'Thirsty' ChatGPT uses four times more water than previously thought

Original Article ↗ Hacker News Discussion ↗

Scope of the Article / Framing

  • Many argue the piece is really about generic data center water use, with “ChatGPT” in the title mainly for clicks.
  • Some see it as a targeted “hit piece” driven by fear of LLMs or innumeracy around big-sounding quantities.
  • Others push back on dismissiveness, arguing that resource allocation and new, rapidly scaling loads (like LLMs) are legitimate topics.

Water Use, Allocation, and Externalities

  • Key point: water doesn’t vanish; cooling often uses evaporation, returning water to the atmosphere, not sewers.
  • Critics respond that the issue is where and when water is used: data centers may compete with households, agriculture, and ecosystems, especially in water‑stressed regions.
  • Concern that “water-positive” pledges may restore water in different locations than where it was withdrawn.
  • Some argue water use should be priced to include local externalities; unclear if this is happening adequately.

Scale and Comparisons

  • Several commenters emphasize scale: data center use is claimed to be tiny relative to industrial usage, municipal pipe leaks, agriculture, and lifestyle choices (e.g., lawns, avocados, T‑shirts, theme parks).
  • Others counter that “everyone else wastes more” doesn’t address whether new high-consumption uses are justified.

Value of LLMs vs. Waste

  • Strong disagreement on utility:
    • Pro-LLM comments cite large personal and professional productivity gains (coding, research, fraud detection, learning).
    • Skeptics see LLMs as hallucination-prone, black-box tools whose output is often low-value or replaceable by search.
  • Some frame LLMs as comparable to or better than other high-consumption tech (e.g., Bitcoin); others see both as wasteful.

Energy Shaming and Ethics

  • Debate over whether “energy/water shaming” is useful or selectively applied only to new tech instead of incumbents (offices, casinos, golf courses, theme parks).
  • Underneath is a larger question: in a finite-resource world, should unnecessary or luxury uses (including LLMs) be socially or politically constrained?

Technical and Policy Notes

  • Distinction between evaporative cooling vs once‑through systems; some confusion about what counts as “consumption.”
  • Training is described as more intensive per run but small compared to total long‑term inference usage.
  • Ideas raised: siting data centers near seas or in deserts with solar and desalination; pairing waste heat with beneficial uses.