Scope-2 Water: The Power Plant Story
Most of what gets called "AI water" is not water that an AI data center touches. It is water that a coal or natural-gas power plant somewhere on the same grid consumed to generate the electricity the data center bought. This is "scope-2" or "indirect" water — and it dominates the totals.
The basic accounting
Every kWh of electricity carries a water footprint, set by the cooling needs of the generation source. Median operational consumption (not withdrawal) values for plants with recirculating cooling towers, from Macknick et al. 2012 (NREL TP-6A20-50900):
| Source | Water consumed (gal/MWh) | mL per kWh |
|---|---|---|
| Coal subcritical | ~479 | ~1,810 |
| Natural gas CC | ~205 | ~780 |
| Nuclear | ~672 | ~2,540 |
| Hydro (reservoir evap) | ~4,491 | ~17,000 |
| Solar PV utility | ~1 | ~4 |
| Wind | ~0 | ~0 |
A note on units: many secondary write-ups quote withdrawal numbers — water cycled through the plant — and label them as consumption. Withdrawal can be 10–50× higher (especially for once-through cooling) but most of it returns to the river. Consumption is what evaporates and leaves the watershed, and is the right metric for "how much water did this query use." See Macknick et al. for the full review.
The Ren et al. 2023 paper, applied to a US-average GPT-3 query, broke out: 2.2 mL on-site cooling + 14.7 mL off-site at the power plant = 16.9 mL total. That's roughly 87% of the per-query water footprint coming from generation, not from the data center itself.
Why this matters more than anything else on this wiki
Three implications follow, all of which destabilise the "AI is uniquely thirsty" story:
1. Most "AI water" is fungible with all other electrical loads
The water consumed at a coal plant to power 1 kWh of GPU inference is identical to the water that would have been consumed for 1 kWh of streaming Netflix, running an air conditioner, or smelting aluminum. There is nothing distinctively AI about it. If the editorial argument is "AI is uniquely water-intensive," scope-2 dominance falsifies that — it's "electricity is water-intensive on a fossil grid, and AI uses electricity."
2. Decarbonising the grid eliminates ~80% of the AI water footprint automatically
A coal-to-solar transition reduces water per kWh by a factor of ~500. A coal-to-wind transition reduces it by ~∞. The single largest lever on AI's water footprint is one nobody has to pull at the data center. (OECD.AI on the indirect dominance.)
This is exactly what Altman and OpenAI have been pushing publicly: "move very quickly to nuclear, wind, or solar." (CNBC, Feb 2026.)
3. The on-site/off-site cooling trade-off can run backwards
This is covered in cooling_explained.md, but the headline: switching from evaporative cooling to air cooling reduces direct water by ~100% and increases electricity demand by ~10%. On a fossil-heavy grid, that means a net increase in total water consumed — the savings on-site are more than wiped out by the extra fossil burn upstream.
The "use less water" rhetoric, taken literally, can make the problem it claims to address worse.
Disclosure asymmetry
Most major tech companies (MSFT, GOOG, AWS) now disclose direct on-site water use in their sustainability reports. Almost none disclose scope-2 water. That asymmetry is a real critique — Ren's central policy ask is better reporting, not "stop using water" — and the editorial should not paper over it. (See also OECD.AI on disclosure gaps.)
The honest framing: the per-query numbers companies put in press releases are direct-only and are the small piece. Independent analysts who include scope-2 land 5–10× higher, and that's still tiny compared to comparison footprints.
Open question for the piece
Is there a clean visualization that shows the same physical water — moving from agriculture (70%+, USGS) to thermoelectric power (~40% of US withdrawals; ~3% of US consumption) to data centers (~0.2%, Masley) to AI specifically (~0.008%, Masley)? A nested treemap or sankey would make the "AI is a tiny slice of a tiny slice" framing immediate.
What the editorial should not try to argue
- "Scope-2 water doesn't count." It does count. It just doesn't make AI uniquely culpable.
- "OpenAI's 0.32 mL number is the right one to anchor on." It's the right number for direct on-site water on a clean grid. For total environmental footprint, the 5–17 mL range is more honest.
Sources cited on this page
- Macknick et al. (2012), Operational water consumption and withdrawal factors — NREL TP-6A20-50900 / IOPscience — per-source water consumption table
- Li, Yang, Islam, Ren (2023), Making AI Less "Thirsty" — arXiv:2304.03271 — 2.2 mL direct + 14.7 mL scope-2 = 16.9 mL US-average GPT-3 query; disclosure policy ask
- OECD.AI, How much water does AI consume? — oecd.ai — indirect dominance and scope-2 disclosure
- Andy Masley (Oct 2025), The AI water issue is fake — andymasley.com — US-aggregate shares
- CNBC (Feb 2026), Sam Altman defends AI resource usage — cnbc.com — Altman on grid decarbonisation
- Full bibliography: sources.md.