Every data center that trains or runs modern AI faces the same unglamorous problem: the chips get hot, and there are only two ways to cool them. You can pour water over the heat, millions of gallons of it, or you can spend enormous amounts of electricity moving air. Neither option is clean, and as a POLITICO investigation lays out, the industry has quietly turned that trade-off into a public-relations minefield.
The choice is close to zero-sum. Save water, and power consumption climbs, along with the carbon emissions and grid strain that follow. Cut electricity, and water use surges, competing with farms, ranches and households, often in the places that can least spare it. Two-thirds of US data centers built since 2022 sit in areas already under high water stress. Google's own math shows that using 264 million gallons of water in central Europe spares 41,000 megawatt-hours of power, enough to run 10,000 homes. Water, one Google executive said, has a "uniquely impressive power of reducing energy consumption."
So the companies have picked sides. Microsoft and Quality Technology Services promise zero-water facilities and accept the higher power draw and emissions in return. Google and Amazon use water in wetter climates to shave peak load when the grid is most strained, then switch to air where drought makes water the scarcer resource. But "zero water" is slipperier than it sounds. Even sealed closed-loop systems still have to reject heat from the building, and above roughly 85 degrees the most effective way to do that is, again, evaporating water. Amazon's own water strategist calls the zero-water promise a source of "a lot of confusion."
The public is not waiting for the fine print. Seven in ten Americans oppose data centers, with water use their top concern, according to a recent Gallup poll. In the first quarter of 2026 alone, at least 75 projects worth 130 billion dollars were disrupted by local opposition, and county-level moratoriums are spreading. Fresh Gartner figures explain the urgency: global data center electricity use will jump 26 percent this year to 565 terawatt-hours, and by 2027 AI-optimized servers will draw more power than all conventional data center hardware combined.
Underneath the backlash is an information vacuum. There is no federal requirement to disclose how a facility is cooled, and the operators that do report use mismatched metrics that make comparison nearly impossible. One expert told Congress that no other fast-growing energy sector suffers from as many "public data blind spots." Communities often cannot answer the single question that matters, which a Google engineer says they should always ask: how, exactly, is your cooling plant rejecting heat?
That secrecy, several researchers argue, is precisely what fuels the mistrust. Operators are chasing nuclear deals and exotic cooling to buy their way out of the bind, Meta alone has signed for more than six gigawatts of nuclear power, but those plants are years away. In the meantime the industry is stuck choosing which scarce resource to burn, and choosing, so far, to say very little about it. As one rancher put it, weighing his water against the cloud: "I'd prefer to be able to still have a steak and some salad than ChatGPT."