Water is the metric everyone skips. Here's why that's the bigger gap.

Carbon gets the headlines. Water doesn't — and that's backwards for AI inference specifically.

A data center's water draw is direct and physical: cooling towers evaporate water to reject heat, and that number doesn't average out across a grid the way carbon intensity does. Texas data centers are projected to use roughly 49 billion gallons in 2025, climbing toward 399 billion by 2030. That's not a marginal footnote — it's a resource constraint showing up in the same regions already fighting over water rights.

Yet almost no inference-level accounting includes it. Carbon dashboards report kWh and CO2e; water shows up, if at all, as a single company-wide ESG line once a year — disconnected from any specific workload, model, or request.

The reason is structural, not laziness: water intensity varies by cooling architecture (evaporative vs. closed-loop vs. air), by region, and by time of year — more variables than grid carbon intensity, with far less public data to draw from. So most platforms just... don't measure it, and the silence gets read as "negligible." It isn't.

Same principle as the carbon boundary problem: an unstated omission isn't a finding, it's a gap wearing the shape of a non-issue. If a sustainability claim covers carbon but is silent on water, that's not a footprint — it's half of one.

This is the same floor we're building for carbon: not a promise of exact liters per query, but a disclosed method for estimating water draw per inference, tied to the cooling architecture and region where that inference actually ran.