How much water is AI using in 2026?
There is no measured total for AI's water use in 2026; there never is until reports land the following year. Our running estimate: US data centers will directly consume roughly 26–40 billion gallons in 2026 (Low, cooling-only boundary; site inference from LBNL's measured 2023 baseline and 2028 projection), with AI the dominant growth driver. Per chat query: 0.3–45 mL, depending on the accounting boundary.
Last reviewed: September 2026. This page is updated as 2026 disclosures arrive; revision notes are at the bottom. The 2025 page holds the historical record.
Why this is an estimate, and why the range is wide
Water data runs on a lag. The most rigorous public accounting of US data centers, the Lawrence Berkeley National Laboratory's 2024 report, measured 2023: 176 TWh of electricity and about 17 billion gallons of directly consumed water (LBNL 2024). Corporate sustainability reports covering 2025 began arriving in mid-2026, and we fold them in at our quarterly reviews. So any number for the year currently underway is a projection, and we label it that way.
The range is wide for a second reason: the answer depends on the boundary you draw. Count only cooling water evaporated at the data center (Low, cooling only) and a chat query costs a fraction of a milliliter; that's the boundary behind Google's official 0.26 mL per median Gemini prompt (Google 2025) and OpenAI's ~0.32 mL per ChatGPT query (Altman 2025). Add the water consumed generating the electricity (Mid, + electricity) and a typical query is a few milliliters: about 3 mL on a standard-class model, 7 mL on a frontier one like ChatGPT (site estimate; Siddik 2021 grid factor). Count the full lifecycle (training runs amortized per query, plus chip manufacturing) and Mistral's lifecycle figure is 45 mL per 400-token response (High, full lifecycle; Mistral LCA 2025), while the widely quoted ~519 mL per 100-word email (Washington Post with UC Riverside, 2024) counts cooling plus electricity at a high 2024 GPT-4 energy estimate, not the full lifecycle.
frontier class · 2 Wh per 1,000 tokens (mid benchmark)
The same prompt can be “five drops” or “a bottle of water” depending on what you count: just data-center cooling (Low), the water behind the electricity (Mid), or training and hardware too (High).
Read the methodology16.9gallons
= 63.9 liters
ChatGPT (standard), 25 prompts/day · per year, Mid scenario
Energy
12.8 kWh
Carbon
4.5 kg CO₂e
The 2026 numbers, scenario by scenario
Top-down (whole data centers). LBNL projects US data-center electricity reaching 325–580 TWh by 2028. Interpolating from the 2023 measurement puts 2026 at roughly 265–420 TWh. If direct water consumption scales proportionally (assuming cooling efficiency holds at 2023 levels), US data centers consume roughly 26–40 billion gallons directly in 2026 (Low, cooling-only boundary; site inference). Including the water consumed by power generation (~4.5 L/kWh, Siddik 2021), the all-in figure (direct plus electricity water) is on the order of 340–540 billion gallons (Mid, + electricity; site inference). Both figures cover all data-center workloads; AI is the fastest-growing slice, not the whole pie.
Bottom-up (per query). We estimate ~5 billion AI chat queries per day globally in 2026 (estimated; scaled from OpenAI's disclosed ~2.5B/day in 2025, likely conservative by now). At a blended ~0.8 Wh per query (a site assumption between our standard-class ~0.56 Wh and frontier-class ~1.4 Wh per exchange; operator disclosures run 0.24–0.34 Wh), annualized:
- Low (cooling only): ~0.24 mL/query → roughly 0.4 billion liters (~0.1 billion gallons) per year
- Mid (+ electricity): ~4 mL/query → roughly 7 billion liters (~1.9 billion gallons) per year
- High (full lifecycle): ~24 mL/query → roughly 44 billion liters (~12 billion gallons) per year
Chat is the visible tip: training, image and video generation, coding agents, and enterprise API loads sit underneath these figures and dominate the data-center totals above.
For scale, 26–40 billion gallons of direct consumption is about a year of home water use for 0.9–1.3 million Americans (82 gal/person/day, EPA WaterSense). The national total is modest next to agriculture or power generation; the concern is concentration: new capacity clustering in regions that were short on water before the servers arrived.
What's different about 2026
Three trends distinguish this year from the 2025 picture:
- Efficiency disclosures are now table stakes. After Google's and Mistral's 2025 publications, per-prompt figures have become a competitive claim. Expect more operator numbers, and remember they will almost always use the narrowest boundary (Low, cooling only). That isn't deception; it's just the part of the system they own.
- Reasoning models pull the per-query average up. Models that "think" before answering use several times the energy of a standard reply (this site's reasoning class runs 3–15 Wh per 1,000 tokens versus ~0.8 for standard chat, a band built on Jegham et al. 2025's reasoning-model benchmarks). More water per query at every boundary.
- The build-out is front-loaded. The construction wave announced in 2024–2025 starts drawing water when facilities energize, which is happening through 2026–2028. This year's totals reflect decisions made two years ago.
What we'll revise as 2026 data lands
This page is built for refresh. Specifically, we will update when: corporate environmental reports for calendar 2025 are folded in (they began arriving in mid-2026; Google's 2024 figure was about 7.7 billion gallons across its data centers, and its 2025 figure will recalibrate our trajectory); any LBNL or successor federal accounting update appears; new LCAs join Mistral's; and operators publish revised per-prompt disclosures. When the audited 2026 numbers eventually land in 2027, this page will state them and show what our estimate got wrong: in which direction, and why.
Revision log: June 2026: initial estimate published. September 2026: corrected Google's 2024 figure (about 7.7 billion gallons; the 6.1 billion was 2023), relabeled the ~519 mL email figure as cooling plus electricity, and disclosed the ~0.8 Wh per query as a blended site assumption.