Madison Pitsch, reporting for a Lincoln newsroom, put it about as plainly as it can be put: improper redaction is letting the public see exactly how much water and electricity the Lincoln Google data center is using. Not a leak. Not a whistleblower. A black box that didn’t quite cover the numbers underneath it.
I review AI tools for a living, which means I spend most of my week looking at dashboards that promise visibility and deliver vibes. So I have a professional appreciation for what happened in Nebraska. The most useful disclosure in AI infrastructure this year came from someone’s PDF editing mistake.
What the numbers actually say
Google’s Lincoln facility, 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, the total annual water figure lands near 765 million gallons. The Fireball Group LLC site in Papillion carries the heaviest load at about 548 million gallons in 2025.
Run the division and the shape of the thing becomes clear. One facility accounts for roughly 72 percent of that 765 million gallon total. Lincoln, the site that triggered the story, is under two percent of it. If you had only seen the Lincoln figure, you would have walked away with a wildly incomplete picture of what these buildings consume across a single state.
That is the part worth sitting with. These reports are required. Nebraska companies submit electricity and water consumption annually to the state Department of Water, Energy, and Environment. State law also allows chunks of those filings to be withheld. So the data exists, the obligation exists, and the public still ends up relying on a formatting error to read it.
Why a toolkit reviewer cares about gallons
My beat is practical. Does the agent framework hold up under load. Does the vector database bill match the estimate. Does the model router actually save money or just move the line item. Water and megawatts feel like a different conversation until you remember that every one of those tools runs on a building like the ones in these filings.
When I test a hosted AI product, I can usually get a latency number, sometimes a token price, occasionally an uptime history. I have never once been handed an environmental cost per thousand requests. Not because it’s unknowable — the filings in Nebraska prove someone is counting — but because nobody has been asked to pass it down the stack.
Here is what I’d want on a tool page, in order of how much I’d trust it:
- Which regions my workload runs in, by default and under failover
- Facility-level water and power figures for those regions, updated annually
- Whether the vendor publishes those figures voluntarily or only when a state forces the filing
- Some honest statement about what they don’t measure
That last one matters most to me. A vendor who says “we don’t track this yet” is more useful than one who posts a sustainability page with no numbers on it. I review tools the same way. The product that admits its gaps is the one I can plan around.
What works and what doesn’t
What works in Nebraska is the reporting requirement itself. Someone decided these facilities should hand over consumption data every year, and that decision is the only reason there’s a story at all. That’s a solid piece of policy design.
What doesn’t work is the redaction layer bolted on top of it. A disclosure regime where the public gets the numbers only when the ink doesn’t take is not a disclosure regime. It’s a filing cabinet with a lock and a bad hinge. The difference between the Lincoln figure and the Papillion figure shows how much interpretation depends on seeing the full set, and partial visibility can mislead worse than none.
I’d extend the same standard to the AI products I test. Selective transparency reads as confidence until you notice which parts got selected. A model provider that publishes benchmark wins and nothing about failure modes is doing a version of the same move, just with better typography.
The Nebraska filings are a reminder that the physical cost of this technology is already being measured, written down, and submitted to a state agency. The measurement isn’t the hard part. Deciding who gets to read it is. And right now, in at least one case, that decision came down to how well a rectangle was placed over a line of text.
If you build or buy AI tools, ask your vendors where the compute lives and what it drinks. The worst outcome is a shrug. The second worst is a number you only got by accident.
🕒 Published: