Does ChatGPT use more water than Google?

The Viral Claim Everyone Gets Wrong: The 500ml Water Bottle Statistic

Ask ten sources how much water ChatGPT uses per query and you'll get answers spanning three orders of magnitude — from a fraction of a drop to a full bottle. The truth is that nobody, including the people making the claims, fully agrees.

The full disputed range looks like this:

• Sam Altman claimed in June 2025 that the average ChatGPT query uses just 0.3 ml of water
• A detailed independent analysis estimates roughly 5 ml per average conversation
• Researchers at Rochester Institute of Technology calculated 10–25 ml per prompt, with 20–50 prompts equaling one 500 ml bottle
• A UC Riverside estimate puts a single 100-word prompt at 519 ml — an entire bottle

That's a spread of more than 1,000x, which should tell you something about the state of this data. The most famous figure — the "one water bottle per conversation" stat that spread everywhere in 2023 — came from the paper "Making AI Less Thirsty." But a methodological critique from independent blogger Sean Goedecke argues the real figure is closer to 5 ml per conversation, roughly 100x lower, for two reasons.

First, the underlying power figure (0.004 kWh per request) was actually a per-page measurement, not a per-request one — a unit mismatch that inflated the result. Second, modern models are roughly 10x smaller and faster than GPT-3, the model the original calculation was based on. As Goedecke puts it, the 500 ml statistic "was wrong at the time" and is now badly out of date.

For context on the other end of the scale: Google disclosed in 2025 that each query to its Gemini assistant consumes 0.26 ml of water — about five drops. That figure, cited in IE Insights reporting, is a first-party corporate disclosure, and even it comes with caveats about what's counted.

Why does this matter beyond trivia? Because per-query numbers obscure the real story. Even at the most conservative estimate, AI-related data centers are projected to consume more than 1 trillion liters of water annually by 2028 — an elevenfold increase from 2024. And the per-query framing invites the same problem we see across AI tooling: confident, widely-cited claims that fall apart under independent verification. It's the reason SkillzBench exists — published research found roughly 26% of AI agent skills carry at least one vulnerability, and the lesson generalizes: don't trust the headline number, verify the methodology.

The honest takeaway: the bottle-per-conversation stat is contested and likely overstated by orders of magnitude, but the aggregate water footprint of AI is growing fast enough that the per-query debate is almost beside the point.

The Only Honest Comparison We Can Make: ChatGPT vs. Google Gemini

Here's the uncomfortable truth: nobody can tell you how much water a single Google Search drinks, because Google has never published that number — and researchers at Oxford note that Google historically treated its water usage as a trade secret. So the headline question can't be answered head-to-head. But there is one honest comparison we can make, and it's revealing.

In 2025, Google disclosed that each query to its Gemini AI assistant consumes about 0.26 ml of water — roughly five drops, according to reporting on the disclosure. That gives us the only apples-to-apples figure available: an AI assistant versus an AI assistant.

Put that against the ChatGPT estimates, and the picture is clear even at the most conservative end:

• Sam Altman claimed in June 2025 that the average ChatGPT query uses just 0.3 ml of water — still more than Gemini's 0.26 ml.
• An independent technical critique estimates ~5 ml per average ChatGPT conversation, arguing the viral "500 ml" figure was methodologically flawed.
• RIT researchers calculated 10–25 ml per prompt, meaning 20–50 questions drain a full 500 ml bottle.
• UC Riverside's widely-cited estimate reaches 519 ml for a single 100-word prompt.

Even if you take Altman's most favorable figure at face value, ChatGPT appears thirstier than Google's own AI assistant — and the gap widens dramatically under independent estimates. That part of the answer is defensible.

The part that isn't defensible is the popular framing: "ChatGPT uses X hundred times more water than a Google Search." No source provides a per-query water figure for traditional Search, so any such multiplier is invented. What we can say is that a chat query is roughly 200× more energy-intensive than basic text classification per UN researchers, which makes "probably yes, but unproven" the only honest verdict on AI-versus-Search.

There's a broader lesson here for anyone evaluating AI claims: widely-cited statistics can be off by two orders of magnitude, and the only defense is independent verification. It's the same principle SkillzBench is built on — benchmarking AI agent skills with and without the skill to measure real Lift, rather than trusting published claims. In a field where even water numbers swing from 0.3 ml to 519 ml, proven beats promised.

Skills, proven. Join the SkillzBench waitlist to see independent skill benchmarks the moment we launch.

The Hidden Water Cost Nobody Counts: Electricity, Not Cooling

Ask ten people about AI's water footprint and they'll quote the "500 ml per conversation" statistic. But the biggest water cost of AI isn't the water you can see evaporating off a cooling tower — it's the water hiding inside the electricity that powers the whole operation.

Researchers at Oxford and Imperial College London make the case bluntly: the largest source of data centre water usage is actually electricity generation, not on-site cooling. Power plants driven by steam turbines consume enormous volumes of water, and every kilowatt-hour a data centre draws carries that burden upstream. According to analysis from the Environmental and Energy Study Institute, the indirect water footprint from data center electricity reached roughly 211 billion gallons in 2023 — about 1.2 gallons per kWh.

That hidden footprint is why disclosed figures likely understate reality. Most companies report only on-site cooling water, excluding power generation and semiconductor manufacturing. The water footprint has three components — on-site cooling, power plant water use, and chip manufacturing — and the last two are routinely left out of corporate sustainability reports.

Geography compounds the problem. Nearly half of the world's 9,000+ data centers sit in regions of high water stress, meaning the water they draw competes directly with local communities and agriculture. Only 32 countries host AI-specialized data centres, with over 90% of that capacity concentrated in the US and China.

