 38,000-query comparison from the Sources interview and uncertainty about public data-center water data.
- The Gentle Singularity — Sam Altman
Altman's earlier estimate is 0.000085 gallons per average ChatGPT query; the post does not supply a reproducible water-accounting method.
- Water-indexed benefits and impacts of California almonds — Fulton, Norton and Shilling / California State University Sacramento
The study reports an average footprint of 12 liters per California almond kernel with substantial variation; this is a study estimate, not a universal constant.
- Making AI Less Thirsty — Li, Yang, Islam and Ren
Research distinguishes direct and indirect AI water use and geographic and temporal variation; it does not validate a fixed water cost for every present-day ChatGPT query.