If you follow AI data center news, you have probably noticed that most headlines fixate on chips, power capacity, and billion-dollar buildouts. But water is becoming just as important. The International Energy Agency says data centers used about 415 TWh of electricity in 2024 and projects that figure could climb to around 945 TWh by 2030, with AI as a major driver. At the same time, the World Resources Institute says two-thirds of U.S. data centers built or in development since 2022 are in water-stressed areas, which turns cooling from a technical detail into a local political issue. (IEA)

That matters because most people still picture AI as “just software.” In reality, every model runs on physical infrastructure that needs to stay cool. And once you start looking into it, the question is no longer just “How much electricity does AI need?” It becomes “Where will the water come from, who gets priority, and what happens during droughts?” That is why this topic is no longer niche. It is one of the biggest undercovered parts of the current compute boom.

If you want the broader infrastructure backdrop, this pairs well with our coverage of AI data center news today.

What “AI data center water usage” actually includes

When people talk about AI data center water usage, they often mean server cooling. That is the most obvious part, but it is not the whole story. A recent UK government-backed report breaks the footprint into three buckets: water used for cooling, water embedded in electricity generation, and water used in semiconductor manufacturing. In other words, even if a facility lowers on-site water use, its total water footprint can still stay high if its electricity mix or chip supply chain is water-intensive.

Cooling is still the main direct issue. The same report says that with evaporative cooling, roughly 80% of the water used in the process can be lost to evaporation, which means it has to be replenished continuously. It also notes that this replenishment often uses potable water, not some mysterious industrial liquid. That is a big reason communities react strongly when new projects are announced near drought-prone areas.

There is also an energy-water tradeoff that gets ignored in simplified debates. Google says it tries to balance carbon-free energy and responsibly sourced water when choosing cooling approaches, and the IEA notes that cooling can account for about 7% of electricity use in efficient hyperscale data centers and more than 30% in less-efficient enterprise facilities. So the industry is not choosing between a perfect option and a bad one. It is usually choosing between different environmental tradeoffs. (Google Data Centers)

For a related look at the economics pushing all this expansion, see why OpenAI is burning cash while Google and Anthropic aren’t as much.

How much water can an AI data center use?

The big numbers are eye-opening. The UK report, citing IEA figures, says the global data center sector already consumes more than 560 billion liters of water annually, and that this could rise to as much as 1,200 billion liters by 2030. It also says a 100 MW hyperscale facility can consume around 2.5 billion liters of water per year, roughly equivalent to the needs of 80,000 people.

At the U.S. level, WRI says AI-related data centers could require up to 32 billion gallons of water annually by 2028. It also notes that mid-sized facilities can use up to 300,000 gallons a day, while large ones can reach 5 million gallons daily. Those numbers help explain why local water boards, utilities, and residents are paying much more attention now than they did during earlier cloud expansion waves. (World Resources Institute)

Company disclosures also show the direction of travel. According to the UK report, Microsoft’s global water use rose 34% in 2022 to 6.4 million cubic meters, while Google’s data centers consumed 19.5 million cubic meters in 2022, up 20% year over year. Those are not apples-to-apples measures for every single facility, but they do show the broader trend: as AI workloads scale, water demand rises unless efficiency improvements outpace growth.

A concrete example helps. AP reported that in West Des Moines, Microsoft’s local facilities used 11.5 million gallons in July 2022 alone, and that officials later required the company to reduce peak water consumption for future projects. That is the kind of local detail that turns an abstract sustainability discussion into a planning fight. (AP News)

Why AI workloads are making the issue worse

The reason AI changes the conversation is simple: denser racks, hotter chips, and more compute. The IEA says accelerated servers, which are mainly tied to AI adoption, are projected to grow electricity demand by about 30% annually in its base case, and that they account for nearly half of the net increase in global data center electricity consumption through 2030. More power density usually means more heat management pressure, which is where water enters the story again. (IEA)

This is also why per-prompt water estimates can be useful but misleading if they are treated like universal truth. A UC Riverside summary of a widely cited study estimated that roughly 20 to 50 ChatGPT queries may consume about half a liter of freshwater. But later expert reactions have argued that neat “per prompt” numbers can understate reality when they exclude indirect water use from electricity generation or rely on narrow assumptions. The Verge reported researchers criticizing one company-backed estimate for doing exactly that. The fair takeaway is not that the earlier study was fake; it is that water accounting depends heavily on location, grid mix, time of day, hardware, and cooling design. (news.ucr.edu)

That nuance matters for readers because a lot of online discussion swings between two extremes: either “AI is destroying the planet” or “it’s only a few drops.” The actual picture is more complicated. AI’s water footprint is real, measurable, and likely to grow, but it is also highly dependent on how and where systems are built. (arXiv)

If you want the geopolitical angle behind this infrastructure race, can China win the AI race is worth reading alongside this piece.

