Every water and carbon number you see reported for data center operations is, more often than not, a market-based average — blended across a grid region, smoothed across a reporting period, often built from utility-level estimates rather than the facility itself.
That's not malicious. It's just what the existing frameworks (like ISO/IEC 30134) were built to produce: a defensible average, not a traceable measurement.
The problem shows up the moment you try to do per-inference accounting. If your water/carbon number can't tell you which basin or which grid node it came from, it's not measurement — it's a regional estimate wearing a specific-looking decimal point.
At CarbonLayer we split these into two explicit categories instead of blending them:
Modeled — projections built on climate-scenario water-gap data (we work with 5 climate models × 2 warming scenarios), useful for forward risk, explicitly not a claim about what happened
Measured — facility-level, tied to the actual site an inference workload ran on, with regional water-stress scoring (via WRI Aqueduct) layered in
Reporting these as one blended number is how you get compliance theater: technically defensible, operationally meaningless. Reporting them separately is more honest and, frankly, more useful — a facility operator or regulator can act on "here's our modeled exposure vs. here's what we actually measured" in a way they can't act on a single smoothed average.
We think this distinction becomes mandatory as soon as regulation catches up — the two legislative failures we've been tracking (Texas's water-reporting rider, Rhode Island's energy-benchmarking bill) both died in part because there was no measurement standard underneath the mandate. You can't legislate what you haven't defined how to measure.
Top comments (0)