Is ChatGPT Bad for the Environment?

Is ChatGPT Bad for the Environment?

The question gets asked in two very different registers. One is personal: should I feel bad about typing into this box? The other is structural: is the industry building this thing responsibly?

They have almost opposite answers, and conflating them is why the public argument goes nowhere. Per prompt, the energy cost is small enough that your individual usage is close to a rounding error against the rest of your day. In aggregate, the electricity demand behind these systems is the fastest-growing load on several national grids, and the companies involved are spending more on it each year than most countries spend on their own power systems.

What follows is what has actually been published, by whom, with what left out — and what to do with it.

The Short Answer

An average ChatGPT query uses roughly a third of a watt-hour of electricity, according to the only figure OpenAI has ever put in public. That is about what an oven draws in a second, or a few minutes of an efficient lightbulb.

At that scale your personal prompting is not the issue. A hundred queries a day for a year is a few kilowatt-hours — less than a single load of laundry a week.

The aggregate picture is the one worth arguing about. Data centre electricity demand rose 17% in 2025 while global electricity demand rose 3%, and the International Energy Agency expects data centre consumption to double by 2030, with power use from AI-focused facilities tripling. Whether that is "bad for the environment" depends almost entirely on what is generating the electricity, which is a policy question rather than a prompting one.

Your site has an impact you can actually measure. Get a free SEO audit and see every issue ranked by what it costs you.

What OpenAI Has Actually Published

The one number, and where it came from

In June 2025, Sam Altman wrote a blog post called "The Gentle Singularity" containing a single sentence of disclosure: the average query uses about 0.34 watt-hours of electricity and about 0.000085 gallons of water. He compared the water figure to roughly one-fifteenth of a teaspoon.

That is it. That sentence is the entirety of OpenAI's public per-query environmental reporting, and almost every article you will read on this subject is ultimately quoting it.

What the figure leaves out

It has not been peer reviewed and no methodology was published alongside it. Nobody outside OpenAI knows what counts as an "average query" — whether deep research requests and long document analyses are in the average or excluded from it.

That distinction is not a quibble. A short factual answer and a twenty-minute research run differ by orders of magnitude in compute, so an unqualified average can be made to say almost anything depending on which requests are in the denominator.

It also excludes training, which is the largest single energy event in a model's life, and it says nothing about which data centres the figure describes. ChatGPT is served from multiple facilities that differ in age, cooling system, climate and grid mix, and the hardware mix changes as newer accelerators arrive.

What OpenAI has not published

There is no per-query emissions figure. No grams of CO2 equivalent, no breakdown between compute and overhead, no statement of whether the electricity is accounted market-based or location-based.

That absence is itself informative. A company confident in its emissions number tends to publish it.

What Google Published Instead, and Why It Matters

The median Gemini prompt

In August 2025, Google released a technical paper with figures for a median Gemini text prompt as of May 2025: 0.24 watt-hours of energy, 0.03 grams of CO2 equivalent, and 0.26 millilitres of water.

The headline numbers land in the same neighbourhood as OpenAI's. The difference is that Google showed its work.

What is inside the boundary

Google's figure counts metered power for the TPUs and GPUs doing the inference, plus host CPU and memory, plus the idle capacity the company keeps provisioned for reliability and low latency, plus campus-level power usage effectiveness for data centre overhead, plus amortised embodied emissions from the hardware itself.

Its own breakdown puts roughly 58% of the energy in the accelerators, about 24% in CPU and memory, and about 18% in redundancy and the facility. That is the sort of detail that lets an outsider check an answer rather than take it on trust.

The redundancy line is the quietly interesting one. Nearly a fifth of the energy is spent on capacity that is sitting idle so that your prompt answers quickly, which means some of the footprint is a cost of responsiveness rather than of computation.

Where researchers push back

The criticisms are substantive and worth knowing. Google reported market-based emissions, which credit clean-energy purchases and can obscure what a specific local grid actually burned; location-based figures would be more comparable. The water figure counts on-site consumption only, leaving out the water used to generate the electricity, which is a large share of the total.

The paper also reports a median rather than a mean, on the reasonable grounds that the mean is dominated by extreme prompts — but that means the long, expensive requests are precisely the ones the headline number hides. And like OpenAI's figure, it excludes training, data storage, networking and end-user devices, and covers only text prompts in the Gemini apps.

One audit, every fix, ranked by impact. Get your free SEO audit and stop guessing which change matters most.

The Number That Actually Moves

How fast demand is growing

The per-prompt figures are small and the aggregate is not, because the number of prompts is enormous and the infrastructure behind them is being built at a pace that has no recent precedent.

