I Use AI 10 Hours a Day. Here's What I Actually Found About Water.

"You know how much water that uses?"

Lecture incoming. (I hate those.)

I spent five years building an earth-and-straw house, fully off-grid, autonomous water system, solar panels on the roof. Real choices. The kind that cost weekends, savings, and more arguments with building inspectors than I care to remember. So when someone who eats four croissants for breakfast decides to lecture me about my Claude queries, I do one thing: I check the numbers.

That's what I did.

What I found made me more annoyed. Not for the reasons you think.

TL;DR: Your AI usage for an entire year consumes less water than the jeans you're wearing right now. The real problem exists, but it doesn't target you the way they're framing it. Here's how to read this debate without getting played from either side.

AI's Thirst: When Productivity Meets Hydration Hysteria

The Number Everyone Quotes Is Wrong

The "500ml per query" claim. You've seen it. Reddit threads, LinkedIn posts, conference talks, your colleague who just discovered environmental impact statistics.

It comes from a 2023 paper by Shaolei Ren at UC Riverside, Making AI Less Thirsty. Good paper, genuinely useful. But the number that went viral isn't what the paper actually says.

The study measured 500ml per 20 to 50 questions, for GPT-3, running on Microsoft Azure data centers in 2022, using evaporative cooling. That's 10 to 25ml per query at most. Not per conversation. Per query. On hardware that's three generations old, in facilities that have since been significantly upgraded.

Sam Altman stated in February 2026 that a typical ChatGPT query uses 0.3ml. An independent analyst working through the same methodology with current models landed around 5ml as a realistic figure. Actual number sits somewhere in that range, depending on where your query gets processed and what cooling system the data center uses.

That variance isn't a footnote. A data center in Iceland using free air cooling consumes close to zero water. The same query processed in an Arizona facility using evaporative cooling during summer consumes orders of magnitude more. Same model, same prompt, completely different footprint. The number without the geography is meaningless.

One honest caveat: the big labs don't publish granular per-query water data. OpenAI, Google, Microsoft all report aggregate figures with enough ambiguity that independent researchers have to estimate. So yes, we're working with approximations. The 0.3ml to 25ml range is real, and anyone claiming precision beyond that is working from models, not measurements. Including the people making the 500ml claim.

The viral figure is an extrapolation from a 2022 study applied to 2026 models. That's not science. That's a meme.

Let Me Do the Math on My Own Usage

I run the full infrastructure behind my daily AI workflow on a $5 VPS. Multiple models, multiple sessions, Claude Code running most of the day. Ten hours of active AI usage, most days, for months.

Let's take the pessimistic estimate and be generous with it: 25ml per query. Assume I average one query every six minutes during active sessions. That's 10 queries per hour, 100 per day, roughly 36,500 per year.

36,500 queries times 25ml: 912 liters per year.

So let's compare it to things nobody lectures me about.

1 pair of jeans: 10,000 L

1 beef steak (200g): 3,000 L

1 cotton t-shirt: 2,700 L

your AI habit, 1 year: 912 L

1 cup of coffee: 140 L

1 load of laundry: 72 L

Water usage breakdown showing AI queries versus everyday items and activities

At 25ml per query, a full year of heavy AI usage lands around 900 liters. Less than a pair of jeans. At the more realistic 5ml estimate, it's under 200 liters. A few loads of laundry. At Altman's disclosed 0.3ml figure, it's 11 liters. Two toilet flushes.

One pair of jeans costs 11 times more water than a year of your AI habit.

The croissant guy who lectured me? His breakfast routine (four croissants, butter, coffee, orange juice) clocks somewhere around 600 liters of embedded water. Per morning. That's before we count whatever he's wearing.

Full disclosure: I'm not a neutral party here. I build with AI every day, I run infrastructure that depends on these models staying available and affordable, and I have an obvious interest in this not being a problem. I'm telling you that upfront because I think the numbers hold anyway, and I'd rather you check them yourself than take my word for it.

The Real Problem Is Real, Just Not Where They're Pointing

The individual footprint is negligible. The systemic footprint is not. Both are true, and mixing them up is either confused or convenient.

In 2023, Microsoft disclosed that 41% of its water withdrawals came from areas with high water stress. Google's data center in Council Bluffs, Iowa consumed 1 billion gallons that year, enough to supply all of Iowa's residential water for five days. Amazon didn't disclose comparable figures, which tells you something. A January 2026 report projected AI-related water consumption will grow by roughly 130% through 2050.

The Arizona example gets cited constantly. In Maricopa County: golf courses consume 29 billion gallons per year. Data centers consume 905 million gallons. Data centers use 32 times less water while generating 50 times more tax revenue per gallon. Altman leaned on this comparison, Steinberger made it in a podcast, and it's factually correct.

But also rhetorically useless. "Golf courses also waste water" doesn't fix the decision to build exascale training infrastructure in Nevada during a multi-decade drought. Both can be problems. Pointing at one doesn't dissolve the other.

Zero-water data centers exist. Microsoft deployed them in 2024. Immersion cooling, dry cooling, liquid-cooled chips... these aren't research projects. The problem isn't unsolvable. It's a priority decision that hyperscalers make when governments require them to, and not before. That's a policy problem. Folding it into a user behavior lecture is a category error.

The water argument lands on your plate because individual guilt is cheaper to produce than systemic accountability. "Use less AI" requires no lobbying, no regulation, no corporate pressure. Clean, portable, shareable. It also accomplishes close to nothing structurally.

The problem is real. Just not about you.

Nobody Knows What Comes Next (That's Always Been True)

The disruption is real. The pain is real. And it deserves to be said clearly before the "but productivity" argument shows up.

People in creative industries, paralegal work, customer support, code review... they're absorbing real, immediate pain right now. Not theoretical future pain. The Silicon Valley tendency to lead with long-term gains while people are losing their livelihoods is a genuine failure of empathy. The people building these tools owe some humility to that reality, not just charts about GDP projections.

But on the uncertainty about what comes next, let's be precise.

Nobody saw the asteroid coming. The hand-weavers didn't see the mechanical loom. The typographers didn't see desktop publishing. Travel agents didn't see the internet. Every major technological transition in history was navigated without a reliable map, by people who had no idea how it ended.

What's new about AI isn't the uncertainty. It's the speed.

Since we're being honest about the past: would you actually go back? Not rhetorically. Concretely.

Life a hundred years ago: your child dies from a fever at age three because antibiotics won't exist for another decade. You crack a tooth... no real dentistry, just extraction and pain, and if it gets infected you might die from it. You get a cancer diagnosis that today means six weeks of treatment and a normal life after. In 1925 it means dead in six months because the tools don't exist yet. If you're a woman, you can't vote in most countries and your legal personhood is largely conditional on your husband's.

Every generation before us absorbed brutal transitions. The people who navigated them weren't the ones who predicted the outcome correctly (nobody did). They were the ones who stayed useful while the map was being redrawn.

The uncertainty isn't new. The speed is.

Experts predicting AI catastrophe with total confidence have the same track record as experts predicting AI utopia with total confidence. Zero for zero. The ones worth reading are the ones who say "I don't know, but here's how I'm positioning."

What you can control: staying technical, staying curious, keeping your hands in actual work. Not because it guarantees anything. Because it's the only position from which adaptation is possible, whatever actually happens. 🤷‍♂️

Nobody saw the asteroid coming either.

Build anyway.