I mapped estimated water use across 30 major AI/cloud data centers

Join the discussion

• Hacker News
• Made this after getting curious how the water-use numbers thrown around in AI news articles actually stack up site-by-site. A few notes:

  What it shows: a running estimate of global AI/data-center water use, a map of 30 real campuses (Google, Amazon, Microsoft, Meta, Oracle, Apple, Alibaba) sized by estimated annual water draw, and a comparison chart against things like golf courses, fast fashion, and fossil fuel plants on a log scale.

  Data sources: per-site figures are triangulated from sustainability reports, utility/permit filings, and known cooling tech + climate where companies don't disclose (most don't). The global baseline is anchored to Lawrence Berkeley National Lab's 2024 Data Center Energy Usage Report, linked in the site's Methodology section.

  Tools: React + D3.js for the map, all client-side, no backend.

  Caveat I want to be upfront about: these are order-of-magnitude estimates, not audited numbers, happy to take corrections if anyone has better sourcing on specific sites!

  https://www.thirstymachines.com/

  by senazadeh
• This is good. Have you sent this to non-profits or local community leaders who have this on their docket? The data will give more context and awareness to the citizens they are working with.
  by Zie_Mordecai
• Metrics for this?

  OPS/FLOPS/TOPS/QOPS: OPS/kWhr, OPS/liter_water,

  From https://www.thegreengrid.org/ , whose board includes many industry folks:

  WUE: Water Usage Effectiveness: https://en.wikipedia.org/wiki/Water_usage_effectiveness

  GPUE: Green Power Usage Effectiveness: https://en.wikipedia.org/wiki/Green_Power_Usage_Effectivenes...

  by westurner
• Data sources; Data quality:

  - Is there a suggested bibtex citation for this analysis?

  - BibTeX in git for the data and the estimates can be referenced with citation identifiers with various static site build tools. Schema.org/Dataset and ScholarlyArticle JSON-LD is probably easier with React. It should be possible to generate BibTeX from JSON-LD (e.g. with citeproc-js and n3.js or rdflib.js or solidjs/react-solid-state or a different RDFJS solution that can template BibTeX).

  - DVC is one way to check data into git, and to evaluate sensitivity to data quality and specificity

  Additional features probably worth tracking:

  - Zero Water facility?

  - Types of thermal fluid in use: Water,

  - Heat recovered : Heat and water forfeited to evaporative cooling

  - Water egress: % purple pipe water, % steam

  by westurner
• These data centers need to be responsible for their own water (not other people's drinking water) and pay for or provide their own electricity (not force increases in other people's bills).

  If they want "hand outs" and "welfare" from the state (aka subsidies and tax exemptions), then the state should partially own them or get a slice of any profits.

  by baranul