Finding · July 18, 2026

Corrections: two water figures were wrong by 1,000x, and we deleted four of our own estimates

Building year-over-year charts surfaced a unit error in two seeded water figures and four stale estimates from an earlier model. What was wrong, how we found it, and what changed.

In July 2026 we added year-over-year charts to the entity profiles. Within an hour of rendering them, the charts had exposed two kinds of errors in our own records. This post documents both, because an observatory that grades companies on transparency should hold itself to at least the same standard.

Error one: two water figures off by a factor of 1,000

Google's FY 2023 water withdrawal was stored as 21,183,000 megaliters. That is 21 trillion liters, roughly fifteen times the annual water supply of New York City, and obviously impossible. The correct figure, consistent with Google's own environmental reporting, is 21,183 megaliters (about 21.2 billion liters). Microsoft's FY 2023 entry had the same defect: 6,400,000 ML stored where 6,400 ML was meant. Both were unit errors introduced when the figures were first seeded, liters converted into megaliters incorrectly by a factor of one thousand.

The trend chart made the error impossible to miss: it showed Google's water use falling from 21,183 GL to 44.7 GL in one year, a decline of 99.8%. Nothing in the physical world does that. Both rows have been corrected in place, and each carries a permanent note recording the correction and its date.

Error two: four estimates from a model we no longer stand behind

The charts also showed Anthropic's energy use falling 76% and Equinix's falling 68% between FY 2024 and FY 2026. Neither decline was real. The FY 2024 figures came from an early version of our estimation model, before it was calibrated against companies' disclosed electricity figures. That early model put Anthropic at 31 TWh for 2024, which would have made a startup's footprint larger than Google's, and Equinix at three times its own disclosed consumption.

A superseded estimate is not history, it is noise. Keeping those rows would have manufactured dramatic downward trends that never happened, so we deleted all four (energy and water for both companies) rather than let a model revision masquerade as a change in the world. The current figures come from the calibrated model, which reproduces disclosed electricity totals for Google, TSMC, NVIDIA, and Equinix to within 15 to 20 percent.

What stays the same

Our correction policy is unchanged and now easier to audit: corrections are applied in place with a dated note on the affected row, and every record change appears on the updates page and in the RSS feed. The charts that caught these errors are now part of the site, which means future inconsistencies of this shape will be visible to every reader, not just to us.

Every figure referenced here is in the observatory dataset with its confidence tier and source. Methods are documented on the methodology page. Corrections and record changes are logged on the updates page. If you can show a figure here is wrong, write to research@openbutterfly.com and we will correct it publicly.