Every AI researcher knows the feeling of staring at a model whose internals are hidden from you. You can query it, measure its outputs, infer things about its structure, but you cannot see inside. So you probe. You find the edge cases, the places where the abstraction leaks, and you learn more from a single failure than from a thousand clean responses.
Nebraska’s data center reporting regime just had one of those failures. State law requires data centers to file annual reports on electricity and water consumption, and state law also lets large chunks of those filings be withheld as confidential. Someone applied the redaction improperly. The numbers came through.
What came through is specific. 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, annual water consumption totaled approximately 765 million gallons. The Fireball Group LLC facility in Papillion accounted for the largest share by a wide margin, at roughly 548 million gallons in 2025.
What a Single Leaked Row Tells You
I work on agent architectures, which means I spend most of my time thinking about systems whose behavior emerges from components I can only partially observe. The methodology transfers here directly. One leaked data point is not just one data point. It is a constraint on every estimate you had been carrying.
Consider the Lincoln numbers against the Papillion numbers. Lincoln uses about 13 million gallons a year. Papillion uses about 548 million. That is a ratio of roughly 42 to 1 between two facilities operated by the same company in the same state. These are not two instances of the same thing scaled slightly differently. They are architecturally distinct systems, almost certainly differing in cooling design, in workload mix, or in both.
That spread matters for anyone trying to reason about the resource cost of AI inference. The common framing treats a query as having a water cost and an energy cost, as though those were properties of the model. They are not. They are properties of the facility the model happens to be running in, and the variance between facilities is enormous. A per-query water figure derived from one site tells you very little about another.
The Six-Facility Picture
Aggregate 765 million gallons across six sites and the arithmetic gets interesting. Papillion alone is 548 million of that total, which leaves about 217 million gallons spread across the remaining five. Lincoln’s 13 million sits well below that average. So one facility is doing most of the water consumption while the others operate at a fraction of its draw.
This is the kind of distribution that should make anyone cautious about extrapolation. If you had sampled only Lincoln, you would have badly underestimated the regional footprint. If you had sampled only Papillion, you would have badly overestimated it. Either single-site sample would have produced a confident, wrong answer.
And this is Nebraska. Six facilities in one state, visible only because a redaction tool failed. The same opacity applies nearly everywhere else, which means most published estimates of AI infrastructure resource use are built on samples of exactly this quality, drawn from exactly this kind of distribution.
Why Observability Is an Architecture Problem
In agent design, we treat observability as a first-class requirement, not an afterthought. A system you cannot instrument is a system you cannot debug, cannot optimize, and cannot safely scale. You build in logging, tracing, and metrics from the start because retrofitting them means reconstructing behavior from accidents.
The physical substrate of AI is currently running without that instrumentation, by design. Reports exist. The reports are filed. The contents are then withheld under confidentiality provisions, which converts a measurement system into a nominal one. The data is collected and then made unavailable to anyone who might use it to reason about the system.
That is an architectural choice, and it has the predictable consequence: the only reliable public information arrives through error. The 52.65 megawatt figure for Lincoln is not public because a process was designed to make it public. It is public because a process designed to keep it private broke.
What I Would Want Instead
The useful version of this is not a leak. It is a standard. Facility-level reporting of energy draw and water consumption, published on a consistent schedule, in a consistent format, with enough granularity to distinguish cooling architectures from each other.
That would let researchers do actual work: model the relationship between workload type and resource draw, compare cooling designs empirically, and build per-query estimates that mean something. Right now we are reverse-engineering a continent-scale system from redaction failures in one Midwestern state, and calling the result analysis.
The Nebraska filings showed what visibility looks like. One peak demand number, one annual water figure, one facility. It was enough to reframe a question. The absence of the rest is not a gap in the data. It is the data.
đź•’ Published: