Few infrastructure questions have stirred as much public anxiety in 2026 as this: how much water does artificial intelligence drink? As AI data centers spread across drought-prone regions from the American Southwest to the Netherlands, communities are pushing back. 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.
To engineers designing these facilities, 2026 is an inflection point. The technologies that created the water problem are being replaced. Cooling, power, and fluids are being reinvented from the chip outward.
Mohit Shrivastava, P.E., has tracked this shift closely. A licensed mechanical engineer with fifteen years of experience, he moved from designing heat exchangers to leading engineering analytics for AI data centers. At Amazon Web Services, he built physics-based models for hyperscale facilities. He now serves as 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?
Most questioned data center water use comes from evaporative cooling. Hot server water flows to cooling towers where part of it evaporates, carrying away heat efficiently. The vapor returns to the water cycle as rain, but locally it depletes supplies in stressed areas. This is why the average Water Usage Effectiveness (WUE) of 1.8 liters per kilowatt-hour faces heavy scrutiny. This had power savings advantages due to which it became very popular.
The solution now is counterintuitive: run the cooling water hotter. ASHRAE guidelines define higher liquid-cooling classes (W32 up to 32°C, W45 up to 45°C). When chips accept warmer water, facilities no longer need energy-intensive chillers. They can use dry coolers or closed loops with almost no evaporation.
“If your chips are happy receiving water at 32 or even 40 degrees Celsius, you can cool that water against ambient air almost everywhere on Earth for most of the year — without a compressor and without evaporative cooling,” Shrivastava says. “You trade a little cooling efficiency for an enormous reduction in both water and electricity.”
Results at scale are impressive. Closed-loop dry-cooled systems can cut freshwater use by up to 70 percent. Peak daily water use drops sharply, and leading sites now achieve WUE below 0.3 liters per kilowatt-hour. The loop is filled once and recirculated for years with proper treatment.
The Forever Chemicals Crisis That Stopped Two-Phase Cooling
Two-phase immersion cooling once seemed ideal but relied on PFAS “forever chemicals.” After 3M exited production and regulators tightened rules, the supply chain collapsed. Shrivastava, whose graduate work focused on two-phase heat transfer, regrets the loss of the elegant physics but supports the change.
Practical replacements dominate today: single-phase direct-to-chip cooling with water-glycol mixtures and non-PFAS dielectric fluids for immersion. “Choosing a cooling technology is never just about which one moves the most heat,” Shrivastava says. “It’s about what you can deploy, service, and trust at scale for a decade.”
The Power Revolution with 800-VDC
AI racks now demand hundreds of kilowatts and are heading toward megawatts. Low-voltage (48-54V) distribution requires too much copper and struggles with massive current. The industry is moving to 800-volt DC, reducing copper, losses, and conversion steps. Racks now include batteries and supercapacitors to handle rapid AI power swings.
“Power has quietly become the defining bottleneck,” Shrivastava observes. “We used to design the building around the servers. Now we design everything around getting clean, stable power to the chip.”
The Road to 2030
Liquid cooling is becoming standard. Warm-water loops with dry rejection enable near-zero-water operation. High-voltage DC with distributed storage is growing. Waste heat will be reused, and real-time models will optimize the entire facility.
The field rewards engineers who combine deep expertise with system-level thinking across mechanical, electrical, controls, and data domains. “The engineers who will thrive,” Shrivastava reflects, “are the ones who can hold the whole system in their head — physics, hardware, and data together — and still get their hands dirty in the lab.”
As AI grows, the quiet revolution in physical infrastructure proves that scaling intelligence and respecting resources can go hand in hand.
