What the Musk-OpenAI Trial Isn’t Talking About: AI Is Drinking America’s Water
A different take on AI
What the Musk-OpenAI Trial Isn’t Talking About: AI Is Drinking America’s Water
The trial in Oakland has everything. Elon Musk and Sam Altman, two of the most powerful figures in AI, facing off in federal court. The mother of four of Musk’s children testifying. Secret Tesla work. Stolen talent. Journal entries that were never meant to be public. A $5,000-an-hour AI safety expert warning a jury that superintelligent AI could kill us all.
It is, by any measure, a spectacular show, probably the next limited series on Netflix.
Behind the spectacle, there are real infrastructure challenges quietly reshaping the AI landscape, including odd new partnerships like Anthropic committing to use SpaceX’s Colossus 1 data center, born entirely from GPU scarcity that is throttling the race toward AGI.
But what stopped me this week wasn’t the courtroom drama.
It was a video about data center cooling, and a number that I haven’t been able to shake: a single AI data center can consume up to 5 million gallons of drinking water per day.
Not industrial water. Not reclaimed water. Potable water, the kind you and I drink, drawn from the same municipal supplies serving local communities.
That’s enough to supply thousands of households or farms, every single day, for one facility.
We mere humans can live without electricity, dare I say we can live without AI, but we cannot live without water to drink, and the crops they help to grow.
The irony is that the real damage, the infrastructural, environmental, and social cost of building AI at the scale these men are competing to achieve, is happening right now, mostly unchecked in rural communities in the USA with limited oversight in real time, in communities nowhere near Oakland. And it is accelerating.
How Thirsty Is AI, Exactly?
The numbers are difficult to fully absorb.
Researchers at UC Riverside have estimated that a single AI chat session of around 20 queries uses roughly a bottle of water. That sounds manageable, until you multiply it by the billions of queries entered into systems like ChatGPT, Grok, and Claude every day.
Google reported consuming more than 5 billion gallons of water across its data centers in 2023 alone, with nearly a third of those withdrawals coming from watersheds already rated as medium or high-water scarcity areas. The industry’s response has been largely to build more, faster.
Silicon, Meet Drought
The United States is in the middle of a significant drought across large portions of its territory. America is not alone in this. The UN declared 2026 the year the world entered 'water bankruptcy.' With the potential of a Super El Nino, this could get far worse before it gets better.
The AI arms race that the Oakland trial is, in its own way, a proxy for, has been building its infrastructure in exactly the places least equipped to absorb the demand.
A University of Texas study released this week estimates that data centers could account for between 3% and 9% of Texas’ total water use by 2040, up from less than 1% today. Texas faces a projected $174 billion water shortfall over the next 50 years.
The Pentagon has approved construction of a 3-gigawatt data center on 1,384 acres at Fort Bliss in Texas, expected to consume more electricity and water than the entire city of El Paso once it opens in 2027.
This is not a future problem. It is a present one.
The Water Nobody Counts
What makes this worse is the transparency gap. Much of the water consumed by AI data centers is not publicly reported.
“All water is local,” as one water policy expert put it. The abstraction of cloud computing, the sense that AI exists somewhere up in a frictionless digital ether, has obscured a very physical reality: every query, every training run, every inference call has a water address. And that address is increasingly in a watershed under stress.
IBM’s Different Answer
While the hyperscale GPU data centers at the center of the AI arms race rely on massive evaporative cooling towers, systems that continuously draw and exhaust water from local supplies, IBM took a different path with the Z17 mainframe, its latest enterprise AI platform.
The Z17 uses a closed-loop Radiator Cooling Assembly built directly into the frame. Rather than drawing millions of gallons from municipal or groundwater sources and evaporating them into the atmosphere, the system recirculates coolant within a contained radiator, more analogous to a car engine than a cooling tower. The water stays in the loop. The local watershed is not the variable.
IBM claims the Z17 delivers comparable AI inference workloads at roughly five and a half times better energy efficiency than GPU-based alternatives.
This is not a legacy argument. It is an architectural one. IBM designed a system that does AI close to the data, within the existing footprint of enterprise data centers, without requiring new water-intensive infrastructure or competing for scarce GPU supply. For regulated industries, banks, hospitals, insurers, government agencies, that already operate Z infrastructure, this is not a theoretical advantage. It is an operational reality available today.
The Question the Trial Isn’t Asking
Who decided that training the next frontier AI model was a higher priority use of scarce fresh water than the people and farms and ecosystems that depend on it?
Who is accountable for that tradeoff? And what governance structure, nonprofit, for-profit, or otherwise, is actually designed to answer for it?
At Three Takes on AI, we have argued consistently that AI needs human accountability above commercial interest. The water story is not a peripheral concern to that argument. It is one of its clearest illustrations.
The trial in Oakland is asking who owns the future of artificial intelligence.
The more urgent question may be who is responsible for what building that future is already costing.
Three Takes on AI publishes podcasts bi-weekly. Previous episodes have covered Shadow AI, the MIT ROI study, AI governance frameworks, and the organizational and human dimensions of deploying AI responsibly. We would love to hear your take.