Does AI drink our water?

The short version

• Per use, almost nothing. Google measured a median Gemini text prompt at 0.26 millilitres of cooling water — about five drops. Verified
• At scale, real and rising fast. Google's own water consumption reached 10.9 billion gallons (41 billion litres) in 2025, up 34% in one year, with data centers driving the increase. Verified
• The damage is local, not global. Bloomberg found that about two-thirds of US data centers built since 2022 sit in places already short of water. Modeled A litre taken from a dry aquifer in August is not the same as a litre taken from a rainy river in December.
• In our region the pressure points are Johor and Batam — not Bali. Johor told investors in November 2025 to wait until mid-2027 for water; Batam's planned data centers would take around 8% of the island's supply. Bali has no hyperscale data center; its water crisis is driven by tourism, which uses about 65% of the island's water.
• The "bottle of water per email" claim has been withdrawn by its own author, who now puts a GPT-4-class prompt at about 15 mL including power-plant water — 35 times lower. Modeled
• Running AI on home PCs would make things worse, not better (section 8).
• For scale: every data center on Earth uses roughly a tenth of what golf courses do, and about 1/4,000th of what meat and dairy do — though most farm water is rain on fields, while data-center water is pumped and treated (section 9).
• The chemicals in cooling discharge are real; the fish kills are not documented. Nobody has properly measured data-center effluent anywhere (section 10).

1. How a data center uses water

A data center is a building full of computers that turn electricity into heat. That heat has to leave the building, and there are two ways to do it: blow air through radiators (costs electricity) or evaporate water (costs water). Almost everything about this topic comes down to that trade-off — plus a second water stream most people never see: the water evaporated at the power plants that make the electricity.

Direct water is what the building itself uses. In a cooling tower, water evaporates to carry heat away; operators report that 45–60% of the water they withdraw is consumed this way, while the rest goes back to the sewer warmer and saltier. Modeled Air-cooled designs use no water but about 10% more electricity. The newest AI servers need liquid piped directly to the chips, but that liquid runs in a sealed loop — it is not consumed. What matters is what sits at the end of the loop: a cooling tower (water) or a dry cooler (electricity).

Indirect water is the bigger, hidden stream. Coal, gas and nuclear plants boil water to make steam and cool it with more water; hydro reservoirs lose water to evaporation; wind and solar use almost none. Per kilowatt-hour, this ranges from roughly zero for wind and solar to about 1.8 litres for the average US grid, with coal and nuclear at the high end. Modeled For US data centers, Lawrence Berkeley National Laboratory puts the indirect footprint at nearly 800 billion litres in 2023, against 66 billion litres used on site — twelve times larger. Modeled

This matters for us: Indonesia's electricity is dominated by coal, so a query answered on an Indonesian grid carries more hidden water than the same query on a cleaner grid. And because tropical air is humid, evaporative cooling works poorly here, which pushes operators towards chillers — less water on site, more electricity, more indirect water.

The efficiency yardstick is WUE (water usage effectiveness): litres evaporated per kilowatt-hour of computing. The spread between operators is about tenfold (figure 4 below).

2. How much water, really?

The per-action numbers are small and — unusually for this debate — measured. Google's August 2025 paper reports 0.24 watt-hours and 0.26 mL of water for a median Gemini text prompt, counting idle machines and cooling overhead. Verified OpenAI's Sam Altman wrote that an average ChatGPT query uses 0.34 Wh and 0.000085 gallons (0.32 mL), without publishing a method. Modeled Mistral's peer-reviewed lifecycle study, which includes power-plant water, found 45 mL per 400-token answer. Modeled

So where did "a bottle of water per email" come from? A 2023 University of California Riverside study estimated that GPT-3 "drank" 500 mL per 10–50 medium responses, including power-plant water. A 2024 Washington Post piece turned that into 519 mL for one 100-word email. The study's lead author, Shaolei Ren, now says that figure is outdated and puts a GPT-4-class prompt at roughly 15 mL, of which about 5 mL is on-site cooling. The bottle is gone; the drops remain.

