Every era builds factories for the thing it needs most. Large-scale farms produced the food required to expand populations. The industrial age built them for steel; the digital age built them for chips and code. Factories producing tokens create them for a single use: artificial intelligence. The 1.21 Gigawatt AI Factory will take us back a step to create the future. They are factories in the truest sense: power in, intelligence out.

Factories are good – they produce at the scale that societies consume. The question communities have always asked of them is not whether they should exist, but how they contribute to the communities they exist in. In the US, they expand the tax base. They create jobs for skilled trades. They create a leadership in the most ground-breaking technology leap since the space race.

For AI factories, that conversation has so far been measured in megawatts. But water is every bit as critical, and for the communities that host these facilities, it is fast becoming as decisive as megawatts in determining whether, and where, the industry can grow.

The numbers bear this out. The International Energy Agency forecasts that global data center water withdrawals will exceed 1,200 billion liters annually by 2030, and the Environmental and Energy Study Institute estimates a single large data center can draw up to five million gallons a day – as much as a town of 10,000 to 50,000 people.

And these figures are before AI reaches full stride: LBNL's 2025 Update projects data centers could account for nearly 12 percent of total US electricity by 2030, up from 4.7 percent in 2024. Water is on the same trajectory as power – and it is arriving in communities that measure it in reservoirs, wells, and household bills.

Why does AI data center cooling consume so much water?

Power and cooling are two sides of the same coin; we can’t defy physics and destroy energy. Every watt a server draws must leave the building as heat, moved by a cooling system.

In the last generation of data centers, the dominant way large facilities reject heat economically is evaporative cooling: water is deliberately evaporated in cooling towers, carrying thermal energy into the atmosphere. That water is not borrowed and returned. It is consumed – gone from the local system.

AI multiplies the load and scale. Accelerator roadmaps are climbing past 4,000 watts per chip, and racks that once drew tens of kilowatts are being specified in the hundreds. More watts means more heat; under conventional architectures, more heat means more evaporation – just as utilities and neighbors grow more skeptical.

Three ways to cool an AI factory

The industry has three architectures to choose from, and they form a progression.

• Open loop, evaporative: Facility water is drawn in and evaporated to reject heat. Cheapest per megawatt on paper, it is where the alarming figures come from – and the real subject of permit disputes from Arizona to the Great Lakes.
• Closed loop with dry coolers: Water (typically a water-glycol mix) circulates in a sealed circuit and rejects heat to air. This largely solves consumption at the facility – but trades one water problem for another on the IT equipment: water still flows over multi-billion-dollar compute, bringing leak risk, corrosion, and constant fluid-chemistry maintenance into the white space. And dry coolers lose efficiency precisely where the AI build-out is headed: hot regions where summer temperatures force oversized equipment or evaporative trim cooling on the worst days. Over a facility's 20-to-30-year life, that residual dependency ages badly.
• Completely waterless: The third architecture designs water out entirely, starting at the chip where the heat is born. No water consuming cooling towers, no facility water loop, no water over the silicon – no dependency left to manage.

The progression matters: the industry moved from consuming water to containing it. The endpoint is eliminating it.

What is waterless two-phase direct-to-chip cooling?

In waterless, two-phase direct-to-chip liquid cooling – the approach behind ZutaCore HyperCool – a dielectric heat transfer fluid circulates in a sealed loop. As the processor heats up, the fluid changes from a liquid to vapor state in a cold plate, absorbing large amounts of thermal energy in the phase change. In the sealed system, the vapor condenses and returns in a liquid state to a cold-plate. No water is consumed – or even present – anywhere in the heat-rejection chain.

Because the mechanism is boiling rather than liquid volume and flow, the approach keeps pace with the hottest silicon on the roadmap, while sidestepping the risks of pushing water over the most valuable compute ever deployed. And because the loop is sealed, the same design performs identically in the desert or beside a river – decoupling site selection from water rights altogether.

The factory that gives back

That decoupling is where the AI factory story turns positive. A waterless facility can be built where power, land, and latency make sense, not only where water can be secured. Capturing heat at the chip supports higher rack densities per megawatt granted. And the heat that would have been evaporated away emerges concentrated and usable: warming homes through district heat networks, driving thermal desalination, even converting back to electricity.

The factories of the last century took from their towns and were made to clean up. The AI factory can skip that chapter – arriving as a neighbour that draws no water and has the option to return a useful resource: heat.

AI's growth will be negotiated community by community, and the strongest case for expansion will belong to operators who can show the water question has been engineered away, not managed around. The technology exists. The industry simply has to choose it.