← Back to Blog

Is AI Why Our Water Is Running Out? — Research & Sources

February 27, 2026 #technology
Evergreen
Table of Contents

Is AI Why Our Water Is Running Out? — Research & Sources

Every claim. Every number. Every source.

By Cleve · February 2026

Someone in the comments asked for proof. So here it is — every data point from the video, fact-checked against peer-reviewed studies, government reports, and investigative journalism. If I got something wrong, this is where you’ll find it. If I got it right, this is where you’ll find that too.

Part 1: Claims About AI Water Use

”One AI prompt uses up to 50ml of water”

✅ VERIFIED — 50ml is the high end of peer-reviewed estimates

The real range is 0.26ml to 50ml per prompt depending on the model, data center, and whether you count indirect water from electricity generation. Sam Altman claims 0.3ml (direct only). Google’s Gemini report found 0.26ml (direct only). The UC Riverside study by Shaolei Ren found 10–50ml when including water used to generate electricity. The script says “up to 50ml” which is accurate — it’s the upper bound from the most comprehensive methodology.

Sources:

✅ VERIFIED — based on comparing typical ChatGPT prompts to standard searches

A standard Google search uses roughly 0.3ml of water. At 10ml per AI prompt (the conservative estimate), that’s already 30x. At 50ml, it’s over 150x. The 30x figure specifically compares a typical ChatGPT prompt to a standard Google search and is widely cited in academic literature.

Sources:

”By 2027, data centres could be withdrawing over 6 trillion litres a year”

✅ VERIFIED — cited by WEF, University of Illinois, and industry reports

The World Economic Forum reports AI adoption could result in 4.2 to 6.6 billion cubic metres of water withdrawal by 2027 (that’s 4.2–6.6 trillion litres). The University of Illinois confirms the same range. ESG Times reports the figure could exceed 6.4 trillion litres specifically for AI-focused facilities.

Sources:

“That’s more water than the entire country of Denmark”

✅ VERIFIED — actually understated

The WEF report says the projected withdrawal is “equivalent to four to six times the annual water withdrawal of Denmark.” So the claim “more than Denmark” is technically understated — it’s 4–6x Denmark.

Sources:

”80% of that just evaporates”

✅ VERIFIED — confirmed by multiple academic and government sources

The Environmental and Energy Study Institute (EESI) states: “Approximately 80% of the water withdrawn by data centers evaporates.” Shaolei Ren (UC Riverside), quoted in Bloomberg’s May 2025 investigation, confirms the same figure. The Environmental Law Institute’s January 2026 fact sheet also cites 80% evaporation / 20% discharge.

Sources:

Part 2: The Comparison Claims

”Agriculture uses 70% of all freshwater on earth”

✅ VERIFIED — one of the most established figures in water science

The UN World Water Development Report, the Food and Agriculture Organization (FAO), the World Bank, and FAO AQUASTAT all confirm that agriculture accounts for approximately 70% of global freshwater withdrawals. FAO AQUASTAT’s precise figure is 69%. This has been consistent across decades of measurement.

Sources:

”One burger costs 2,500 litres of water — that’s 50,000 AI prompts”

✅ VERIFIED — within the published range of 1,700–3,000 litres

Published estimates: World Economic Forum says ~2,000L (plain) to 3,000L (with toppings). UNEP says ~1,695L for a quarter-pounder. Water Footprint Network data yields ~2,808L with bun and toppings. 2,500L is a defensible middle estimate. The math: 2,500L ÷ 0.05L per prompt = 50,000 prompts.

Sources:

”A bowl of oatmeal? About 100 litres. 25 times less than a burger.”

✅ VERIFIED — based on Water Footprint Network data

Water Footprint Network data: processed oatmeal = 2,536 litres per kg. A typical bowl uses ~40g of dry oats. At 2,536 L/kg: 40g = ~101 litres. The 25x comparison: 2,500L (burger) ÷ 100L (oatmeal) = 25x. Both calculations check out.

Sources:

”The US has 16,000 golf courses, each using as much water as a data centre”

⚠️ MOSTLY VERIFIED — comparison is closest to mid-size data centres

The National Golf Foundation confirms nearly 16,000 US golf courses. Average golf course water use: ~312,000 gallons/day (multiple sources). Average data centre: ~300,000 gallons/day (Brookings Institution). Large/hyperscale data centres can use up to 5 million gallons/day. So golf courses match average-to-mid-size data centres, not the largest ones. The overall point — that golf courses collectively represent enormous water consumption that receives far less scrutiny — stands.

Sources:

Part 3: The Solution Claims

”Over 160 new AI data centres were built in drought zones in the US in just three years”

✅ VERIFIED — from Bloomberg’s investigative analysis

Bloomberg News (May 2025) conducted an analysis using data from World Resources Institute and DC Byte: “More than 160 new AI data centers have sprung up across the US in the past three years in places with high competition for scarce water resources… That’s a 70% increase from the prior three-year period.” Confirmed by Environmental Law Institute and TechRepublic.

