# Water and Energy Consumption Research: USA 2024-2025

Research conducted: 28 November 2025
Updated: 3 July 2026 - see "July 2026 Update" at the end of this document for corrections and new data. Where the two conflict, the July 2026 section supersedes the original notes below.
Updated: 2 September 2026 - see "September 2026 Update" (Knowable Magazine review) at the end of this document.
All data sources from 2024-2025 unless otherwise noted.

---

## WATER CONSUMPTION

### Golf Courses

**Annual Water Consumption:**
- **547.5 billion gallons per year** (2024 estimate)
  - Calculated: 1.5 billion gallons/day × 365 days
  - Represents approximately 0.5% of total daily U.S. water usage
  - Sources: [USGA Environmental Benefits](https://www.usga.org/content/usga/home-page/course-care/green-section-record/62/issue-06/the-environmental-benefits-of-golf-courses.html), [USGA Water Resource Center](https://www.usga.org/content/dam/usga/pdf/Water%20Resource%20Center/how-much-water-does-golf-use.pdf)

**Alternative Historical Estimates:**
- 759.2 billion gallons/year (based on 2003-2005 data of 2.08 billion gallons/day)
- 606 billion gallons/year (based on 2013 data)

**Context:**
- Golf courses reduced water usage by 29% between 2005 and 2020
  - Source: [GCSAA Water Usage Survey](https://www.gcsaa.org/who-we-are/media/news-release/2022-news-releases/2022/07/26/golf-courses-reduce-water-usage-by-29-percent-according-to-national-survey)

**Number of Golf Course Users:**
- **28 million on-course golfers** in 2024 (playing traditional golf on actual courses)
  - Largest single-year increase (1.5 million) since 2000
  - Seventh consecutive year of growth
  - Sources: [PGA By the Numbers 2024](https://www.pga.com/story/by-the-numbers-golf-in-2024), [LINKS Magazine](https://linksmagazine.com/5-things-to-know-about-the-state-of-recreational-golf-in-2025/), [National Golf Foundation](https://www.ngf.org/short-game/golfs-state-of-industry-in-3-minutes/)

- **47.2 million total golf participants** (including off-course activities like driving ranges, simulators)
  - Sources: [Andalucia Golf](https://andaluciagolf.com/en/more-than-47-million-americans-played-golf-in-2024/), [TeachMe.To](https://teachme.to/blog/record-breaking-year-for-golf-in-america-the-annual-golf-report-cm990ebqt00289eyzkgxjf468)

**Additional Golf Statistics:**
- 15,962 golf courses at 13,952 facilities (2024)
  - Source: [NGF Golf Facilities Report](https://www.ngf.org/member-publication/golf-facilities-in-the-u-s-2024/)
- 545 million rounds played (2024 record)
  - Source: [USGA Golf Scorecard 2024](https://www.usga.org/content/usga/home-page/articles/2024/12/golf-scorecard-2024-statistics-golf-recreational-game.html)

### Water Waste from Leaks

**Total Annual Water Waste:**
- **34.2 trillion gallons per year** (combined infrastructure + residential)
  - Calculated: 33.3 trillion + 0.9 trillion

**Infrastructure Losses (Public Water Systems):**
- **33.3 trillion gallons annually** (126 billion cubic metres)
  - Represents 14-18% of all treated water in the USA
  - Results in over $187 billion in lost revenue annually
  - 240,000 water main breaks occur annually, costing $2.6 billion in repairs
  - Nearly 20% of water mains have exceeded useful life
  - Sources: [ASCE Infrastructure Report Card 2025](https://infrastructurereportcard.org/cat-item/drinking-water-infrastructure/), [Anytime Plumbing Stats 2025](https://anytimeplumbing.net/u-s-water-usage-plumbing-statistics/)

**Residential Household Leaks:**
- **900 billion gallons annually nationwide**
  - Equivalent to annual water use of 11 million homes
  - Average family wastes 9,400 gallons/year from household leaks
  - 10% of homes have leaks wasting 90+ gallons per day
  - Sources: [EPA WaterSense Statistics](https://www.epa.gov/watersense/statistics-and-facts), [Anytime Plumbing Stats 2025](https://anytimeplumbing.net/u-s-water-usage-plumbing-statistics/)

**Commercial/Business Leaks:**
- **Estimated 340 billion gallons annually**
  - Calculated: Commercial sector accounts for 17% of public water supply withdrawals (5.66 trillion gallons)
  - Leaks account for approximately 6% of commercial facility water use
  - Sources: [EPA WaterSense Types of Facilities](https://www.epa.gov/watersense/types-facilities), [EPA Fix a Leak Week](https://www.epa.gov/newsreleases/16th-annual-fix-leak-week-reminds-businesses-reduce-water-waste)

**Specific Leak Sources:**
- Faucets/taps: One drip per second = 3,000+ gallons wasted annually
  - Source: [EPA WaterSense Statistics](https://www.epa.gov/watersense/statistics-and-facts)
- Toilets: Average leaking toilet wastes 200 gallons per day
  - Source: [Mr. Rooter Plumbing Statistics](https://www.mrrooter.com/about/blog/water-usage-statistics/)
- Irrigation systems: Poorly maintained systems waste up to 25,000 gallons/year per household
  - Source: [EPA WaterSense Statistics](https://www.epa.gov/watersense/statistics-and-facts)

---

## ENERGY CONSUMPTION

### Television Power Usage (65-75" TVs)

**Idle/Standby Mode:**
- 65-inch TVs: **1.1 watts average** (0.0011 kWh/hour)
  - Range: 0.5-3 watts typical for smart TVs
- 75-inch TVs: **2.6 watts average** (0.0026 kWh/hour)
  - Most common: 3 watts

**Active Viewing (Per Hour):**
- 65-inch LED: **95 watts** (0.095 kWh/hour)
- 65-inch OLED: **120 watts** (0.12 kWh/hour)
  - OLED uses approximately 26% more power than LED
- 75-inch average: **114.5 watts** (0.1145 kWh/hour)
- 75-inch OLED (e.g., Samsung S95F): **185 watts** (0.185 kWh/hour)

**Technology Comparison:**
- LED TVs consume 30-40% less power than OLED models
- 94% of Energy Star certified TVs are LED
- 89% use direct-lit LED, 11% use edge-lit configurations
- Power consumption increases 15-25% for every 10-inch increase in screen size

