Few infrastructure questions have stirred as much public anxiety in 2026 as a deceptively simple one: how much water does artificial intelligence drink? As AI data centers multiply across drought-prone regions from the American Southwest to the Netherlands, communities have pushed back, worried that the facilities training tomorrow’s models are straining the local supplies that serve their homes and farms. The concern is real — a single large campus can withdraw as much water as a small town, and U.S. data centers consumed an estimated 17 billion gallons in 2023.
But to the engineers designing these facilities, 2026 looks less like a crisis than an inflection point. The technologies that created the water problem are already being replaced, and the larger architecture of the data center — how it is cooled, how it is powered, and even what fluids run through it — is being reinvented from the chip outward.
Among the engineers who have tracked that shift from the inside is Mohit Shrivastava, P.E., a licensed mechanical engineer whose roughly fifteen-year career has moved from designing heat exchangers and thermal systems toward leading engineering analytics for AI-focused data center operations. During his time at Amazon Web Services, he supported thermal engineering across a global fleet of hyperscale data centers, building physics-based energy models and helping standardize engineering methods across large multidisciplinary teams. He is now Director of Engineering Analytics at Switch.
“The water debate is really a thermodynamics debate in disguise,” he explains. “Once you understand the physics of how older facilities reject heat, the path to fixing it becomes a concrete engineering problem.”
Why warmer water is the answer
Most of the water a data center “consumes” leaves through evaporative cooling. In a conventional facility, hot water from the servers flows to cooling towers, where part of it is deliberately evaporated. Evaporation sheds heat efficiently — every gram of vapor carries away a large amount of energy — which is why operators relied on it for decades. The water itself isn’t destroyed; it rejoins the atmosphere as part of the natural water cycle and eventually returns as rain. The problem is local: in a water-stressed region, that moisture leaves the community’s supply faster than it comes back, which is why the industry-average Water Usage Effectiveness (WUE) of roughly 1.8 liters per kilowatt-hour has drawn so much scrutiny.
The counterintuitive fix is to run the cooling water hotter. ASHRAE’s guidelines classify liquid-cooling loops by supply-water temperature, and the higher classes — historically labeled W3 (up to 32°C) and W4 (up to 45°C) — are now central to the conversation. When chips are designed to accept warm water rather than chilled water, the facility n
Most of the water a data center “consumes” leaves through evaporative cooling. In a conventional facility, hot water from the servers flows to cooling towers, where part of it is deliberately evaporated. Evaporation sheds heat efficiently — every gram of vapor carries away a large amount of energy — which is why operators relied on it for decades. The water itself isn’t destroyed; it rejoins the atmosphere as part of the natural water cycle and eventually returns as rain. The problem is local: in a water-stressed region, that moisture leaves the community’s supply faster than it comes back, which is why the industry-average Water Usage Effectiveness (WUE) of roughly 1.8 liters per kilowatt-hour has drawn so much scrutiny.
