The Thirsty Algorithm: Why Israel Should Price Water Into Its AI Ambitions

Who has the data?

Every fashionable anxiety eventually finds its viral statistic, and artificial intelligence has now found water. The claim ricocheting around social media — that a single ChatGPT email “drinks” a bottle of water — is not quite wrong, but it is not quite the study either. The 519-millilitre figure comes from a 2024 Washington Post analysis produced with the UC Riverside researcher Shaolei Ren, not from the underlying peer-reviewed paper, which more cautiously estimated that GPT-3 consumes roughly a 500-millilitre bottle for every ten to fifty responses. The number swings wildly with where and when the data centre runs, and the larger share of that water is not spilled on silicon at all — it is evaporated upstream, generating the electricity the servers consume. The honest version is less frightening and more interesting: AI’s water footprint is real, large at scale, and almost entirely a function of geography.

And geography is exactly where Israel’s position becomes interesting — because no developed country has had to think harder about the cost of water in an arid place, or has more thoroughly engineered its way out of scarcity.

The same Riverside team projects that global AI could withdraw 4.2 to 6.6 billion cubic metres of water annually by 2027 — somewhere between four to six Denmarks and half the United Kingdom. (The figure is for water withdrawn; the volume actually evaporated and lost is roughly a tenth of that — a distinction the headlines invariably drop.) For most countries this is an abstraction. For Israel it is a balance-sheet question, because Israel is the one developed economy that has already solved the problem the rest of the world is only now discovering it has. The country produces the overwhelming majority of its drinking water through reverse-osmosis desalination and reuses more than 87 percent of its treated municipal wastewater for agriculture — the highest rate in the developed world, several times that of any European peer. Israel did not stumble into water security. It manufactured it.

Here is where an economist’s instinct matters more than an environmentalist’s alarm. Water in Israel is not a free natural endowment to be guarded; it is a produced good with a marginal cost denominated in kilowatt-hours. Desalinating a cubic metre of Mediterranean seawater takes energy. A data centre cooled by that water is therefore consuming electricity twice — once to cool the chips, and once, embedded, to have produced the coolant in the first place. The water-versus-energy debate that paralyses American jurisdictions — should the aquifer feed the farms or the servers? — barely arises here, because the two inputs have already been collapsed into one. A country that prices its water against energy can do the arithmetic that Phoenix and Northern Virginia cannot.

That arithmetic should be run before the concrete is poured. Israel’s data-centre operators are already discovering that summer heat throttles processors sized for gentler enterprise loads, and that grid connections lag demand by eighteen to thirty-six months. The temptation will be to treat water as the easy input — abundant, engineered, available — and to let cooling design follow the brochure rather than the August peak. That would be the strategic error. Direct liquid cooling and reclaimed non-potable water are not green virtue-signalling in the Israeli context; they are the rational response to a coolant that arrives with an electricity bill attached.

There is a darker pattern this argument should name. The water cost of AI is increasingly being exported to the countries least able to absorb it — a phenomenon critics now call “data colonialism.” The servers answer prompts in California and London; the cooling water is drawn from Querétaro, Santiago, and Montevideo. In 2023, Uruguay’s capital became, by one environmentalist’s account, the first in the world to reach “day zero” and ran so short of fresh water that authorities mixed brackish river water into the public supply — while Google negotiated a data centre projected to consume the daily equivalent of 55,000 people. Chile’s courts partially revoked Google’s Santiago permit over its cooling draw; the company eventually paused the $200 million project and went back to the drawing board. A June 2026 UN University report warns that by 2030 AI’s water footprint could equal the annual domestic water needs of all 1.3 billion people in Sub-Saharan Africa. The injustice is structural: the value accrues to the wealthy economies whose users generate the queries, while the hydrological risk settles on semi-arid host nations lured by investment headlines.

Israel’s position is the mirror image of these countries. It neither needs to import the water cost nor to export it, because it manufactures its own supply. The viral statistic frames AI water use as a moral problem — water drawn from regions already experiencing severe drought. Israel inverts that logic. It is the country that exports the solution: the desalination engineering, the leak-detection algorithms, the wastewater reuse that even Iran now needs and cannot build for itself. The same scarcity that ought to make Israel disciplined about thirsty data centres is precisely what has given it the toolkit the data-centre world will be buying for the next two decades. The constraint became the comparative advantage.

So the question Israel faces is not whether AI uses too much water. That is the question for places that never had to think about water at all. Israel’s question is whether it will price the coolant honestly, build for the August peak rather than the brochure, and recognise that its hard-won water expertise is not a defensive moat but an export. The thirsty algorithm is coming everywhere. Israel is one of the few countries that already knows what a litre actually costs — and one of the very few that can sell the answer.

Appendix: Water stress and AI data-centre conflict in selected host economies

Uruguay (Canelones / Montevideo) — Worst drought in roughly 70 years in 2023; the capital reached “day zero,” with brackish river water added to the public supply. Google’s project was estimated at around 7.6 million litres per day for cooling, equal to the daily domestic use of about 55,000 people. Public protests followed; the project was scaled back to less water-intensive cooling.

Chile (Santiago — Cerrillos and Quilicura) — Severe multi-year drought; around half of the country’s 19 million people under water stress by 2022, with rationing in Santiago. Google’s planned cooling draw was about 169 litres per second; Microsoft also proposed a facility in Quilicura. A local referendum opposed the project, a court partially revoked the permit, and Google paused its $200 million plan to redesign the cooling system.

Mexico (Querétaro — Colón) — A semi-arid region subject to climate-intensified drought. Roughly $10 billion in Google, Microsoft and AWS investment; three proposed centres could draw up to about 4 million litres per day in a town of 65,000. Resident protests continue over industrial versus domestic priority.

United States (Loudoun County, Virginia; Silicon Valley, California) — Localised stress and recurrent California drought. Water use in Loudoun’s “Data Centre Alley” is up roughly 250 percent since 2019; a single 15-megawatt facility can use about 1.36 million litres per day. As a wealthy economy, the US absorbs the cost domestically.

Malaysia — Grid and water strain emerged alongside hub ambitions in Johor. The government moved in September 2025 to slow approvals.

Israel — Engineered water security: the majority of drinking water is desalinated and more than 87 percent of effluent is reused, the highest rate in the OECD. The domestic build-out faces summer-heat throttling and an 18–36 month grid lag, but no water-scarcity conflict. Israel is positioned as an exporter of desalination, reuse and leak-detection technology.

Sources: Li, Yang, Islam and Ren, “Making AI Less ‘Thirsty'” (2023/2025); UNU-INWEH, “Environmental Cost of AI’s Energy Use” (June 2026); Mongabay; Nearshore Americas; PBS/AP; OECD. Figures are cooling estimates and vary by source and design assumptions.