**Sources:**
- [TV Wattage 2024 Analysis](https://ecocostsavings.com/tv-wattage/) (analysis of 107 Energy Star TVs, February 2024)
- [TV Electricity Usage Guide 2025](https://solartechonline.com/blog/tv-electricity-usage-guide/)
- [Jackery TV Power Guide](https://www.jackery.com/blogs/knowledge/how-many-watts-does-a-tv-use)
- [Others Electric 2024](https://otherselectric.com/2024/11/07/how-much-electricity-does-a-tv-use/)

---

## FUEL/TRANSPORTATION CONSUMPTION

### Commuting Fuel Usage

**Personal Cars - Annual Fuel Consumption:**
- **281.6 gallons per year average** (based on 27.1 MPG fleet average for Model Year 2023)
  - Range: 282-300 gallons depending on vehicle fuel economy
  - Calculated: 7,632 annual commute miles (15.9 miles one-way × 2 trips × 240 workdays) ÷ 27.1 MPG
  - 69.2% of Americans drive personal vehicles to work (2024)
  - Sources: [US Gas Price News](https://news.usgasprice.com/how-much-fuel-does-the-average-american-commute-use-each-year/), [EPA Automotive Trends](https://www.epa.gov/automotive-trends/highlights-automotive-trends-report)

**Breakdown by Vehicle Type:**
- Compact car (32 MPG): **238 gallons/year**
- Midsize SUV (24 MPG): **318 gallons/year**
- Full-size pickup (18 MPG): **424 gallons/year**
- Hybrid (50 MPG): **153 gallons/year**
- Average across types: **283.25 gallons/year**

**Additional Context:**
- National total: Over $90 billion spent annually on commuting fuel
- Additional waste: 15-20 gallons/year from idling
- Average cost: $2,043/year for car commuting (fuel, maintenance, insurance)

**Public Transport - Equivalent Fuel Consumption:**
(Passengers share the fuel consumption of transit vehicles; calculated for 7,632 annual commute miles)
- Transit rail: **54.0 gallons/year equivalent** (141.4 passenger-miles per gallon)
- Commuter rail: **100.4 gallons/year equivalent** (76.0 passenger-miles per gallon)
- Transit buses: **294.7 gallons/year equivalent** (25.9 passenger-miles per gallon)
  - Only 3.69% of Americans use public transit for commuting (2024)
  - Transit ridership at 85% of pre-pandemic levels
  - Sources: [Alternative Fuels Data Center](https://afdc.energy.gov/data/10311), [Bureau of Transportation Statistics](https://www.bts.gov/content/energy-intensity-passenger-modes), [Census Commuting Report 2022](https://www2.census.gov/library/publications/2024/demo/acsbr-018.pdf)

**Rideshare/Private Transport:**
- Uber holds 76% of US rideshare market, Lyft 24% (March 2024)
- 72% of Americans do not use ride-hailing apps; only 8% are frequent users
- Rideshare comprises 6% of national vehicle miles travelled
- Rideshare adds 2.6 vehicle miles for every mile of personal driving replaced (160% increase)
  - Source: [AutoInsurance Rideshare Statistics](https://www.autoinsurance.com/research/rideshare-statistics/)

**Environmental Impact:**
- Passenger cars emitted 370 Mt CO₂e in 2022
- Light trucks emitted 660 Mt CO₂e in 2022
- Together represent 57% of US transportation emissions
  - Source: [Center for Sustainable Systems](https://css.umich.edu/publications/factsheets/mobility/personal-transportation-factsheet)

---

## AI/TECHNOLOGY CONSUMPTION

### ChatGPT Usage Statistics

**User Numbers:**

*Global Weekly Active Users:*
- December 2024: **300 million**
- August 2025: **700 million**
- October 2025: **800 million**
  - Sources: [TechCrunch October 2025](https://techcrunch.com/2025/10/06/sam-altman-says-chatgpt-has-hit-800m-weekly-active-users/), [TechCrunch March 2025](https://techcrunch.com/2025/03/06/chatgpt-doubled-its-weekly-active-users-in-under-6-months-thanks-to-new-releases/)

*USA-Specific Users:*
- **70 million monthly active users** (2025)
  - USA represents 19.01% of global user base (largest country)
  - Sources: [Backlinko ChatGPT Statistics](https://backlinko.com/chatgpt-stats), [DemandSage Statistics](https://www.demandsage.com/chatgpt-statistics/)

*USA Weekly Active Users (Calculated):*
- December 2024: **~57 million** (300M × 0.1901)
- August 2025: **~133 million** (700M × 0.1901)
- October 2025: **~152 million** (800M × 0.1901)

*Adoption Rates (USA, 2024):*
- 23% of US adults have ever used ChatGPT (as of February 2024)
- 22% use ChatGPT at least once per month
- Only 7% report frequent daily use
- 43% of adults aged 18-29 have used ChatGPT
  - Source: [ElectroIQ ChatGPT Statistics](https://electroiq.com/stats/chatgpt-statistics/)

### Power Consumption Per Query

**Official OpenAI Figure:**
- **0.34 watt-hours (0.00034 kWh)** per query
  - Stated by Sam Altman
  - Source: [DCD Sam Altman Interview](https://www.datacenterdynamics.com/en/news/sam-altman-chatgpt-queries-consume-034-watt-hours-of-electricity-and-0000085-gallons-of-water/)

**Third-Party Estimates:**
- Electric Power Research Institute (EPRI): **0.0029 kWh** per query
- Epoch AI (2024): **~0.3 watt-hours** for GPT-4o queries
- Range: **0.0017 kWh to 0.0026 kWh** depending on model and query complexity
- Complex queries can consume over **20 Wh** for the smartest models
  - Sources: [Epoch AI Energy Analysis](https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use), [Business Energy UK](https://www.businessenergyuk.com/knowledge-hub/chatgpt-energy-consumption-visualized/), [Dev Sustainability](https://www.devsustainability.com/p/chatgpt-energy-usage-is-034-wh-per)

**Comparison:**
- Google Search: 0.0003 kWh per query
- **ChatGPT uses approximately 10× more energy than Google Search**
  - Source: [University of Washington Research](https://www.washington.edu/news/2023/07/27/how-much-energy-does-chatgpt-use/)

