← Privacy policy

Droplet methodology

Model version 0.1.0 · Last updated: August 8, 2026

View the Droplet source code on GitHub

How the estimate is calculated

Droplet estimates output tokens from the character length of each completed assistant response. It then estimates the energy for that response and converts the energy estimate into on-site and electricity-generation water estimates.

tokens_out    = char_count / 4
E_Wh          = (Wh_base + tokens_out × Wh_per_token) × reasoning_multiplier
water_onsite  = E_Wh × WUE_dc
water_grid    = E_Wh × EWIF_grid
water_total   = water_onsite + water_grid

On-site and grid water

On-site water is used directly for data-center cooling. Grid water is used upstream to generate the electricity that powers a request. “On-site cooling only” shows the direct estimate; “including electricity generation” combines both values.

Coefficients

CoefficientLowMidHighUnit
Wh_base0.020.050.30Wh / request
Wh_per_token0.00020.000630.009Wh / token
reasoning_multiplier1.5310×
WUE_dc0.21.11.9mL / Wh
EWIF_grid1.04.56.5mL / Wh

The low, mid, and high ranges reflect uncertainty across hardware efficiency, response length, data-center water use, and electricity-generation water use.

Research anchors

Our low estimate follows Google’s production measurements of on-site cooling water. Our high estimate follows Li et al., which adds the water consumed generating the electricity. The truth for any given prompt sits somewhere between, and depends on which data center served you—something neither we nor you can know.

  1. Elsworth, C., Huang, K., Patterson, D., Schneider, I., Sedivy, R., Goodman, S., Townsend, B., Ranganathan, P., Dean, J., Vahdat, A., Gomes, B., & Manyika, J. (2025). Measuring the environmental impact of delivering AI at Google Scale. arXiv:2508.15734.

    arXiv abstract · Google PDF

  2. Li, P., Yang, J., Islam, M. A., & Ren, S. (2025). Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models. Communications of the ACM, 68(7), 54–63.

    Communications of the ACM (peer-reviewed version) · arXiv preprint

Known limitations