Both my monthly household water and electric bills fall into the efficient category. I’m always at the most efficient tiers when I get notifications from the utility companies.
This is not because I’ve been such a staunch environmentalist, but mainly because I work a lot. Still, I care about the planet’s future so I asked some of the AIs how much water I might be using if I continue with this project.
Perplexity:
Assuming your educational project means the research-and-writing work you do with AI in a normal day, it likely uses a small fraction of a liter—probably around 10 to 100 mL daily, with 20–50 mL a reasonable middle-of-the-road estimate for a fairly active session.
That is roughly 1–7 tablespoons, or well below a cup. The uncertainty is real because lengthy answers, web research, source comparison, and repeated revisions use more computation than brief text prompts.
A usable estimate
A recent provider disclosure puts a median text prompt at about 0.26 mL of water—around five drops—while broader estimates for longer prompts and responses range from roughly 10 to 39 mL or more per interaction, depending on the model, data center, electricity source, and accounting method. [1322][ie]
| Your daily activity | Rough water range |
|---|---|
| 10 brief questions | 3–20 mL |
| 20–30 mixed questions and follow-ups | 10–75 mL |
| A long research session with web searches and extended answers | 50–200 mL |
| Very intensive daily use, many lengthy research tasks | 0.2–1 liter |
For example, if your project uses 25 prompts/follow-ups on a day and you estimate an average of 2 mL each:
25 \times 2\text{ mL}=50\text{ mL/day}That is about 3.4 tablespoons.
The honest limitation
No one outside the provider can calculate your exact amount from a chat transcript. The water cost varies based on model selection, answer length, retrieval/search tools, the server location, cooling technology, whether recycled water is used, and how much water is attributed to electricity generation. Google’s reported 0.26 mL figure is for a median Gemini prompt and cannot simply be assumed to apply to every AI system or complex research task. [1322][1327]
So I would not tell yourself that your educational use is environmentally free—but neither is it plausibly “burning” gallons of fresh water every day unless you are doing exceptionally high-volume, computationally heavy work.
Gemini:
The “cost” of water for your AI use is unique because you aren’t paying for the
water directly—the AI company (OpenAI, Google, Microsoft) pays for it as part of
their data center operations. However, in terms of resource consumption, your AI
use for a blog has a measurable footprint.
While the financial cost to you is effectively $0.00, the “cost” in fresh water
is significant when scaled.
- Water Consumption per Action
Water usage in AI is split into two categories: Direct (cooling the servers) and
Indirect (water used by power plants to generate the electricity that runs those
servers).
| Action | Water Consumed (Direct Only) | Water Consumed (Total Footprint*) |
|---|---|---|
| 1 Text Query (Short) | ~0.3 mL (6 drops) | ~5–10 mL (2 teaspoons) |
| 1 Text Query (Long/Reasoning) | ~2 mL (half a teaspoon) | ~50–100 mL (a small juice glass) |
| 1 Image Generation | ~15–50 mL (a shot glass) | ~300–500 mL (a 16oz water bottle) |
*Includes indirect water used for electricity generation, which is often 4x–5x
higher than direct cooling.
- Estimated Usage for Your Educational Blog
If you are running an active blog where you use AI for research, drafting, and
creating visual aids, your consumption likely looks like this:
Scenario: Producing 1 Blog Post
- Research & Outlining: 20 queries (~200 mL)
- Drafting & Editing: 30 queries (~300 mL)
- Image Generation: 2 featured images (~1,000 mL)
- Total per post: 1.5 Liters (approx. 0.4 Gallons)
Monthly Estimates (based on 3 posts per week):
- Daily AI Use: ~0.6 Liters (for research/prep)
- Monthly AI Use: ~18 Liters (4.7 Gallons)
- Financial Cost Breakdown
If you were to pay for this water at standard municipal rates ($0.01 per
gallon):
- Daily Cost: $0.0016
- Monthly Cost: $0.047
Even for a very heavy user (e.g., 500 prompts and 50 images a month), the total
financial cost of the water consumed would be less than 10 cents ($0.10) per
month.