**Annual Energy Per User (Estimated):**
- Assuming 10 queries/day: **1.24 kWh/year**
- Assuming 30 queries/day: **3.72 kWh/year**

### Training Power Consumption

**GPT-3 Training (175 billion parameters):**
- **1.287 GWh** (1,287 MWh)
  - Some estimates suggest up to 10 GWh including infrastructure
  - Source: [Medium GPT-3 Energy](https://medium.com/@rogt.x1997/ais-dirty-secret-how-gpt-3-consumed-1-287-mwh-and-emitted-the-same-co%E2%82%82-as-112-cars-5e43b85eb600)

**GPT-4 Training (estimated 280 billion parameters):**
- **62.32 GWh** (62,319 MWh) - most cited estimate
  - Range: 51.77 to 62.32 GWh
  - **40-48× higher than GPT-3** despite only ~10× parameters
  - Sources: [Baeldung LLM Power Consumption](https://www.baeldung.com/cs/chatgpt-large-language-models-power-consumption), [Towards Data Science GPT-4 Carbon](https://towardsdatascience.com/the-carbon-footprint-of-gpt-4-d6c676eb21ae/)

**Training Energy Per User (Calculated):**

*Using 800 million global users (2025):*
- GPT-4: **77.9 Wh per user** (62.32 GWh ÷ 800M users)
- GPT-3: **1.61 Wh per user** (1.287 GWh ÷ 800M users)

*Using 300 million users (December 2024):*
- GPT-4: **207.7 Wh per user** (62.32 GWh ÷ 300M users)
- GPT-3: **4.29 Wh per user** (1.287 GWh ÷ 300M users)

**Note:** Training is a one-time cost amortised across all users, while inference (query) costs are ongoing.

### Total ChatGPT Energy Consumption

**Daily/Annual Energy:**
- December 2024: **2.9 million kWh/day** (with 1 billion queries/day)
- Annual (2024): **1,058.5 GWh/year**
- Cost: Approximately **$139.72 million/year** at $0.132/kWh (US commercial rate, October 2024)
  - Source: [Business Energy UK](https://www.businessenergyuk.com/knowledge-hub/chatgpt-energy-consumption-visualized/)

### Water Consumption (AI Data Centres)

**Per Query:**
- OpenAI (Altman): **0.000085 gallons** (~0.3 mL) per query
  - Source: [DCD Sam Altman Interview](https://www.datacenterdynamics.com/en/news/sam-altman-chatgpt-queries-consume-034-watt-hours-of-electricity-and-0000085-gallons-of-water/)
- UC Riverside estimate: **519 mL** per 100-word prompt (one water bottle)
  - Source: [Washington Post AI Water Usage](https://www.washingtonpost.com/technology/2024/09/18/energy-ai-use-electricity-water-data-centers/)
- General estimate: **~0.5 litres** per 20-50 queries

**Daily/Annual Water Usage:**
- Estimated **148.28 million litres/day** (39.16 million gallons/day)
  - Source: [Carbon Credits Environmental Cost](https://carboncredits.com/chatgpt-hits-700m-weekly-users-but-at-what-environmental-cost/)

**Data Centre Context:**
- Large data centres: up to 5 million gallons/day
- Google data centres: 8.1 billion gallons in 2024
- Microsoft water use increased 34% between 2021-2022
  - Source: [Medium ChatGPT Water Usage](https://medium.com/readers-club/chatgpt-feels-free-but-its-burning-through-water-you-ll-never-see-1a1167244a5a)

---

## ADDITIONAL INTERESTING DATA POINTS

### Fleet Fuel Efficiency Trends
- MY 2023 represents largest fuel economy improvement in 9 years
- Electric vehicles (BEV, PHEV, fuel cell) reached 11.5% of production in MY 2023
- Projected to reach 14.8% in MY 2024
- Fuel economy nearly doubled since 1975 (13.1 MPG to 27.1 MPG)
  - Source: [EPA Report Fuel Economy Record](https://www.epa.gov/newsreleases/epa-report-shows-us-fuel-economy-hits-record-high-and-co2-emissions-reach-record-low)

### Commuting Time and Economic Impact
- Average one-way commute: 27.6 minutes
- Round-trip commute: 50.8 minutes
- Annual time spent commuting: 221 hours
- Average household transportation costs: $3,119.74/year for petrol and fuel
- 93% of household transport spending goes to buying and operating cars
- Miami is most expensive city for car commuting ($2,656/year)
  - Source: [US Gas Price News](https://news.usgasprice.com/how-much-fuel-does-the-average-american-commute-use-each-year/)

### Golf Demographics (2024)
- 3.7 million juniors played golf on a course (highest since 2006)
- 8 million female golfers played on-course
- 6.8 million golfers aged 18-34 (dominant age group)
- 3.35 million golfers actively posting scores to GHIN handicap system
- 77 million scores posted in 2024 (6% increase vs 2023, 30% increase since 2020)
  - Sources: [PGA By the Numbers 2024](https://www.pga.com/story/by-the-numbers-golf-in-2024), [USGA Golf Scorecard 2024](https://www.usga.org/content/usga/home-page/articles/2024/12/golf-scorecard-2024-statistics-golf-recreational-game.html)

### Household Water Waste Financial Impact
- $6 billion/year cost from household leaks nationally
- Average family could save $380/year by fixing leaks
  - Source: [EPA WaterSense Statistics](https://www.epa.gov/watersense/statistics-and-facts)

---

## RESEARCH METHODOLOGY

**Data Collection:**
- Five parallel sub-agents conducted focused research using web search
- All mathematical calculations verified using calculator tools
- Multiple sources consulted for each data point where available
- Priority given to government sources (EPA, USGA, Census) and industry organisations

**Data Quality:**
- All data from 2024-2025 unless explicitly noted
- Where estimates vary, multiple sources and ranges provided
- Calculations shown for derived statistics
- USA-specific data prioritised; global data included where relevant for context

**Research Completed:** 28 November 2025

---

# July 2026 Update

Research conducted: 3 July 2026 via six parallel research agents, with load-bearing figures re-verified against primary sources. This section supersedes the November 2025 notes where they conflict.