Per-prompt numbers are tiny, but there are billions of prompts. Google's water consumption rose from 6.4 billion gallons in 2023 to 8.1 billion in 2024 and 10.9 billion in 2025. Verified Amazon published an absolute figure for the first time in June 2026: 2.5 billion gallons withdrawn in 2025. Microsoft reports a fleet WUE of 0.27 L/kWh and says it became "water positive" in 2025, a claim repeated in its annual report to the SEC. Meta's last confirmed consumption figure is 813 million gallons for 2023; the "1.6 billion gallons" sometimes quoted for 2024 is water Meta restored through projects, not water it used. The AI-native companies — xAI, OpenAI, Anthropic — publish no facility water data at all.

Nationally and globally the totals are modest. The IEA estimates that data centers worldwide accounted for about 560 billion litres in 2023, heading towards 1,200 billion by 2030, including power-plant water. Modeled For perspective, agriculture takes roughly 70% of all freshwater withdrawn on Earth, and one kilogram of beef embodies about 15,400 litres. The IEA's rule of thumb is more useful than the global total: a 100-megawatt data center uses about 2 million litres a day, the water of 6,500 households. Whether that is a problem depends entirely on where those 2 million litres come from.

3. "But the evaporated water comes back as rain, doesn't it?"

Yes — and no. Water is not destroyed; it re-enters the water cycle. But it usually rains down hundreds of kilometres away or over the sea, weeks later, and the aquifer it was pumped from refills over years to centuries. That is why hydrologists count evaporation as "consumptive use": the water is gone from that basin in that season, which is exactly when a dry region also needs it most. Much of it is also treated drinking water, and the part that does return is warmer and saltier.

Both sides of the public argument are partly right. Critics who say the AI water issue is fake are right that nationally and per person the volumes are trivial and water is renewable. Critics who say it is a serious problem are right locally: roughly two-thirds of new US data centers since 2022 sit in water-stressed regions. AI is not a global water emergency; specific data centers in specific basins are a real local problem.

4. Where it actually hurts

Follow the conflicts and you find the same pattern: a big facility, a dry place, a drinking-water source.

• United States. In Virginia, Arizona and Texas, 72% of the new data centers in water-stressed areas were built. Loudoun County, Virginia — the densest data-center cluster on Earth — saw its facilities use about 10% of county water in 2023. In Memphis, xAI's Colossus supercomputer draws drinking water from the Memphis Sand aquifer; figures range from under 1% of city pumping today to 5.7 million gallons a day at full build-out, and an $80 million recycling plant was paused in 2026 with construction promised for 2027.
• Chile and Uruguay. Google's planned Santiago center would have needed 7.6 million litres of drinking water a day; an environmental court partly reversed the permit in 2024 and Google redesigned it for air cooling. In Uruguay, during the worst drought in 70 years, Google shrank its planned center to a third of the size and switched to air cooling after protests.
• Spain. In drought-hit Aragón, Amazon's three planned centers are licensed for 755,720 cubic metres of water a year — enough to irrigate 233 hectares of corn, and Amazon asked for 48% more water for its existing sites. The same investigation counted 62 Amazon, Microsoft and Google data centers in water-scarce regions, from India and Chile to the Gulf states.
• Who is most exposed. Bloomberg notes that China and India have even larger shares of data centers in high-stress areas than the US; Saudi Arabia and the UAE are building in extreme water stress. A February 2026 paper in AGU Advances concludes that data centers remain among the least transparent industrial water users, which is why local fights so often start with a public-records lawsuit.

5. Southeast Asia: Singapore's moratorium, Johor's boom, Batam's reservoirs — and Bali

Singapore froze new data centers from 2019 to 2022 because of electricity and water limits, then reopened with conditions — a Green Data Centre Roadmap with efficiency requirements and, in 2025–26, capacity awards tied to green energy. The freeze pushed the boom across the strait.

Johor, Malaysia is now the regional epicentre, with capacity approaching 6 gigawatts. In 2024, 101 applications asked for more than 808 million litres of water a day; only 45 were approved, for 142 million litres — against a state production of 2,352 million litres a day (the same rivers supply Singapore under the 1962 agreement). In November 2025, amid drought, the state tightened approvals and stopped approving the thirstiest water-cooled designs, asking investors to wait roughly 18 months. Operators are switching to reclaimed water.

Batam, Indonesia is the flashpoint closest to home. The island depends on rain-fed reservoirs that supply about 382 million litres a day, and it has rationed water before. Nine data centers planned for Nongsa Digital Park would need about 29 million litres a day by 2032; with one more facility at Kabil, around 8% of today's supply. The Jakarta Post counted at least 16 data centers at various stages in August 2026, and a Batam authority official has said openly that development will have to be limited because "they consume too much water, while residents are still struggling to get enough". In December 2024, villagers in Teluk Mata Ikan protested when water flowed to the economic zone while their taps ran dry.

