The Real AI (Resource Depleting) Load Is In Cooling, Not Compute.
My friend, let’s call him Andy, recently relocated from the fancy and growing Mueller district in Austin Texas and bought a house in San Marcos. San Marcos, with a population of 74k, is a town between Austin and San Antonio and fast becoming the ‘connector city’ between these two hubs that have million people populations. Andy works in tech in downtown Austin. He moved because he could no longer keep renting in Mueller and his family is growing. He’s expecting his first child in November of this year and his wife wants a yard. She also wants an extra room for her mother who will be coming to help with the baby. Andy bought a home in San Marcos because it is impossible for the average worker to buy in Austin. The same tech companies that (unintentionally) influenced the rising home prices in Austin are now threatening the well-being (and affordability) that Andy sought in San Marcos…
———
Compute itself - the conversion of energy into intelligence - is the star of the show when we talk about the energy demands as a result of data centers. But the real story is of the ecological impacts of this increased load/demand on sectors like water. According to IEA and Energy.gov Data centers already consume about 4–5% of U.S. electricity (~176 TWh in 2023), a share projected to climb to 7–12% (325–580 TWh) by 2028, with AI workloads driving much of this explosive growth in demand.
The underlying story we are not paying attention to is that of the water (thermal cooling) component of this growth. In Texas, data centers will use about 54 TWh of electricity in 2025, and depending on how they’re cooled, and attributing all infrastructure overhead to cooling, roughly 25-40% of that power goes just to cooling the machines rather than compute. This has a huge effect on the water used wherever that data center is located also. A 100 MW AI data center can use 46 - 417M gallons/yr of cooling water (design-dependent), which is roughly the annual usage of 500-5000 homes in Texas —and that’s before counting the additional indirect water used to generate its electricity.
———
As demand for more and more compute threatens our power grid, data center developers (AI and cloud regionals) are in an aggressive arms race to find, acquire and generate power on land close to major cities. There isn’t land or available power in the middle of our major cities. Some cities, even as they desire the revenue from new business coming into their cities, are starting to turn away data center development. This has led to increasing interest in places like San Marcos evidenced by the recent purchase of 1515 acres of land by a developer for a 2GW data center.
But small towns like San Marcos do not have the infrastructure to support some of the resource demands of these data centers. San Marcos is currently under Stage 3 drought restrictions, a classification that was simplified from 5 stages to 3, a town under serious strain. The Stage 3 restriction is due to extended drought across Central Texas, falling aquifer levels, and reservoir depletion, all of which threaten the city’s primary water sources.
That yard in the home that Andy bought? He can only water it once between 8:00pm - 10:00am every two weeks. He shouldn’t really wash his car. But he can, thankfully, keep a small vegetable garden lightly watered every day. At the same time a site that uses 500-5k times the water Andy uses is about to be built less than 20 miles from his home. And it will not have the same restrictions. What are we actually doing?
———
Look. I like my AI tools. A lot. Some of the data for this post came from research carried out by several of these AI tools. What I am offering is that we are scaling data center infrastructure at the expense of deep considerations for the interconnected systems that are at play here. Will there be water to cool the San Marcos data center? A calculation of the water budget (calculator tool found here) and the consequent water stress should be implemented before an application is even made. A simple run of the too would should that the live condition of San Marcos ‘is experiencing D4 (Exceptional Drought) conditions. This severe drought is characterized by significant precipitation deficits, decreased soil moisture, and low reservoir levels.”
AI infrastructure planning must seriously consider water budgeting and not just power. Cooling design choices are becoming ever more critical is the decision of what to build and design, closed loop air vs dielectric vs liquid immersion cooling questions should be answered early with due consideration for drought or water availability. These are equally as important as power and fiber availability. The real computing cost is fast becoming thermal management. We should think of it that way. Andy and his incoming newborn need us to pay more attention.
Seyi Fabode is an expert at the intersection of power, utilities, technology and cities. He a partner at www.asha-labs.com and runs www.asimovgrid.com