## Corrections to the original research

- **"US Infrastructure Leaks" (126.1 trillion litres / 33.3 trillion gallons) is a GLOBAL figure, not US.** ASCE's drinking water page quotes 126 billion m³ without a geographic label, but the same page states US public supply withdraws only ~39 billion gallons/day (~14.2T gallons/yr) - you cannot lose 33.3T gallons from a 14.2T gallon supply. The figure matches the widely-cited global non-revenue-water estimate. US-specific losses: ~6 billion gallons of treated water per day ≈ 2.19T gallons ≈ **8.3 trillion litres/year** (EPA/ASCE lineage). The dashboard now shows both, labelled.
  - Source: [ASCE Drinking Water](https://infrastructurereportcard.org/cat-item/drinking-water-infrastructure/) (checked 3 Jul 2026)
- **ChatGPT annual water value contained an arithmetic error.** The stated formula (0.32 mL x 2.5B queries/day x 365) gives **~292 million litres/year**, not the 111.2 billion previously shown (~380x overstatement). Corrected; labelled "on-site cooling only".
- **Sora video energy is per 10 seconds, not 5.** The ~1 kWh estimate traces to a Nov 2025 SemiAnalysis-derived calculation for a 10-second Sora 2 clip. Ratio vs text query: ~2,900x (not 4,000x). Measured open video models span 0.14-109 Wh/clip ([Hugging Face, Jul 2025](https://huggingface.co/blog/jdelavande/text-to-video-energy-cost)).
- **"100 queries/day = 0.3% of household electricity" recomputes to ~0.12%** (34 Wh ÷ ~28.8 kWh/day household average).
- **Netflix note:** the "IEA says ~36" figure sometimes cited against the 77 Wh/hour estimate is 36 g CO₂/hour (emissions), not 36 Wh - not a competing energy figure. 77 Wh/hour stands; ~72% of it is the viewing device, not the data centre.

## Key new data (mid-2026)

### Usage and adoption
- ChatGPT: **900M+ weekly active users** (Feb 2026, [TechCrunch](https://techcrunch.com/2026/02/27/chatgpt-reaches-900m-weekly-active-users/)); ~1B **monthly** actives (Jun 2026, Reuters/Sensor Tower - different metric). Latest officially confirmed volume remains **2.5B+ prompts/day** (Jul 2025); the "2.8B/day Q1 2026" figure circulating could not be verified on OpenAI's own Q1 2026 post (checked directly, 3 Jul 2026).
- **Pew Research (survey Feb 2026, n=5,119, published 17 Jun 2026):** 49% of US adults have used AI chatbots (33% in 2024, 23% in 2023); 24% use daily. ChatGPT 44%, Gemini 24%, Copilot 17%, Meta AI 14%, Grok 8%, Claude 6%. [Source](https://www.pewresearch.org/internet/2026/06/17/americans-and-ai-2026-chatbots-smart-devices-and-views-on-impact/)
- Google Gemini: 750M MAU (Q4 2025 earnings), reportedly 900M+ by May 2026.

### Per-query figures (unchanged, still the latest disclosures)
- Google Gemini: 0.24 Wh / 0.26 mL / 0.03 g CO₂e per median text prompt (Aug 2025, [Google Cloud](https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference)) - no newer disclosure through mid-2026. The accelerator-only boundary would be 0.10 Wh / 0.12 mL; the headline figure is full-stack (idle + overhead included). Note: the widely-quoted "9 Wh in May 2024" before-value doesn't reconcile with Google's own 33x claim (0.24 x 33 ≈ 8 Wh); the dashboard now shows ~8 Wh, derived.
- OpenAI: 0.34 Wh / 0.32 mL (Altman, Jun 2025; reaffirmed Feb 2026 - he called per-query water concerns "fake" and said OpenAI has largely moved off evaporative cooling, [CNBC](https://www.cnbc.com/2026/02/23/openai-altman-defends-ai-resource-usage-water-concerns-fake-humans-use-energy-summit.html)). No methodology published.
- Mistral Large 2 LCA (Jul 2025): 1.14 g CO₂e / **45 mL** water per 400-token reply - includes Scope 2 electricity water, hence ~170x Google's number. Boundary, not thirst.
- Anthropic, Meta, xAI, DeepSeek: no per-query disclosures as of mid-2026.

### Reasoning models (new headline insight; measured)
- Reasoning modes average **25-30x more energy** than standard queries; same-model reasoning on/off measured at **150-700x**; driven by ~10x more output tokens. Two independent sources: [HF AI Energy Score v2](https://huggingface.co/blog/sasha/ai-energy-score-v2) (Dec 2025) and [ML.Energy v3](https://ml.energy/blog/measurement/energy/diagnosing-inference-energy-consumption-with-the-mlenergy-leaderboard-v30/).
- Newer ≠ greener: 8 of 14 comparable models used equal-or-more energy than early-2025 predecessors (HF v2). MoE: measured 3.6x lower energy/token vs similar-size dense (ML.Energy).
- A frontier per-model energy comparison (GPT-5.x vs Claude vs Gemini) is NOT honestly sourceable - no closed-lab disclosures; third-party estimates (Jegham et al. arXiv 2505.09598; URI ~18 Wh GPT-5 claim) are modelled, not measured, and mix methodologies.

### Water accounting and boundaries
- Morgan Stanley (Sep 2025): AI data centre water ~**1,068 billion L/yr by 2028** (range 637-1,485B; ~11x 2024), counting Scope 1 cooling + Scope 2 power-generation + Scope 3 chip-fab water.
- All global AI today (Scope 1+2): ~312-765 billion L/yr (de Vries; Ren via [IEEE Spectrum](https://spectrum.ieee.org/ai-water-usage)) - wide uncertainty. The "~260B gallons global AI" figure traces to Mordor Intelligence market research via aggregators: low quality, avoid.
- The "training ≈ 50% of a model's resource use" claim (Hank Green video) is a misattribution of UC Riverside's work; current consensus is inference dominates lifetime footprint (60-90%).
- USGS: thermoelectric = 41% of total US water withdrawals (34% of freshwater), but only ~3-4% consumed.
- Corn (verified for the comparison anchor): US corn **irrigation** ~5.3-5.5T gallons ≈ 20.4T litres/yr (USDA NASS 2023); total evapotranspiration ~45-60T gallons. Green's "20T gallons" matches neither (likely conflation with all-crop irrigation). ~35-40% of corn to ethanol (USDA); 1,174-1,492 gal water/gal ethanol (Water Resources Research).
- ~2/3 of new US data centres since 2022 in high water-stress areas (Bloomberg/S&P Global).
- Microsoft FY2024: 5.81B L consumed / 10.71B withdrawn. Google 2024: 8.1B gal (30.7B L) consumed; ~9.9B gal derived for 2025.

