ChatGPT · Impact
How Much Water and Energy ChatGPT Uses
This question has become unusually hard to answer honestly, because the widely circulated numbers disagree by orders of magnitude and almost every article picks the one that suits its angle. The disagreement is real but it is not mysterious, and once you understand what each figure counts, both stop looking like propaganda.
GPTPrompts.AI Editorial
Written from primary sources, with disagreements shown rather than hidden · Last updated August 2026
How much water does one ChatGPT query use?
Sam Altman, OpenAI
Approximately 0.3 millilitres
Counts water evaporated on site in data centre cooling for an average query.
Researchers at UC Riverside
Substantially higher per prompt
Adds indirect water: the water consumed by power plants generating the electricity the data centre uses.
Why they differ. Neither side is fabricating. Direct water use counts what evaporates in cooling. Indirect water use also counts what is consumed producing the power, which is often the larger share and varies enormously with the local electricity mix. A data centre on hydro-heavy grid power and one on thermal power have very different indirect footprints for identical compute. Any figure quoted without stating which boundary it uses is close to meaningless.
01
Why a single query is the wrong unit anyway
Per-query figures are compelling because they are relatable, and misleading for the same reason. They make the impact feel like a personal choice about whether you ask one more question.
The larger numbers sit elsewhere: in training runs, in the total volume of queries across hundreds of millions of users, and in the construction and power provisioning of data centres. Your individual usage is a rounding error against that, which is an argument about where scrutiny belongs, not an argument that the impact is zero.
02
What actually drives the footprint
Three variables matter far more than how many prompts you personally send.
- The electricity mix where the data centre sits. Same compute, very different emissions and indirect water.
- Cooling design. Evaporative cooling uses water to save electricity; other designs trade the reverse.
- Model size and how the request is served. A large model answering a trivial question costs far more than a small one doing the same job.
03
Reading claims about AI and the environment
Both directions of exaggeration are common. Some coverage takes the highest available per-query estimate and multiplies it into an alarming annual total. Some industry communication quotes only direct cooling water and omits the electricity entirely.
A few questions separate a solid claim from a weak one.
- Does it say whether the figure is direct only, or direct plus indirect?
- Is the source the company, an academic group, or a blog citing a blog?
- Is it per query, per response length, or per training run? These differ by many orders of magnitude.
- Does it name the region and its power mix, or treat all data centres as identical?
Frequently asked questions
So how much water does ChatGPT actually use per query?
There is no single honest number without stating the boundary. Counting only on-site cooling, OpenAI's Sam Altman has cited about 0.3 millilitres. Counting the water used to generate the electricity as well, academic estimates run far higher. If you need one figure, quote it with its boundary, because a bare number will be wrong for half the audience.
Is ChatGPT bad for the environment?
It has a real and non-trivial footprint in electricity, water and hardware, concentrated in training and in aggregate usage rather than in any individual query. Whether that is 'bad' depends on what it displaces and how the power is generated, which is a genuine policy argument rather than a settled fact.
Does using a smaller model reduce the impact?
Yes, and it is one of the few levers an individual actually controls. Serving a request from a smaller model consumes less compute than routing everything through the largest available one. Matching the model to the difficulty of the task is both cheaper and lighter.
Why do published numbers vary so much?
Mostly because of what is counted. Direct versus indirect water, one query versus a long response, inference alone versus a share of training, and one region's power mix versus another. Comparing two figures without checking these is comparing different quantities that happen to share a unit.