The aggregate trajectory is what turns per-query trivia into a systemic question:

• Morgan Stanley projects AI-related data centers could consume more than 1 trillion liters annually by 2028 — an elevenfold increase from 2024.
• A UN report projects a 2030 data center water footprint of 9.3 trillion liters, roughly equal to the annual domestic water needs of 1.3 billion people in Sub-Saharan Africa.
• A peer-reviewed study projects global data center water consumption could increase more than sevenfold by 2050 without mitigation.

There's a deeper lesson here about verification. A statistic can be repeated a thousand times and still be wrong by orders of magnitude — as the disputed 500 ml figure demonstrates. That's precisely why SkillzBench is building independent verification for AI agent skills: published research found roughly 26% of skills carry at least one vulnerability, and benchmarking every skill under controlled conditions beats trusting claims at face value. Whether the subject is water usage or agent skills, the principle holds: measure it yourself before you believe it.

If you want verified answers on which AI agent skills actually deliver, join the SkillzBench waitlist — skills, proven.

What's Actually Being Done: Zero-Water Cooling and the Rebound Effect

The good news is that the AI industry isn't ignoring its water problem — engineers are already deploying cooling technologies that could cut data center water use dramatically. The catch is that solving the engineering problem may not solve the actual problem.

Liquid and direct-to-chip cooling can reduce water use by up to 90% compared with traditional methods, according to industry analysis, and closed-loop systems can cut freshwater consumption by up to 70%, per policy research from EESI. Microsoft has gone further, announcing next-generation datacenters designed to consume zero water for cooling altogether.

These aren't lab experiments — they're being built into the next generation of AI infrastructure. And they matter, because the status quo is expensive: a 1 MW data centre using traditional cooling consumes roughly 26 million litres of water per year, and up to 43% of US data centre electricity goes to cooling, according to Oxford engineering researchers.

But here's where optimism meets reality. Prof. Kaveh Madani, Director of the UN's water institute, warns of a classic rebound effect: efficiency gains get absorbed by volume growth. "More efficient and affordable AI and energy mean more consumption of AI, making the overall footprint far bigger than what we save through efficiency gains," he noted in the UNU-INWEH report.

The math behind his concern is straightforward:

• Inference, not training, accounts for 80–90% of AI energy use once models deploy — so water demand scales directly with adoption.
• Morgan Stanley projects AI data center consumption above 1 trillion liters annually by 2028 — an elevenfold increase from 2024.
• A peer-reviewed study projects global data center water consumption could grow more than sevenfold by 2050 without mitigation.

This is the Jevons Paradox in action: when something becomes cheaper and more efficient, we simply do more of it. A 90% reduction in cooling water per query means little if query volume grows fiftyfold. ChatGPT already handles roughly 2.5 billion prompts per day, according to the UN analysis — and agent workloads, where AI tools run autonomously and chain many calls together, multiply that further.

There's also a verification gap here. Just as per-query water estimates vary from 0.3 ml to 519 ml depending on methodology, claims about "efficient" or "sustainable" AI infrastructure often go unmeasured and untested. It's the same trust problem SkillzBench is tackling in the agent-skills ecosystem — independent verification beats taking vendor claims at face value, whether the claim is a skill's Lift or a data center's water footprint.

So is the problem solvable? Technically, yes — the mitigation toolkit is real and improving. Behaviorally, it depends on whether efficiency outpaces growth. The honest answer: zero-water cooling is necessary but not sufficient, and the numbers will keep moving as AI adoption accelerates.

Why This Story Is Really About Trust: Verifying Claims in the AI Era

If this article has taught you anything, it should be skepticism. The most viral AI water statistic — the "500 ml per conversation" claim — was wrong by roughly 100x, and the "corrected" numbers still span three orders of magnitude. In the AI era, the real lesson isn't about water at all. It's about trust.

Consider the track record. A detailed critique by independent technologist Sean Goedecke traced the 500 ml figure to a methodological error: the underlying power estimate was per-page, not per-request, and modern models run roughly 10x smaller and faster than GPT-3. His revised estimate: about 5 ml per average conversation. Meanwhile, Sam Altman claimed in June 2025 that the average ChatGPT query uses just 0.3 ml — a figure nobody outside OpenAI can independently verify.

The opacity runs deeper than one statistic. Oxford researchers note that Google has historically considered its water usage a trade secret, and disclosed figures likely understate the full footprint by excluding electricity generation and chip manufacturing. When the people making the claims control the data, face-value trust is not a strategy.

This matters far beyond environmental trivia. If you're building with AI — deploying agents, installing skills, wiring up tools — you're consuming other people's claims constantly: "this skill is safe," "this benchmark proves it works," "this integration is reliable." The water story shows exactly how those claims can go wrong:

• Viral statistics can be off by orders of magnitude — the 500 ml claim versus the ~5 ml critique.
• Vendor self-reporting can't be verified — Altman's 0.3 ml figure has no independent confirmation.
• Corporations can simply refuse to disclose — as Google did with water data for years.
• Even researchers disagree wildly — per-prompt estimates range from 0.3 ml to 519 ml.

Agent skills face the same trust problem, amplified. These SKILL.md instruction packages are spreading across GitHub faster than anyone can evaluate them, and published research suggests roughly 26% carry at least one vulnerability. A skill's README is a claim. A star count is a claim. Neither is evidence.

That's the gap SkillzBench is being built to close. As an independent testing lab for AI agent skills, it's constructing a pipeline that security-scans every skill, pins each version to a commit SHA, and benchmarks skills in a sandbox — with and without the skill installed — so the Lift is measured, not asserted. Trust scores derive from security, quality, license, and maintenance signals, and money never touches the numbers.

The water debate eventually got its correction because someone did the math independently. AI claims deserve the same treatment. Skills, proven — join the SkillzBench waitlist to see what verification looks like when it arrives.