### Data centre macro (IEA/LBNL/PJM)
- **2025 actual: 485 TWh global data centre electricity, +17% YoY; AI-focused facilities +50%** ([IEA Electricity 2026](https://www.iea.org/reports/electricity-2026), Dec 2025). ~950 TWh by 2030 (~3% of global electricity); ~1,200 TWh by 2035; higher scenarios judged less likely due to bottlenecks.
- US: 4.4% of electricity is 2023 data (LBNL, 176 TWh); 6.7-12% projected by 2028. 105 Mt CO₂ / 2.18% of emissions (2023). Carbon intensity 548 vs 369 g CO₂e/kWh (48% premium) independently reproduced (~545 g) by a June 2026 Harvard/UCLA hyperscale preprint (arXiv 2606.05420).
- **Electricity prices (new 2026 story):** PJM capacity auction record $333.44/MW-day (Dec 2025); market monitor attributes 63% of the 2025/26 increase ($9.3B) to data centres; PJM-state residential bills +1.5-5% from June 2026; NRDC scenario: up to $163B through 2033.
- Overbuild risk mainstream: Microsoft froze ~1.5 GW self-build; Nadella: "there will be an overbuild"; BIS Annual Report 2026 lists AI capex bust as a top systemic risk (~$1T/yr capex vs ~$50-60B AI revenue).
- Proportion: data centres ≈ 8% of projected global demand GROWTH to 2030 - less than EVs or air conditioning (Carbon Brief).

### Training
- Grok 4: ~310 GWh, 754M L water, 154 kt CO₂e - largest credibly-estimated run, ~6x GPT-4 ([Epoch AI](https://epoch.ai/data-insights/grok-4-training-resources), significant uncertainty flagged).
- GPT-5: ~5x10²⁵ FLOP median - LESS training compute than GPT-4.5 ([Epoch AI](https://epoch.ai/gradient-updates/why-gpt5-used-less-training-compute-than-gpt45-but-gpt6-probably-wont)). Frontier training is not monotonically growing.
- Training power demand grows ~2.2x/yr; hardware efficiency +~40%/yr; GW announcements (Stargate, Colossus 2) are capacity targets, not consumption.

### Everyday anchors (re-verified)
- Golf: 2.01T litres (531.1B gal, 2024 survey released Dec 2025, -31% vs 2005); ~16,000 courses; 29.1M on-course golfers (GCSAA/USGA/NGF).
- TV/streaming 6h45m/day confirmed (eMarketer 2025, reported as the peak year); 77 Wh/hr streaming confirmed; commuting 1,066 L/yr confirmed (EPA MY2024: 27.2 mpg); household leaks 3.41T L confirmed (EPA, with a noted internal EPA inconsistency: "nearly 1 trillion gallons" on its Fix-a-Leak page vs 900B on Statistics and Facts).
- US residential outdoor watering: EPA's current figure is ~8B gallons/day (~11T L/yr), superseding the older 9B figure.

### Anthropic / Claude (researched 3 July 2026)

- No official environmental disclosures as of Jul 2026: no per-query or fleet energy/water/carbon figures, no PUE/WUE, no sustainability report; the Transparency Hub is safety-only. Anthropic declined environmental disclosure in [Stanford's FMTI](https://crfm.stanford.edu/fmti/December-2025/company-reports/Anthropic_FinalReport_FMTI2025.html) (Dec 2025) citing proprietary concerns. First climate move: joined the Frontier carbon-removal coalition (Jun 2026); 2026 hires in non-financial/emissions reporting suggest formal disclosure is coming ([Heatmap](https://heatmap.news/climate/anthropic-carbon-emissions)).
- Official but qualitative/forward-looking: pays 100% of its grid-interconnection upgrade costs and cites "water-efficient cooling" ([Feb 2026](https://www.anthropic.com/news/covering-electricity-price-increases)); up to 5 GW Amazon compute; 2-5 GW single training runs projected for 2027-28; $50B Fluidstack data centre buildout (TX/NY, Nov 2025).
- Project Rainier (AWS New Carlisle, Indiana; Amazon's disclosures, not Anthropic's): ~2.2 GW planned, 1M+ Trainium2 chips, air-cooled ~98% of the year, estimated PUE ~1.15, Phase 1 operational cooling water permitted ~1.6M gal/day; the 31-35 MGD figures in local hearings are construction dewatering. AWS declines to quantify finished-campus operational water ([MeasuredAI](https://measuredai.substack.com/p/aws-new-carlisle-data-center-campus), from permits and utility filings).
- Only third-party per-query estimate: Jegham et al. ([arXiv 2505.09598](https://arxiv.org/abs/2505.09598)) - Claude 3.7 Sonnet ~0.84 Wh (short) / ~2.78 (medium) / ~5.52 Wh (long) per query; extended thinking ~3.5-17 Wh. Fully modelled (API latency + statistically inferred hardware, no telemetry) and covers only the superseded Claude 3.7, so kept off the page. No measured closed-model figures exist anywhere - HF AI Energy Score and ML.Energy can only test open models.
- Training compute (Epoch AI, benchmark-imputed, self-labelled low-precision/speculative): Claude 3.5 Sonnet ~3.6e25 FLOP; Claude Opus 4 ~1.5e26 FLOP. No third-party training energy or water estimate exists for any Claude model.
- Scale: $14B run-rate revenue ([official, Feb 2026](https://www.anthropic.com/news/anthropic-raises-30-billion-series-g-funding-380-billion-post-money-valuation)); reported ~$47B by May 2026 (secondary). ~40% of enterprise LLM API spend vs OpenAI 27% (Menlo Ventures, year-end 2025); ~54% of the coding market; Claude Code 2M+ weekly actives (Feb 2026); 6% US adult consumer reach (Pew). Consumer MAU estimates (12M-56M) are aggregator-grade, unusable. No disclosed query/token volume (contrast OpenAI's 2.5B+ prompts/day), so Claude's total footprint cannot be independently estimated.

### Training energy per model (researched 3 July 2026)

Three parallel research passes: developer disclosures, third-party estimates, and user-base denominators for the per-user amortisation chart.