Sources:

”Closed-loop cooling cuts water use by 99%”

✅ VERIFIED — specifically for next-gen fully closed-loop systems

Prime Data Centers states their closed-loop design “uses less than one percent of the water consumed by traditional adiabatic and evaporative cooling systems” — that’s a 99%+ reduction. Microsoft’s zero-water-evaporation design (deploying 2026–2027) eliminates evaporative water entirely. Important caveat: not all systems marketed as “closed-loop” achieve this. Older designs still have open cooling towers. The truly zero-water systems from Prime, Microsoft, and Oracle represent the new standard.

Sources:

FAQ: Common Questions & Pushback

These are the most common responses I’ve gotten in the comments, DMs, and replies. Here’s where I stand — with receipts.

”You’re just a tech bro defending AI companies.”

Rewatch the video. I said AI IS making it worse. I said the Johor protest is valid. I said companies choose cheaper cooling at the community’s expense. I’m not defending big tech — I’m saying we should regulate ALL water-wasting industries, including AI. Singling out data centres while ignoring the 70% of freshwater going to agriculture isn’t activism. It’s selective outrage.

”Your 50ml number is wrong! One ChatGPT email uses a whole bottle of water!”

The 500ml figure is for a 20–50 prompt conversation, not a single prompt. Even the study’s own authors clarify this. Per prompt, the range is 0.26ml (Google Gemini, direct only) to 50ml (UC Riverside, full lifecycle). I said “up to 50ml” — that’s the high end of per-prompt estimates. The real scandal isn’t which number is right. It’s that most AI companies refuse to disclose which number applies to their systems.

”Comparing AI to burgers is a false equivalence. We need food to survive.”

I’m not saying don’t eat. I’m showing scale. Agriculture uses 70% of all freshwater (UN/FAO). Data centres use roughly 1–2%. Even if you shut down every AI data centre tomorrow, you’ve barely moved the needle on global water scarcity. The UN declared water bankruptcy because of decades of industrial agriculture, deforestation, and climate change — not because of ChatGPT. Also: you don’t need a beef burger. A bowl of oatmeal uses 25x less water. The argument isn’t “AI vs. food.” It’s: why are we only outraged about the 1–2% and silent about the 70%?

”Closed-loop cooling is just greenwashing. It doesn’t really work.”

Partly fair. Many older “closed-loop” systems still have an open cooling tower on the outer loop that evaporates water. But the new generation is different. Prime Data Centers’ fully closed design uses less than 1% of water compared to traditional systems. Microsoft’s zero-water design (launching 2026–2027) eliminates evaporative cooling entirely. Oracle has deployed similar systems. The tech genuinely exists and is verified. The problem is most companies still choose evaporative cooling because it’s cheaper. That’s exactly why the video says we need to push for it — not trust companies to volunteer it.

”AI water use is growing exponentially. Stop minimising it!”

I never said AI isn’t a crisis. I said it’s not the only crisis. US data centres consumed about 17 billion gallons directly in 2023. US golf courses use over 500 billion gallons. Agricultural irrigation withdraws trillions. AI is growing fast — absolutely. But the water crisis predates AI by decades. Both things are true: AI is making it worse AND other industries have been draining water for far longer with far less scrutiny. Asking for equal accountability isn’t minimising. It’s being thorough.

”If data centres power everything (Netflix, banking, etc.), then isn’t all tech the problem?”

Yes — and that’s the point. Data centres existed long before AI. Every time you scroll Instagram, stream a show, or check your bank balance, you’re using one. AI has increased the load, but the infrastructure was already there. The question isn’t “should data centres exist” — it’s “should they be built in drought zones, using the cheapest cooling, with zero transparency?” The answer to that is no, regardless of whether the workload is AI or Netflix.

”You mentioned the Johor protest. What actually happened there?”

In February 2026, over 50 residents representing roughly 1,000 people across four housing estates in Gelang Patah, Johor protested a Zdata Technologies data centre being built less than 1km from their homes. Complaints included constant construction dust, concerns about water supply, and zero transparency from the developer about water usage. Johor state had previously paused approvals for Tier 1/2 data centres (which can use up to 50 million litres per day). As of mid-2025, Johor had approved RM164.45 billion in data centre investments.

Full Source List

Institutional & Government Sources

Academic & Research Sources

Investigative Journalism

Industry & Corporate Sources

Food & Agriculture Water Footprint

Critical / Counter Perspectives (included for balance)

If you think I got something wrong, tell me. I’ll update this document.

The goal isn’t to be right. It’s to be accurate.