Jakarta hosts most of Indonesia's data centers — around 185 of them — and is backed by big commitments: Microsoft's US$1.7 billion and a 1.2-gigawatt power deal for BDx, the largest in the country. Jakarta's problem is groundwater pumping and sinking land rather than cooling towers, but every new gigawatt on a coal-heavy grid adds indirect water somewhere in Java.

6. AI data centers versus the "ordinary" cloud

The buildings that run Google Search, WhatsApp or Netflix and the ones that train and run ChatGPT, Gemini or Claude are increasingly the same buildings — but the AI halls are different in degree. A conventional rack of servers draws 10–20 kilowatts; the latest AI racks draw over 100 and require liquid cooling to the chip. Campuses are planned at 1–5 gigawatts rather than tens of megawatts. Training runs flat-out for months; answering queries runs around the clock.

Two trends are true at once. Water intensity is falling: Amazon reports 0.12 L/kWh, about 90% of its sites on free-air cooling; Microsoft's new designs evaporate no water at all, saving 125 million litres a year per site, with the first ones opening in 2026. Verified But absolute volumes are rising because the build-out is so large: LBNL projects US hyperscale direct water alone at 60–124 billion litres by 2028, roughly one to two times today's entire sector. Modeled

Disclosure is where AI-native builders fall short. Google, Microsoft, Meta and now Amazon publish totals and WUE. OpenAI's Stargate sites and Anthropic's Project Rainier (built with Amazon) claim closed-loop cooling but publish no facility figures; Anthropic does not operate its own data centers, so its footprint sits inside Amazon's and Google's numbers. xAI's Memphis figures come from utility contracts and activists' records requests, not from the company. Meta's 5-gigawatt Hyperion campus in Louisiana holds a permit for up to 23 million gallons a day from six wells, though Meta says actual use will be far lower.

Is AI the reason the numbers are climbing? Google attributes most of its 34% jump to data centers during the AI build-out, and the IEA reports that data-center electricity use surged 17% in 2025 and is set to double by 2030 — and electricity is where most of the water is. Modeled

7. Which app uses the most water?

Nobody knows precisely, because no company reports water per product. What we have are company totals (known) and per-query figures for two AI products (known or claimed). Everything else is arithmetic with stated assumptions.

The honest ranking: Google as a company is by far the largest disclosed water user, and within it Search — simply because of its volume — likely outweighs Gemini today. ChatGPT is the single AI product most likely to rival Google Search, with a range wide enough to sit either side of it. WhatsApp, Telegram and other text apps are near the bottom: text is cheap, video and AI are not.

8. Would running AI at home on a powerful PC help?

No. Three things work against it. First, efficiency: a hyperscale data center runs at a power overhead of around 10% (Google's fleet PUE is about 1.09) and serves thousands of requests per chip at once; a home GPU serves one person and idles the rest of the day. Google's own paper notes that consumer hardware lacks this batching advantage, so energy per answer is higher. Our illustrative estimate: a 400-word answer from a small 8-billion-parameter model on a gaming PC drawing ~500 W takes 1–3 Wh; a 70-billion-parameter model 7–20 Wh — versus 0.24 Wh for Gemini. Estimated

Second, water: a home PC evaporates nothing, but every watt-hour still consumed water at the power plant — and on Indonesia's coal-heavy grid that is around 2 litres per kWh, so a 70B-model answer at home carries roughly 15–40 mL of power-plant water, ten to thirty times a cloud prompt. Third, the hardware: hundreds of millions of extra graphics cards would have to be manufactured, and chip fabs are among the thirstiest factories on Earth — the IEA counts hardware manufacturing as a distinct part of the data-center water footprint. Local AI has real merits (privacy, offline use); saving water is not one of them.

9. Put it in proportion: golf, beef, and every data center on Earth

A recurring objection in these debates is that the AI water story is being blown out of proportion while much thirstier things go unmentioned. That objection is largely correct on the arithmetic — and it still doesn't settle the argument, for a reason worth understanding.