**Developer-disclosed (energy derived from GPU-hours x TDP where marked; GPU-only unless stated):**

- Llama 3.1 405B: 30.84M H100-hours x 700 W = ~21.6 GWh GPU-only; 8,930 tCO₂e location-based, offset to 0 market-based ([Meta model card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_1/MODEL_CARD.md), Jul 2024). Full 3.1 collection: 39.3M GPU-h, ~27.5 GWh, 11,390 tCO₂e. Llama 4 Scout ~3.5 GWh (5.0M GPU-h), Maverick ~1.67 GWh (2.38M GPU-h); Behemoth undisclosed.
- DeepSeek-V3: 2.788M H800-hours x 700 W = ~1.95 GWh GPU-only; no emissions disclosed ([technical report](https://arxiv.org/abs/2412.19437), Dec 2024). R1's RL stage is not itemised.
- BLOOM: 433 MWh measured (dynamic GPU power), 24.7 tCO₂e dynamic / 50.5 full lifecycle - the canonical fully-documented run, low carbon due to the French nuclear grid ([Luccioni et al.](https://arxiv.org/abs/2211.02001), 2022).
- AI2 OLMo: measured data-centre energy including PUE (the strongest boundary of any disclosure): OLMo 2 7B 131 MWh / 13B 257 MWh ([arXiv 2503.05804](https://arxiv.org/abs/2503.05804)); OLMo 3 32B pretrain 1.42 GWh, whole series 1.95 GWh, 647 tCO₂e final runs (Nov 2025).
- Falcon-180B: ~7M A100-hours, ~2.8 GWh derived (HF blog, 2023). Gemma discloses tCO₂e (~131 t for Gemma 1) but no energy. Mistral's LCA (Jul 2025) covers Large 2 training + 18 months of inference combined (20.4 ktCO₂e, 281,000 m³ water) - not separable into a training-only figure.
- Confirmed absent: no official training energy/emissions from OpenAI, Anthropic, Google (Gemini), xAI, Alibaba or Nvidia.

**Third-party estimates (Epoch AI is effectively the sole source; energy = power draw x duration):**

- Epoch's [models database](https://epoch.ai/data/notable-ai-models) power-draw field is peak facility power (GPU TDP x count x 1.82 server overhead x PUE); the Grok 4 headline additionally applies x0.75 average-to-peak. Derived figures: Grok 3 ~237 GWh peak / ~178 GWh average; Gemini 1.0 Ultra ~92 GWh peak / ~69 GWh average (the only Gemini estimate in existence); GPT-4 ~45 GWh peak (corroborates the page's ~50 GWh); Llama 3.1 405B ~48 GWh facility-level (vs 21.6 GWh GPU-only - illustrates the boundary gap).
- Sanity check: GPT-4 ~50 GWh and Grok 4 ~310 GWh both stand unrevised as of Jul 2026.
- No credible training-energy estimate exists for GPT-4.5, GPT-5, o-series, Gemini 2/3, or any Claude model - Epoch has FLOP estimates only, because hardware type/count are not public.

**User-base denominators (for amortisation):**

- Firm and official: ChatGPT 900M WAU ([Feb 2026](https://techcrunch.com/2026/02/27/chatgpt-reaches-900m-weekly-active-users/)); Gemini app 900M MAU ([Google I/O, May 2026](https://blog.google/innovation-and-ai/sundar-pichai-io-2026/)) - different metrics (WAU vs MAU), not directly comparable.
- Soft or third-party: Meta AI 1B MAU (official but counts passive surfacing inside FB/IG/WhatsApp); DeepSeek ~125-130M MAU and Grok ~117M MAU (aggregator-grade); Claude has no official consumer figure. Llama's 1.2B is cumulative downloads, not users - wrong denominator for a per-user axis.
- Nobody has published a training-energy-per-user curve; the nearest reference is Epoch's [inference-vs-training crossover](https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use) (cumulative inference passes training energy after ~150-200 days at 1B queries/day), which is consistent with the page's "inference dominates" framing.

**Chart decision:** the Training Energy Per User chart now plots six curves, each labelled with its year - Grok 4 (2025) ~310 GWh, Gemini 1.0 Ultra (2023) ~70 GWh (average basis, consistent with Grok 4's), GPT-4 (2023) ~50 GWh, Llama 3.1 405B (2024) 21.6 GWh, DeepSeek-V3 (2024) ~1.95 GWh, Llama 4 Maverick (2025) ~1.67 GWh - over the same hypothetical 100M-900M user axis. The vintage skew is unavoidable and the chart note says so: after 2024 the frontier stopped being estimable (hardware type/count not public), so Grok 4 is the newest frontier-scale figure anywhere; Llama 4 Maverick is the newest disclosed run and shows energy falling, not rising, below the frontier. The note also flags the facility-level vs GPU-only boundary mix. GPT-3 (1.287 GWh) stays in the data but off the chart. Real user-base markers were considered and rejected: only two denominators are official and they use different metrics. BLOOM/OLMo are cited in the section's highlight box as the only measured runs.

### Token efficiency of the current generation (researched 3 July 2026)

Three parallel research passes: leaderboard measurements, lab-official claims, and recent analyses (Jan-Jul 2026). Purpose: give "Reasoning Models Changed the Maths" current-generation token data.

**Headline (verified against the primary source):** Epoch AI's ["How persistent is the inference cost burden?"](https://epoch.ai/gradient-updates/how-persistent-is-the-inference-cost-burden) - reaching ~27% accuracy on FrontierMath took ~43M output tokens with o4-mini (high effort, Apr 2025) but ~5M with GPT-5.2 (low effort, Dec 2025): ~9x fewer tokens, "roughly a 3x cost reduction over eight months" after per-token prices, and "very roughly a 5-10x cost reduction per year for reaching a given capability level". Note: a research agent initially reported this as "3x fewer tokens"; the fetched page shows 3x is the _cost_ figure, 43M to 5M (~9x) is the token figure. Epoch also cites SemiAnalysis: Anthropic substantially cut Claude's reasoning verbosity between Sonnet 3.7 and Sonnet 4, and Claude had the lowest total output tokens among leading reasoning models at that time.

