Honestly? The best thing to happen to AI transparency this year was somebody’s sloppy PDF work.
That sounds like a joke, but I mean it. For years, the conversation about AI’s environmental cost has run on estimates, extrapolations, and whatever companies chose to publish in their sustainability reports. Then a redaction in Nebraska didn’t take, and suddenly we had actual numbers attached to actual buildings. Not a model. Not a projection. A filing.
Here’s what came out. Google’s Lincoln data center, filed under the name Agate LLC, reported a peak electricity demand of 52.65 megawatts and annual water consumption of roughly 13 million gallons. Across six Nebraska data centers, annual water use came to approximately 765 million gallons. The biggest single user was Google’s Fireball Group LLC facility in Papillion, at about 548 million gallons in 2025.
Nebraska requires data centers to submit these annual reports to the state Department of Water, Energy, and Environment. State law also lets companies shield much of that information from public view. In this case, the shielding failed.
Why I care about this on a site about AI agents
If you read agent101.net, you’re probably here because you want to understand what AI agents actually are and how they work, without a computer science degree getting in the way. Fair enough. But I’d argue the physical side of AI is part of that explanation, not a separate topic.
When you ask an AI agent to book a flight, summarize a contract, or sort through your inbox, the work doesn’t happen on your laptop. It happens in a building. That building has a transformer, a cooling system, and a water bill. Agents are a particularly interesting case because they don’t do one thing and stop. They chain steps together. One request can mean several model calls, each one a tiny draw on a very real power grid.
So when someone asks me “what does an AI agent cost?”, the honest answer has two halves. There’s the price on your invoice. And there’s the half that shows up in a county utility report that you were never supposed to read.
Reading the numbers without panicking
Let me be careful here, because big numbers invite bad conclusions.
Thirteen million gallons a year sounds enormous. Compared to what, though? Compared to a household, it’s astronomical. Compared to agriculture in the same state, it’s a different conversation entirely. Nebraska is farm country, and irrigation moves water at a scale that dwarfs most industrial users. Context matters, and I’d rather give you the tools to ask good questions than hand you an outrage number.
The numbers worth sitting with:
- 52.65 megawatts peak demand at one facility, which is a meaningful chunk of a local grid’s capacity
- 765 million gallons across six facilities in a single state, annually
- 548 million gallons at Papillion in 2025, meaning one site accounts for the majority of that six-site total
That last point is the one I keep coming back to. The distribution is lopsided. One facility, not six evenly, drives most of the water use. Which tells you that “data center” isn’t a single category with a single footprint. Design choices, cooling methods, and local climate all change the math enormously. Some sites sip. Some sites gulp.
The transparency problem is the actual story
Strip away the specific figures and what’s left is a process question. These reports exist. The state collects them. The public, by design, mostly can’t see them.
I don’t think that’s a conspiracy. Companies have legitimate reasons to protect operational details, and competitive concerns around infrastructure are real. But there’s a gap between “protecting trade secrets” and “citizens can’t find out how much water the building down the road uses.” Those are different things, and the current setup treats them as the same thing.
The irony is thick. The industry asking us to trust AI agents with our calendars, our spending, and increasingly our decisions is also the industry that would prefer we not see its water bill. Trust tends to be reciprocal. If you want people comfortable handing off tasks to autonomous software, giving them a clear picture of what powers that software seems like a reasonable starting point.
What I’d actually do with this
Nothing dramatic. I’m not going to tell you to stop using AI tools, because I use them constantly and that advice would be hollow.
What I’d do is adjust the mental model. AI agents aren’t weightless. They’re a service delivered from physical infrastructure in specific places with specific resource costs. When a company publishes an efficiency claim, ask which facilities it covers and whether the underlying data is public. When a new data center gets announced near you, ask what its reporting obligations look like and who gets to read the results.
The Nebraska filings got out by accident. The next useful set of numbers probably won’t. Which means the people asking for them have to be more persistent than the people hoping a redaction holds.
đ Published: