Key Numbers
The good faith range for "water per ChatGPT query" spans about 50× from low to high. Picking a single number is a values choice — what to include (training? construction? scope-2?), what generation of hardware, what location. The honest move is to publish all the headline figures with their assumptions, which is what this page does.
Per query / per conversation
| Estimate | Value | Scope | Year | Source |
|---|---|---|---|---|
| OpenAI (Sam Altman) | 0.000085 gal / 0.32 mL | Direct on-site, average global query | Jun 2025 | Altman, Gentle Singularity / DCD |
| Goedecke (independent) | ~5 mL | Per conversation, modern model, includes both scopes | Oct 2024 | seangoedecke.com |
| Ren et al. (corrected read) | 10–50 mL | Per medium-length GPT-3 response, Azure 2022 | 2023 | arXiv:2304.03271 |
| Ren et al. (US average) | 16.9 mL | Per GPT-3 query: 2.2 mL on-site + 14.7 mL off-site | 2023 | arXiv:2304.03271 |
| Viral WaPo / Hao framing | ~500 mL | "Per email" — see the_500ml_myth.md | 2023+ | WaPo / Hao Empire of AI — Masley critique |
The Altman number is unaudited and unclear about deep-research / Code Interpreter / image-gen requests. The Goedecke number is conservative and methodologically transparent. The Ren numbers are the only peer-reviewed figures and now describe a model that is two generations old.
For editorial purposes, 5 mL per query is the safest defensible "high" number for a typical text query in 2026.
Aggregate scale (United States)
| Quantity | Value | Source |
|---|---|---|
| All US data centers, daily freshwater (2023) | 200–250 million gal | Masley / EESI |
| All US data centers, daily freshwater (2021) | 449 million gal | EESI |
| Total US daily freshwater consumption | ~132 billion gal | USGS |
| Data centers as share of US freshwater (2023) | ~0.2% | derived from EESI / USGS |
| AI-specific share of US freshwater (2023) | ~0.008% | Masley |
| AI-specific share of US freshwater (2030 proj.) | ~0.08% | Masley |
| US agriculture share of US freshwater | ~70–80% | USGS |
The 2021 figure is higher than 2023 not because data centers shrank, but because the EESI methodology counted indirect (power-plant) water in the 2021 figure and direct only in 2023 — yet another reason to always check what scope is being measured.
Aggregate scale (regional)
| Region | Value | Source |
|---|---|---|
| Phoenix-area data centers (~60 sites) | 177 million gal/day | Stanford & The West |
| Phoenix metro agriculture share of water | 86% | Stanford & The West |
| Texas data centers, projected 2025 | 49 billion gal/year (~134M gal/day) | HARC / U Houston (PDF pending — see sources.md) |
| Texas data centers, projected 2030 | 399 billion gal/year | HARC / U Houston (PDF pending) |
| US data centers in high water stress | ~40% of facilities | Lincoln Institute / WRI Aqueduct |
The Texas 2030 projection is the most aggressive credible aggregate number in the discourse. Even at 399 B gal/year — 1.1 B gal/day — it is roughly 0.8% of US daily freshwater consumption, and concentrated in Texas where the agricultural baseline is also enormous.
Power-plant (indirect / scope-2) water intensities
Median operational water consumption (not withdrawal) for plants with recirculating cooling towers, from Macknick et al. 2012 (NREL TP-6A20-50900):
| Generation source | Water consumed (gal / MWh) | Notes |
|---|---|---|
| Coal subcritical | ~479 | Most water-intensive thermal |
| Natural gas combined-cycle | ~205 | Modern dispatchable workhorse |
| Nuclear | ~672 | Once-through withdrawal much higher; consumption similar |
| Hydroelectric (reservoir evap) | ~4,491 | Wide range 0–18,000; allocation contested |
| Solar PV utility | ~1 | Effectively negligible |
| Wind | ~0 | Effectively zero operational water |
Two implications: (1) most of the "AI water footprint" is actually a coal / gas footprint plus, in some grids, hydro reservoir evaporation; (2) decarbonising the thermal share of the grid eliminates the bulk of the apparent AI scope-2 water problem without anyone touching a data center. Note the gap between withdrawal (water cycled through) and consumption (water evaporated): older popular write-ups often quote withdrawal numbers (10–50× higher), which is one source of the inflated AI-water headlines.
Cooling tech operational tradeoffs
| Cooling approach | On-site water | Electricity overhead | Net climate effect (current grid) |
|---|---|---|---|
| Evaporative (towers) | High | Baseline | Less CO₂ |
| Closed-loop liquid | Very low | ~5% | Slightly more CO₂ |
| Air cooling | ~Zero | ~10% | Notably more CO₂ + indirect water |
Reducing on-site water at a data center on a fossil-heavy grid generally increases total water consumed, by shifting it to the power plant.
Comparisons (full table on its own page)
For a few headline reference points (Water Footprint Calculator, FoodPrint): a hamburger is ~660 gal of water; a t-shirt ~713 gal; a pair of jeans ~1,800 gal; a single almond ~1 gal; a 10-minute shower ~25 gal. See comparisons.md for the full set with prompt-equivalents.
Sources cited on this page
- Altman, The Gentle Singularity (Jun 2025) — 0.32 mL / 0.34 Wh per query
- Goedecke, Talking to ChatGPT costs 5 mL (Oct 2024) — 5 mL/conversation independent
- Ren et al., Making AI Less "Thirsty" (2023) — 10–50 mL/query GPT-3 + 16.9 mL US-avg breakdown
- Masley, The AI water issue is fake (Oct 2025) — 0.008%/0.08% national share figures
- Masley, Empire of AI is wildly misleading — Hao critique
- EESI, Data Centers and Water Consumption — 200–250 M gal/day (2023), 449 M gal/day (2021)
- USGS, Water Use in the United States — 132 B gal/day national; agriculture share
- Stanford & The West — Phoenix-metro figures
- Lincoln Institute, Data Drain — 40% in high-stress siting
- Macknick et al. (NREL TP-6A20-50900) — per-source water consumption
- Water Footprint Calculator and FoodPrint — comparison footprints
- Full bibliography: sources.md.