**Cross-model measurement used for the new chart** - Artificial Analysis "output tokens per Intelligence Index task" (60-eval suite, answer + reasoning tokens, measured via each model's API; [artificialanalysis.ai/models](https://artificialanalysis.ai/models), fetched 3 Jul 2026): DeepSeek V4 Pro (max) 36,963 tokens @ index 44.3; Kimi K2.6 35,022 @ 42.8; Claude Fable 5 33,127 @ 59.9 (highest index measured); Gemini 3.5 Flash 27,733 @ 50.2; Qwen3.7 Max 22,373 @ 46.0; GPT-5.5 (xhigh) 16,064 @ 54.8; Grok 4.3 (high) 14,365 @ 37.6; Gemini 3.1 Pro (preview) 13,171 @ 46.5. A ~2.8x token spread on the same suite; reasoning tokens are 51-80% of output; GPT-5.5 and Gemini 3.1 Pro deliver ~2x the intelligence-per-token of DeepSeek V4 Pro/Kimi K2.6. Related AA data: cost to run the full index ranges from $3,752 (Opus 4.8) and $2,819 (GPT-5.5 xhigh) down to $176 (DeepSeek V4 Pro). Claude Opus 4.8 and Sonnet 5 sit outside AA's default token chart, hence their absence from ours. Re-verified directly against AA's embedded schema.org dataset (answer + reasoning sums reproduce every figure, e.g. Gemini 3.1 Pro Preview 2,741 + 10,430 = 13,171). Configuration matters when reading these: each model is measured at the setting AA chose - Claude Fable 5's config is "Adaptive Reasoning, Max Effort, Opus 4.8 Fallback", its most deliberate mode - and single-task benchmark token counts do not capture multi-turn agentic behaviour, where verbosity and retry loops can rank models very differently from a one-shot suite. Both caveats are stated under the chart.

**Why tokens and not energy:** ML.Energy and HF AI Energy Score meter Joules on their own GPUs, so they can only ever test open-weight models - no energy leaderboard can compare today's closed flagships. API-metered token counts (Artificial Analysis, Aider) are the only cross-vendor efficiency axis that includes them. No HF Energy Score v3 exists (v2, Dec 2025, remains latest); ML.Energy v3 likewise unrevised as fetched.

**Lab claims (vendor-grade, flagged as such on the page):** Anthropic Opus 4.5 ([Nov 2025](https://www.anthropic.com/news/claude-opus-4-5)) is the only quantified first-party figure: medium effort matched Sonnet 4.5's SWE-bench Verified score with 76% fewer output tokens; highest effort beat it by 4.3pp with 48% fewer. Opus 4.8 (May 2026) defaults to high effort with adaptive thinking ("fewer wasted thinking tokens"); Claude Fable 5 (Jun 2026) makes adaptive thinking the only mode. GPT-5.5 claims "fewer reasoning tokens than prior models at the same reasoning effort" (no figure); Gemini 3.1 Pro adds a medium thinking_level and reportedly "uses reasoning more efficiently" (secondary source); DeepSeek V4's Compressed Sparse Attention and Qwen3.5's 8.6-19x decoding throughput are compute/throughput claims, not tokens-per-task. Every major lab now exposes an effort/thinking dial; OpenAI adds a verbosity parameter.

**Independent context:** Nous Research ([Aug 2025](https://nousresearch.com/measuring-thinking-efficiency-in-reasoning-models-the-missing-benchmark)) - open-weight models use 1.5-4x more reasoning tokens than closed for equivalent tasks (up to ~10x on simple knowledge questions); consistent with AA's 2026 snapshot where DeepSeek/Kimi top the token chart. OckBench ([arXiv 2511.05722](https://arxiv.org/abs/2511.05722), Nov 2025) - up to 5x token variance at equal accuracy; "token efficiency remains largely unoptimized". Aider polyglot (not updated past Oct 2025, previous-gen only): GPT-5 effort sweep on identical tasks - low 4,260 tokens/solved @ 81.3%, medium 7,532 @ 86.7%, high 13,250 @ 88.0% (3.1x tokens for +6.7pp), the cleanest illustration that effort dials trade tokens for points. Stanford AI Index 2026/IEEE Spectrum: inference to GPT-3.5-level capability fell $20 to $0.07 per M tokens (Nov 2022-Oct 2024); energy efficiency improving ~40%/yr; per-token energy ~0.0001-0.002 Wh.

**What went on the page:** a "The maths is moving again" highlight box (Epoch fixed-capability collapse ~9x/8mo, effort dials, Opus 4.5 76% flagged as vendor claim, and the catch: frontier per-query spend keeps climbing because savings are spent on harder problems). The existing 25-30x / 150-700x measured multipliers stay - they are snapshots of reasoning-on-vs-off, which the new data complements rather than contradicts. An AA output-tokens-per-task bar chart (eight current models) was published, then reworked for clarity, then removed the same day by editorial decision: a single-suite, single-configuration token count is too easily read as a model-quality ranking, and it contradicted real-world multi-turn experience (a model can be terse on one-shot benchmarks yet wasteful in agentic use). The per-model AA figures remain recorded above for reference; only the Epoch fixed-capability trend and the effort-dial shift are page-worthy.

