MTS
MTS Editorial No. 02— AI water consumption · May 2026

How much water does your AI prompt drink?

The viral claim — that ChatGPT consumes 500 mL of water per query, or a “whole bottle per email” — is wrong by about a hundredfold. The actual per-query cost depends on the model, the hardware, and the data center it runs in. Pick the inputs below and see for yourself.

§ Live calculatorPick four inputs · result updates live
Model
Hardware
Data center
Query type
Water per query · mid-case
5.7 mL
GPT-5.5 · Azure US East (modern)
Low0.25 mL
Mid5.7 mL
High114 mL
Energy / query
2.784 Wh
≈ a high-efficiency LED bulb running for a few minutes
Direct cooling
1.3 mL
22% of total · hits the data-center cooling tower
Scope-2 power
4.5 mL
78% of total · consumed at the generating plant

That's about

  • 1 sip of water5.3 prompts
  • 1 toilet flush (modern)1,062 prompts
  • 1 cup of coffee23,217 prompts
  • 10-minute shower16,576 prompts
  • 1 hamburger437,700 prompts
  • 1 cotton t-shirt473,094 prompts
  • 1 pair of jeans1,193,949 prompts
Inputs & assumptions

GPT-5.5 — active params 80200400 B; total 2500 B; arch moe. Released Apr 23, 2026; OpenAI's first new pre-train since GPT-4.5 ('Spud'). SemiAnalysis-implied floor: above V4-Pro (49B / 1.6T).

NVIDIA H100 SXM — peak 989 TFLOPS BF16, 700 W TDP, 3.35 TB/s HBM.

Azure US East (modern) — PUE 1.18; on-site WUE 0.45 L/kWh; scope-2 WUE 1.6 L/kWh; serving overhead 5×. PJM grid; mixed gas/nuclear/coal; modern adiabatic cooling. WUE_direct mid 0.45 anchored on Microsoft FY2024 fleet-wide ~0.49 L/kWh.

Typical chat turn3002000 input tokens, 1501000 output, reasoning multiplier 11×.

§ The argument in three beats
01 · The myth

It’s a sip, not a bottle.

For a typical chat turn on a typical 2026 frontier model on a modern hyperscaler region, the mid-case water cost is roughly 0.2 to 3.3 mL — about a sip. (Worst-case sites like xAI Memphis push the high end up to ~9 mL.) The 500 mL viral figure requires GPT-3 hardware, evap cooling, and reading the source paper as “per query” when it actually said “per 10–50 queries.”

Read · the 500 mL myth →
02 · The lever

Region dominates the answer.

The biggest single lever isn’t the model or the hardware — it’s the data center’s cooling system and the local grid mix. The same query on a clean PPA-matched site uses ~94% less water than on a fossil + evap site.

Read · cooling, explained →
03 · The honest part

Local stress is real. National alarm isn’t.

All consumer + B2B AI combined uses roughly 0.001% of US daily freshwater on our mid-case estimate. But Memphis, Phoenix, and Newton County are facing real local conflicts — about siting and permitting, not your prompt count.

Read · local vs national →