The IEA's assessment is the one to anchor on: data centre electricity demand up 17% in 2025 against 3% growth in global electricity demand overall, a doubling of data centre consumption expected by 2030, and a tripling of the AI-focused slice.

Where the money is going

Capital expenditure by five large technology companies exceeded $400 billion in 2025, and was expected to rise a further 75% in 2026. Conditional offtake agreements between data centres and small modular reactor projects grew from about 25 gigawatts at the end of 2024 to roughly 45 gigawatts.

Those are not numbers you can offset with careful prompting. They describe a construction programme.

That matters for how you read any per-prompt figure. Most of the environmental decision has already been taken by the time a request reaches a server: the facility exists, the hardware is installed, the grid connection is contracted, and the idle capacity is powered whether or not you type anything.

Why the grid, not the prompt, is the variable

A watt-hour of electricity in a region running on hydro and nuclear carries a fraction of the emissions of the same watt-hour on a coal-heavy grid. The same prompt can differ by an order of magnitude in carbon terms depending on where and when it was served.

That is why the useful questions are about siting, grid mix, water-stressed regions and whether new demand is met with new clean generation or by keeping existing fossil plants online. Technology companies accounted for roughly 40% of corporate renewable power purchase agreements signed in 2025, which is real — and also not the same thing as the new demand being carbon-free at the moment it is drawn.

How to Think About Your Own Use

Per-prompt maths is mostly a trap

If you do the arithmetic on 0.34 watt-hours, you will conclude that your usage is negligible, and you will be right. You will also have answered a question that does not matter much.

Individual restraint does not change a data centre buildout driven by enterprise contracts and model training. Treating it as the main lever is the same error as framing climate policy around personal recycling habits.

There is a second reason the arithmetic misleads. Per-query figures are medians or unspecified averages, so multiplying one by your own usage assumes your prompts look like the typical prompt — and if you lean on long-context work or reasoning modes, they do not.

The choices that do matter

If you want your use of these tools to be defensible, three things carry more weight than prompt volume:

• Pick the smallest model that does the job. Reasoning-heavy modes and long-context requests cost substantially more compute than a short answer from a small model. This is the one per-query choice with real leverage.
• Do not run things twice. Batch work, cache results, and stop re-generating the same output. Automated pipelines that re-run on every page load are where real waste accumulates.
• Ask vendors for location-based numbers. If you buy AI at organisational scale, procurement questions about grid mix and water stress move more than anything you do in the chat box.

Measured once, fixed once. A free SEO audit finds what is quietly costing you traffic.

What the Honest Uncertainties Are

Training is missing from every per-prompt figure

Both published figures are inference only. Training a frontier model is an enormous, concentrated energy event, and amortising it across subsequent queries would raise the per-query number — by how much, nobody outside the labs can say, because neither the training energy nor the query count is public.

Accounting choices flatter everyone

Market-based emissions accounting is standard practice and also the most generous available framing. Until vendors publish location-based figures alongside, cross-company comparisons are not really comparisons.

Water is a separate question

Water has its own accounting problems, its own geography and its own viral claims, and it does not reduce to the energy figure. We covered that in detail in our guide to how much water ChatGPT uses, including where the widely repeated bottle-of-water comparison came from.

Hardware and e-waste barely get measured

Embodied emissions from manufacturing accelerators appear in Google's figure as an amortised component and in nobody else's at all. The disposal end of that cycle is largely undocumented.

How ChatGPT Compares to the Alternatives

• Gemini is the only major assistant whose vendor has published a methodology you can argue with. That is a meaningful transparency advantage, independent of whether the numbers are flattering.
• ChatGPT has one unverified sentence of disclosure and no emissions figure. Its per-query energy claim is in the same range as Google's, but you are taking it on trust.
• Claude, Perplexity and Copilot have published no comparable per-query figures at all, so an absence of numbers should not be read as a lower footprint.
• A plain web search is roughly an order of magnitude cheaper per query than an AI answer on most published estimates — which matters mainly because AI answers are now appended to searches whether or not you asked for one.

Want the whole picture, not one metric? Start with a free SEO audit — one pass, every issue listed.

Final Thoughts

Is ChatGPT bad for the environment? As a personal habit, no, not in any way that should change your behaviour. As an industry, the answer is unresolved and depends on decisions being made right now about where data centres get built and what generates their electricity.

The thing worth being annoyed about is the state of the disclosure. One company has published a sentence; another has published a paper with real gaps in it; most have published nothing. Every confident figure you see circulating traces back to one of those two sources, usually without the caveats.

So be sceptical of precise claims in both directions, treat the aggregate growth numbers as the serious part of the story, and judge vendors on what they are willing to show you rather than on how good the headline number sounds.