Why communities and regulators suddenly care so much

This is where AI data center water usage stops being a technical sustainability KPI and becomes a public issue. WRI says more than $64 billion in data center projects were delayed or canceled between May 2024 and March 2025 due to organized opposition. Water is not the only reason, but it is clearly one of the major flashpoints, especially in places already dealing with drought, aquifer stress, or fast population growth. (World Resources Institute)

Transparency is a big part of the problem. The UK report says only two-fifths of data center operators actively track water usage metrics, and it recommends mandatory, location-based reporting. That is a striking number. If operators are asking communities for land, transmission upgrades, water access, and tax incentives, it is not unreasonable for residents to ask for site-level water data in return.

I think this is the part many AI boosters still underestimate. People will tolerate a lot when a project promises jobs and investment. They get much less enthusiastic when they hear that the facility may be drawing potable water in a stressed watershed, while still employing relatively few permanent workers. That tension is only going to intensify as more AI campuses get announced. The evidence from WRI suggests this is already happening. (World Resources Institute)

What companies are doing to reduce water use

To be fair, the industry is not standing still. Google says it evaluates the local water source for each data center, uses a science-based watershed health framework, and in high-risk locations looks for alternatives such as air-cooling or recycled water. It also says it aims to replenish more freshwater than it consumes, on average, across data centers and offices by 2030. (Google Data Centers)

Microsoft is also publishing efficiency data more openly than many operators. Its data center site defines WUE, or water usage effectiveness, as liters of water used for cooling and humidification per kilowatt-hour of IT energy. Microsoft says its global WUE improved from 0.30 to 0.27 L/kWh between FY24 and FY25. On top of that, Microsoft said in its 2025 sustainability update that a new data center design optimized for AI workloads uses zero water for cooling and can avoid an estimated 125,000 cubic meters of water annually per facility. (Microsoft Datacenters)

There are also promising technical shifts. The UK report says some advanced cooling approaches, including direct-to-chip and certain immersion systems, can reduce water use by 31% to 52% over their life cycles compared with traditional air cooling approaches cited in the report. It also highlights closed-loop systems, non-potable water, rainwater harvesting, and hybrid cooling as practical levers. The key point is that “AI data center water usage” is not fixed. It can be shaped by engineering decisions and site selection.

Still, the hard part is scale. Google’s sustainability pages also show that even while companies improve efficiency, their overall infrastructure keeps growing. Google said its data center electricity consumption increased 27% year over year in 2024 due to business growth and rising product adoption, including AI. That is the classic problem: efficiency gains help, but demand can rise faster than efficiency improves. (Google)

The real bottom line on AI data center water usage

The cleanest way to say it is this: AI data center water usage is no longer a side issue. It is becoming a core infrastructure constraint. The sector’s water footprint already sits in the hundreds of billions of liters globally, hyperscale sites can rival towns in daily use, and many new builds are still landing in water-stressed regions. At the same time, operators do have real tools to reduce the damage, from recycled water and hybrid cooling to site-level reporting and zero-water designs.

So the smartest question for readers is probably not “Does AI use water?” That answer is obviously yes. The better question is: Which companies are willing to disclose the numbers, which sites are being built in the wrong places, and which engineering choices actually reduce pressure on local communities? That is where the next phase of AI data center news will get much more interesting. (World Resources Institute)

And honestly, this is one of those topics where the comments matter. If a major AI data center were proposed near your city, would you care more about jobs, tax revenue, electricity demand, or water withdrawals? And do you think site-level water reporting should be mandatory?

FAQ

Do AI data centers use drinking water?
Often, yes. The UK report says evaporative cooling systems can rely on potable water to avoid clogging and contamination, while Google says it is increasingly exploring alternatives such as recycled water in higher-risk locations.

Is water use only about cooling?
No. Direct cooling is the biggest visible use, but the broader footprint also includes water used in electricity generationand semiconductor manufacturing. That is why narrow site-level numbers do not always tell the whole story.

Are all AI data centers equally water-intensive?
Not even close. Water use varies by climate, cooling design, local grid mix, hardware density, and the availability of recycled or non-potable water. Microsoft’s WUE framework and Google’s site-by-site water-risk approach both reflect that reality. (Microsoft Datacenters)

Can the industry reduce water use without slowing AI growth?
It can reduce water intensity, but that does not automatically reduce total water consumption. Better cooling, closed-loop systems, and smarter siting all help, yet overall demand can still rise if AI infrastructure grows faster than efficiency improves.