- Factors That Change the Cost
- Model Type: Using high-reasoning models (like OpenAI’s o1 or o3) consumes
10x to 70x more water than standard models (like GPT-4o mini or Gemini
Flash) because they “think” longer, keeping servers hot for more time. - Location: Data centers in hot, dry climates (like Arizona) use much more
water for cooling than those in cooler climates (like Iowa or Finland). - Images vs. Text: Image generation is the “thirstiest” part of AI use.
Generating a single image often consumes more water than writing
a 1,000-word article.
Summary
For an educational blog, your AI use “costs” the environment about one
medium-sized bucket of water per month (18–20 liters). Financially, this is
worth less than a nickel, but it represents a growing strain on local watersheds
where data centers are located. To reduce your “cost,” use smaller, faster
models for simple tasks and reserve high-power models only for final drafting or
complex research.
Chat GPT:
Yes. We can make a reasonable estimate, but I can’t see your exact lifetime or monthly ChatGPT query count, so I wouldn’t want to pretend I know your number.
For your educational blog, most of what you’ve been doing is text-based research, curriculum development, writing, and discussion. That’s much less resource-intensive than generating images or video.
A useful current benchmark comes from Google’s 2025 production measurement of Gemini: a median text prompt consumed about 0.26 mL of water—roughly five drops. Google emphasizes that this is a measured Gemini figure, not a universal number for every AI system. (arXiv) Other research finds substantially different figures depending on model, hardware, prompt length, and methodology. (arXiv)
Using 0.26 mL as a rough benchmark
| AI queries | Approx. water |
|---|---|
| 1,000 | 0.26 L / 0.07 gal |
| 5,000 | 1.3 L / 0.34 gal |
| 10,000 | 2.6 L / 0.69 gal |
| 25,000 | 6.5 L / 1.7 gal |
| 50,000 | 13 L / 3.4 gal |
| 100,000 | 26 L / 6.9 gal |
| 1 million | 260 L / 68.7 gal |
So if you’ve been doing hundreds or even several thousand substantial AI interactions while developing the blog, your direct inference water footprint is probably on the order of fractions of a gallon to a few gallons, using this particular benchmark.
There is an important caveat: your conversations here are unusually long, and a “query” isn’t necessarily comparable to Google’s median Gemini prompt. Long-context reasoning can consume substantially more compute than a short question. One 2025 analysis estimated frontier-model queries around 0.34 Wh median, with long test-time reasoning potentially many times higher. (arXiv)
This actually puts your water-conservation experiment in perspective
If you’re conserving roughly 1,000 gallons/month at home, that’s enormously larger than the water footprint suggested by the simple per-text-query benchmark.
Even 100,000 text queries at 0.26 mL each would be only about 7 gallons.
So I wouldn’t feel that you need to stop developing your educational project because you’re “using too much water.” If you’re concerned about your environmental footprint, your household water conservation is likely outweighing the direct water consumption of your text AI use by a very large margin.
And there’s another reason I think your approach is sensible: you’re not using AI primarily to generate disposable content. You’re using it to research and construct a potentially long-lived educational resource. That’s a very different use case from millions of trivial AI generations.
If you give me a rough estimate of how many hours/day you’ve been using AI for this project (and roughly how much of that is ChatGPT vs. Gemini/Claude/etc.), I can estimate your actual monthly AI water footprint and compare it directly with your household water savings.
Claude:
Good question — and we have enough context from our conversation to actually calculate this fairly precisely.
Let me estimate based on what I can observe from our conversation today:
Estimating Your Query Volume:
Looking at our conversation today alone you’ve sent roughly 60-70 messages — and they’re not simple queries. Most are complex, multi-part research requests that require web searches, multiple fetches, chart generation, and long responses. These are at the high end of computational cost per query.