Golf. American golf courses applied about 1.68 million acre-feet of water in 2020 across 14,145 facilities — roughly 2.07 billion cubic metres, a 29% reduction since 2005. Verified That is about 31 times the cooling water used by every data center in the United States (66 billion litres, or 0.066 billion m³) and still around 2.4 times their total including power-plant water. Globally there are roughly 38,900 golf courses; scaling the US average gives an estimated 4–6 billion m³ a year, close to ten times the 0.56 billion m³ the IEA attributes to all the world's data centers. Estimated

Meat. The comparison gets starker. The standard reference for animal agriculture, Mekonnen and Hoekstra (2012), puts the global water footprint of animal production at 2,422 billion cubic metres a year — about 4,000 times all data-center water — with 98% of it going to grow feed, and beef alone taking roughly a third. Modeled A single 150-gram beef burger embodies around 2,300 litres of water, which at Google's measured rate is about 8.9 million Gemini prompts. Estimated

10. Do data centers pollute the water they return?

A separate worry, raised often and rarely examined: that data centers discharge chlorine, copper, zinc, nitrates, phosphates, PFAS and glycol into rivers, and that warm effluent feeds algae and kills fish. Here is what stands up.

The starting point is blowdown. In an evaporative cooling tower, pure water leaves as vapour and the dissolved minerals stay behind, concentrating with each cycle. Periodically that concentrated water is dumped and replaced — that discharge is blowdown, and it is saltier, warmer and more chemically loaded than the water that came in.

The most important finding in this section is an absence. Despite the volume of public concern, there is essentially no peer-reviewed assessment of data-center effluent quality or its ecological effects anywhere, and no documented data-center river fish kill. That vacuum is not proof of safety — it is proof that almost nobody has looked, which is exactly what the authors of a February 2026 paper in AGU Advances argue when they call data centers among the least transparent industrial water users.

One structural point cuts through all of it: this entire category of problem belongs to evaporative cooling. Closed-loop and zero-water designs have no blowdown to discharge, so as those designs spread the discharge question shrinks on its own.

11. Quick checks on claims going around

12. What actually reduces the harm

1. Where it is built. Siting matters more than any per-prompt number. A cooling tower in a rainy, humid place is a rounding error; the same tower on Batam's reservoirs or Memphis's aquifer is a community fight. IEEE Spectrum's rule: in water-stressed regions use low-water cooling plus renewables; in wet regions with dirty grids, evaporative cooling may be the better trade.
2. What powers it. Because indirect water dominates, a cleaner grid cuts the water footprint faster than any cooling redesign. For Indonesia, this is the lever.
3. How it is cooled. Closed-loop and zero-evaporation designs (Microsoft), reclaimed water (Johor, Singapore's NEWater) and free-air cooling (Amazon) are all in production now.
4. Whether we can see it. Transparency is the precondition for everything else; the AGU Advances authors call for site-level disclosure to support regulation and community planning. Ask any developer for litres per day, the source, and the dry-season plan.

What to watch from Bali: Johor's mid-2027 reassessment; Batam's facility count and whether its promised treatment capacity arrives before the data centers do; any proposal for a large water-cooled facility in Bali or drawing on Balinese aquifers — that would change this article's conclusion overnight.

Method and caveats

This brief rests on two rounds of research (August 2026) that checked figures against primary documents: the LBNL 2024 report (direct water 66 billion litres and indirect ~800 billion litres in 2023; hyperscale 60–124 billion litres projected for 2028; sector WUE just over 0.36 L/kWh), the IEA's Energy and AI, company reports, court records and local reporting. Known limits:

• The IEA's split of its 560 billion litres into indirect, direct and manufacturing water circulates in secondary sources and was not confirmed against the report page, so it is not charted.
• Company figures are not like-for-like: Google and Meta report water consumed; Microsoft and Amazon report water withdrawn (a larger number). Meta's fleet WUE is variously reported as 0.18, 0.20 or 0.26 L/kWh depending on year and scope.
• xAI's Memphis water use has been reported at 30,000 gallons a day (early), ~381,000 (summer 2025 peak), ~812,000 (activist estimate), 1.2–1.3 million (utility contract cap) and 5–10 million (projections); we present it as a rising range.
• Per-app totals are estimates by construction — no company splits water by product — and per-query figures span roughly 40× depending on scope. Hatched bars mean "order of magnitude".
• Chile's court sequence is dated differently by sources (rulings in February and September 2024, suspension in July 2024); we describe the sequence rather than a single date.