Here’s something that doesn’t get discussed enough in developer circles: every time we train a large language model, deploy a new model to production, or even run inference on existing models, we’re drawing on infrastructure that has real physical consequences. I mean, we’re talking about buildings that span 20,000 acres, use massive amounts of water, and consume enough electricity to power small cities.
Kevin O’Leary, the guy from Shark Tank, just agreed to shrink his massive Utah data center project, Project Stratos, from 40,000 acres down to around 20,000 acres. That’s still larger than Manhattan, and it still has people concerned about water usage in a state that’s been dealing with serious drought issues. Utah’s Senate President asked for a 75% reduction, but O’Leary only went for about half that. Even so, this is a big deal.
Why This Matters for AI Developers
The connection between artificial intelligence and these massive facilities is direct. Modern AI models, especially the large ones everyone is excited about, require enormous computational resources to train and run. We’re not talking about a Python script running on your laptop. We’re talking about clusters of specialized hardware running 24/7, generating heat, and consuming both electricity and water for cooling.
The water part is what gets me. Data centers need cooling systems, and in dry environments like Utah, that means evaporative cooling which consumes water. The project is located near the Locomotive Springs Waterfowl Management Area, and there’s already concerns about the Great Salt Lake shrinking. O’Leary was asked to implement technology that minimizes water consumption and to divert excess water to the lake. These aren’t trivial requests in a region where water is becoming increasingly scarce.
The Industry’s Reckoning
What I’m seeing here is a microcosm of a larger issue the entire tech industry is going to face. As demand for AI capabilities grows, so does the need for infrastructure. Companies keep building bigger, more powerful models, but the physical footprint of the infrastructure required to support them is often an afterthought in the excitement.
Developers, we’re the ones building on top of these systems. We should be asking hard questions about where our models run, what resources they consume, and what the long-term implications are. This isn’t just an environmental concern, it’s becoming a business and regulatory one too. If states start imposing stricter requirements on data center construction, if water usage becomes heavily regulated, if there’s pushback from communities the way there was in Utah, that affects everyone building AI products.
What We Can Do
I’m not saying we should stop building AI applications. That would be absurd. But there are things we can think about. Training smaller, more efficient models that get the job done without requiring massive infrastructure. Deploying models closer to where they’re used to reduce latency and infrastructure needs. Being mindful of inference costs and optimization. Supporting companies and projects that are investing in greener infrastructure.
The Project Stratos situation tells me the era of unchecked data center growth is ending. Environmental concerns, community pushback, and resource scarcity are all converging. The developers who think about these things now will be better positioned for whatever comes next.
The conversation about AI’s environmental footprint is just getting started, and it will shape how we build these systems for years to come.