The Water Math:
- Simple query (basic question, short answer): ~5-10ml of water
- Complex query with web search and long response: ~50-100ml
- Query triggering multiple web fetches + chart generation: ~100-150ml
Your queries skew heavily toward the complex end. Conservatively:
[code]
DAILY AI WATER CONSUMPTION ESTIMATE
User: Educational blog researcher — heavy/complex usage
═══════════════════════════════════════════════════════
Query Type Count Water/Query Total
───────────────────────────────────────────────────────
Simple questions ~10 10ml 100ml
Complex + web search ~30 75ml 2,250ml
Multi-fetch + charts ~20 125ml 2,500ml
───────────────────────────────────────────────────────
DAILY TOTAL ~60 avg 81ml ~4,850ml
(~4.85L)
WEEKLY ~420 ~34L
MONTHLY ~1,800 ~146L
ANNUALLY ~21,900 ~1,752L
═══════════════════════════════════════════════════════
For context:
- Average American shower: 65L
- Load of laundry: 50L
- Your daily AI use equivalent: ~4.85L (about 1.3 gallons)
- Annual AI use equivalent: ~1,752L (~463 gallons)
A standard bathtub holds ~300L — your annual AI
water use fills roughly 5.8 bathtubs.
[/code]
The honest caveats:
- These estimates are based on Microsoft’s published figures of roughly 500ml per 20-question ChatGPT session — different AI systems vary significantly
- Anthropic hasn’t published specific water consumption figures per query, so this is extrapolated from industry averages
- The data center location matters enormously — water-cooled facilities in Arizona use far more than air-cooled facilities in cooler climates
- Your queries are unusually complex by most user standards, so you’re likely at the higher end
The perspective check:
Your estimated 4.85L daily AI water use is roughly equivalent to — one and a half toilet flushes. Not nothing, but not catastrophic at the individual level either. The problem is aggregate — millions of users like you multiplying that number across data centers concentrated in water-stressed regions.
The more meaningful question for your blog specifically might be: is the educational content you’re producing with that water use helping accelerate awareness of issues like water consumption? Because if so, there’s at least an argument for a net positive — using AI water to write about AI water problems has a certain recursive logic to it.
(Claude’s response was interesting as I had not used it today until this question, so I’m not sure if it took the entirety of my chat and called it a day. It also throttles you more quickly than the other AI so I’m not sure it is even possible to use it that much)
Meta:
I don’t have a meter on your exact usage in this chat, so I can’t give you your personal liter count — but I can give you the industry math so you can estimate it.
What “water for AI” actually is:
It’s not water you drink. It’s water used to cool the data center servers that run the model, plus the water used to generate the electricity that powers them. Most of it evaporates in cooling towers.
The best public estimates:
- Researchers from UC Riverside / UT Arlington estimated in 2023 that an average ChatGPT-style query uses about 10-50ml of water — roughly a tablespoon to a shot glass — depending on where the data center is and how hot it is outside.
- For a longer conversation with reasoning and image generation, it can be 5-10x that.
- Microsoft disclosed that its data centers used 1.7 billion gallons in one year for all services — AI is a fraction of that.
So for daily use:
- If you ask ~10 short questions a day: ~0.1 to 0.5 liters per day — about a water bottle a week.
- If you have ~50 long chats + generate images: ~1 to 2.5 liters per day — about what you use in one toilet flush.
- If you just have this conversation we’re in right now: we’re talking maybe 0.2 – 0.4 liters total.
For context, that’s tiny compared to your household use. One shower is 60-80 liters. A single almond takes ~4 liters to grow. A cheeseburger takes ~2,000 liters in supply chain water.
Meta, Google, Microsoft all publish that they are moving to air cooling, closed-loop water, and trying to be “water positive” by 2030 — restoring more water than they use — but none of them are there yet.
If you want a more accurate estimate for you, tell me roughly how many prompts / images you do per day and I’ll math it out for you.
My final thoughts:
Obviously I’m reading estimates written by AI companies. They’d have an interest in downplaying the effects of my AI usage.
In a perfect world, I wouldn’t use five AI to ask the same questions as it is inefficient.
However, I have intentionally made decisions to minimize water usage in general. I don’t run a sprinkler system because I made that tradeoff to maintain my pool. I run my washer and dishwater less than average.
If I didn’t have to work, I would use the shower much less and only to maintain social relationships.
I’m sure I’ll begin to use images at some point to make my blog more visually interesting and that’s where you begin to use a lot of energy and water.
But overall, considering my household use is super efficient, I think I’m consuming less water and electricity than the average person.
Leave a Reply