Picture a state office in Lincoln, Nebraska. A PDF lands in the annual reporting queue, the same way it does every year. Electricity consumption, water consumption, the usual boxes. The numbers are there, dutifully covered by black rectangles because state law lets data center operators treat that information as proprietary. Then somebody opens the file, selects the text underneath the black box, and copies it out. The redaction was decoration. The numbers were always right there.
That is how the public learned that 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. It is also how we learned the bigger number: across six Nebraska data centers, annual water consumption came to approximately 765 million gallons. Google’s Fireball Group LLC facility in Papillion led the pack at roughly 548 million gallons in 2025.
I review AI tools for a living. I poke at agents, benchmark them, and complain when the demo does not match the product. So let me say the uncomfortable part plainly: the most important spec sheet in this industry is the one nobody publishes, and we only got a peek at it because of a copy-paste failure.
The transparency problem is the actual story
Forget the gallons for a second. The mechanism here deserves more attention than the measurement. Nebraska requires these facilities to report electricity and water use to the state Department of Water, Energy, and Environment. Good. That is the kind of accountability you want when a single building draws tens of megawatts off a shared grid. But the same framework lets operators redact the figures as confidential business information. So the reporting requirement exists, and the public benefit of it mostly does not.
Which means the honest description of the status quo is this: the data gets collected, the data gets hidden, and the only reason anyone outside the filing office saw real numbers in 2026 is that a black rectangle is not encryption. If your compliance regime depends on correctly applying a graphics tool, you do not have a compliance regime. You have a formality.
I keep thinking about how we talk about AI tools on this site. We argue over token costs, latency, context windows, whether an agent actually completes the task or just claims it did. We demand numbers. We get annoyed when a vendor publishes a benchmark without the methodology. And yet the physical layer underneath every one of those tools — the buildings, the cooling, the draw on local utilities — operates with less disclosure than a mid-tier SaaS pricing page.
What the numbers actually tell us
The distribution is more interesting than the total. Of that roughly 765 million gallons across six facilities, Papillion’s Fireball Group accounted for about 548 million. That is the large majority of the state’s reported data center water use sitting in one facility. Lincoln’s 13 million gallons, by comparison, is a rounding error in the same ledger.
That spread matters because it undercuts the lazy version of this debate on both sides. The “data centers are drying out your town” framing does not hold when one site uses 13 million gallons and another uses forty times that. Neither does the industry-friendly “it’s all fine, it’s just a building” line, when a single campus is pulling hundreds of millions of gallons a year. The answer depends entirely on which facility, which cooling design, which local water system. Which is exactly the kind of nuance you cannot have without per-site numbers.
Same with the 52.65 megawatt peak demand figure from Lincoln. On its own it is a data point. In context — alongside grid capacity, other industrial load, and planned expansion — it becomes something a city council or a ratepayer can actually reason about. Without disclosure, nobody gets to reason about anything. They just get a press release about jobs and a utility bill that moved.
Where this leaves the rest of us
I am not going to pretend one Nebraska filing error rewrites how AI infrastructure gets built. It does not. What it does is demonstrate, cleanly, that the information exists, that it is specific, that it is per-facility, and that the only thing standing between it and the public is a legal exemption plus a layer of digital ink.
If you build on these models, you are a customer of that infrastructure whether you think about it or not. Every agent run has a physical footprint somewhere, and right now you cannot see it unless a clerk fumbles a PDF. I would rather know. Not because knowing settles the argument about whether these facilities are worth it, but because you cannot have that argument honestly when one side has the numbers and the other side has a black box.
The redaction failed. That is the one piece of good news in the whole story, and it was an accident.
🕒 Published: