The Hidden Cost of AI: Why Data Centers Are Drinking the Future
We’ve spent the last decade talking about cloud like it was infinite, endless scale, unlimited compute, constant innovation. But the reality is, nothing about infrastructure is truly infinite. Especially not water.
When we think about data centers, we usually think about power, electricity, GPUs, and compute density. What we don’t think about is cooling, and that’s where the real story begins. Data centers run hot, really hot. To keep them operational, operators rely heavily on water-based cooling systems, and the numbers are staggering. A single data center can use around 300,000 gallons of water per day. Large facilities can reach up to 5 million gallons daily. By 2028, U.S. data centers could consume tens of billions of gallons annually. That’s not a rounding error. That’s infrastructure at the scale of cities.
At a high level, the reason is simple: heat. Every AI model, every API call, every agent loop generates heat, and water is one of the most efficient ways to remove it. Most systems rely on evaporative cooling, where water absorbs heat, evaporates, and is gone, not reused, not recycled. And that’s just direct usage. There’s also indirect water consumption, the water required to generate the electricity powering these systems, which often exceeds the cooling itself.
This is where the conversation shifts from sustainability to scale. AI didn’t just increase demand—it changed the curve entirely. Workloads run longer, models iterate continuously, and agents don’t stop unless you tell them to. Even something as simple as interacting with AI has a water footprint. Small per interaction, but massive at scale. Multiply that across enterprises, consumers, and autonomous systems, and what you get isn’t linear growth, it’s exponential demand across compute, power, and water.
We’ve seen this pattern before. Cloud wasn’t constrained by servers, it was constrained by operating models. AI won’t be constrained by GPUs alone. It will be constrained by water availability, local infrastructure capacity, and community impact. We’re already seeing data center projects delayed due to resource concerns, communities pushing back on water usage, and investors demanding transparency around environmental impact. Because water is local, and you can’t scale a river.
There’s also an uncomfortable trade-off at play. You can reduce water usage, but it usually comes at the cost of increased energy consumption. Air cooling uses less water but more power. Liquid cooling uses less power but more water. There is no perfect solution, only trade-offs.
Which brings this back to a familiar theme: technology isn’t the problem. The operating model is. We’re deploying AI the same way we deployed cloud in 2012; fast, excited, and largely uncontrolled. But this time, the cost isn’t just financial. It’s environmental, regional, and physical.
If AI is the next platform shift, then the way we build around it has to evolve. Cost awareness needs to become resource awareness, where FinOps expands beyond dollars to include water and energy. Architecture matters again, as workload placement, cooling strategies, and region selection directly impact sustainability. Accountability must exist at scale, where every model, agent, and loop carries a measurable real-world cost. And most importantly, we need to design before we deploy; because once something is live, its footprint is already locked in.
We used to ask, “How fast can we scale this?” Now we need to ask, “What does it cost the world when we do?” Because the next bottleneck in cloud and AI isn’t compute. It’s water.
(Part of my cloud + AI operating model series. More at michaelearls.com/cloud)
Data facts pulled from https://www.networkworld.com/article/4138052/why-do-data-centers-need-so-much-water.html