### Post-publication refinements (3 July 2026)

- Both bar charts (water and energy) now draw bar lengths on a linear scale, true to the numbers. The earlier log-scale widths preserved rank but destroyed proportion (Google's 30.7B litre bar rendered at ~45% of the 126T bar when it is 1/4,000th of it). Bars that would be thinner than a pixel are clamped to a minimum visible width and labelled as such; the energy chart's Sora-only log special case was removed.
- The ~126 trillion litre global network losses figure is attributed to Liemberger & Wyatt 2019, "Quantifying the global non-revenue water problem", Water Supply ([DOI 10.2166/ws.2018.129](https://doi.org/10.2166/ws.2018.129)): a modelled estimate of ~346 million m³/day extrapolated from utility-reported non-revenue water rates, not a per-country measurement. The page now labels it "Modelled estimate, not a census" and displays "~126 trillion".
- research.md and data.json are now linked from the page's header navigation, not only the footer.
- The "often misquoted as US-only" phrasing was dropped from the global losses sublabel; the misattribution history stays in the data.json notes field.
- Added Anthropic/Claude coverage (see the Anthropic / Claude section above): a note in the disclosure-gap box, Project Rainier as an operator-response example in the boundaries section, and a perspectives card on Anthropic's undisclosed API-shaped footprint.
- Water locality gained a second independent source: a [Guardian analysis](https://www.theguardian.com/us-news/2026/jun/08/datacenter-ai-drought-water) (8 Jun 2026, Cleanview + US Drought Monitor data) found 517 of 809 planned US data centres (~2/3) on land in drought throughout the preceding year, corroborating the Bloomberg/S&P water-stress finding for existing builds. It also documents why developers pick dry land (cheap land, tax breaks, less equipment corrosion) and scale (a large site can draw up to 5M gallons/day, the water use of ~50,000 people).
- The Training Energy Per User chart went from one curve (GPT-4) to six (see the Training energy per model section above): Grok 4, Gemini 1.0 Ultra, GPT-4, Llama 3.1 405B, DeepSeek-V3 and Llama 4 Maverick, spanning ~185x, each legend entry labelled with the model's year. The note under the chart explains the vintage skew (post-2024 frontier models are not estimable - a disclosure gap, not a data-selection choice) and the facility-level vs GPU-only boundary mix; the measured-runs contrast (BLOOM, OLMo) went into the section's highlight box.
- The reasoning section gained current-generation token-efficiency framing (see the Token efficiency section above): a "The maths is moving again" highlight box on the fixed-capability cost collapse (Epoch AI, ~9x fewer tokens for the same FrontierMath accuracy in eight months) and the industry-wide shift to effort/thinking dials. The 25-30x energy multiplier framing is unchanged. An accompanying bar chart of Artificial Analysis output-tokens-per-task across eight frontier models was published, reworked for direction/colour clarity, and then removed the same day: single-suite token counts at one configuration invite a model-quality reading the data cannot support (details in the Token efficiency section).
- Section order reworked so data leads and interpretation follows: Water, golf big-number and Boundaries now open the page, then Daily Energy (with the 0.24 Wh big-number), Task Types, Reasoning, and only then AI Efficiency is Improving (previously the first section); the header nav matches. Rationale: the page's core argument is comparative consumption data, so the comparisons come before the efficiency-trend commentary.
- The footer's data-provenance line (sources + last-updated date) is now also shown at the top of the hero, pinned below the nav so it's visible on load and scrolls away with the page; the Research and data.json nav links open in new tabs.
- Two hero stat cards added: ~11,700x more water lost to US household leaks than ChatGPT's global cooling water (EPA 3.41T L / 292M L, same on-site boundary as the water chart) and ~35x more energy for an hour of big-screen TV than a 10-prompt ChatGPT conversation (120 Wh / 3.4 Wh). Hero numbers are now tone-coloured - green where AI consumption is smaller than the everyday comparison or improving, orange-red for genuine costs.
- Hero card set finalised at six, ordered water-energy-cautions-trajectory: 3-6x golf, ~11,700x leaks, ~35x TV vs 10 prompts, ~2,900x video vs text (caution), 2/3 of new US data centres in water-stressed areas (caution), 33x Gemini efficiency. Dropped as redundant or off-message: 227 queries per Netflix hour and ~550x daily TV (both restated the ~35x point), the 25-30x reasoning card (covered in its own section), and 49% chatbot adoption (not a consumption stat). The 4-green/2-orange mix mirrors the page's stance: smaller than assumed, with real outliers named.
- The ~1,000 Wh Sora 2 bar was removed from the Daily Energy chart - at true scale it flattened every other bar to slivers (an hour of OLED TV, the next-largest, is 120 Wh). Video generation now lives in the chart footnote with the ~2,900x figure and a pointer to the Energy Varies by AI Task Type section, which keeps the full Sora comparison and the measured 0.14-109 Wh open-model range.

## Framing credit

The measurement-boundary framing (why honest numbers disagree by 1,000x; withdrawal vs consumption; water locality) was prompted by Hank Green's video ["Why is Everyone So Wrong About AI Water Use?"](https://youtu.be/H_c6MWk7PQc) (June 2026). His specific figures were independently verified: the boundary argument and locality thesis hold; his corn total and the "training is ~50%" claim did not survive verification and were replaced with primary-sourced figures.

**Update completed:** 3 July 2026

## September 2026 Update

Trigger: Knowable Magazine, ["How much of a problem is AI's water use?"](https://knowablemagazine.org/content/article/technology/2026/how-much-water-do-ai-data-centers-use) (Katarina Zimmer, 24 Aug 2026), reviewed against this page with three of its cited sources read in full. No corrections were needed; the article and this page agree on the 0.26 mL Gemini figure, the origin and obsolescence of the 500 mL claim (Ren's GPT-4 short-email estimate, power-generation water included), locality over aggregate volume, and the state of company disclosure. Four additions and one supporting figure followed.

- **Peer-reviewed US projection** ([You et al., Nature Sustainability 2025](https://www.nature.com/articles/s41893-025-01681-y), Cornell): US AI servers 731-1,125 million m³/yr (= billion L) by 2030 depending on buildout, Scope 1 on-site cooling + Scope 2 power-generation water, no chip fab. Upper bound roughly New York City's annual drinking water supply. Added as a third bookend (`boundaries.bookends.peerReviewed`) and to the Big Number card; corroborates Morgan Stanley's ~1,068B global 2028 figure from an academic source on a slightly narrower boundary. Same paper: best-practice WUE improvement cuts the water footprint ~29-32%; combined best practices cut it 86%; industry-average PUE 1.58 / WUE 1.8 are the reference points.
- **Siting swing** (same paper): best vs worst distribution of AI servers across US states moves the 2030 water footprint by -52% to +354%. Added to the Water is local box as the number behind "where matters more than how much".
- **US on-site baseline**: all US data centre cooling consumed 66 billion L in 2023, under 1% of national consumption (LBNL 2024 US Data Center Energy Usage Report, as cited by Knowable). Added as a water bar (`us_datacenters_onsite`) on the same on-site boundary as the ChatGPT bar. The LBNL escholarship page returned 403 during this pass, so the figure is taken from Knowable's citation, not re-read at source; the report's 2028 direct-water projection was not added for that reason.
- **Texas** ([UT Austin Bureau of Economic Geology white paper](https://compass.beg.utexas.edu/files/publications/Water_Requirements_for_DC_White_Paper.pdf), 2025): data centres ~0.75% of statewide water withdrawals in 2025 (43.5B gallons, direct + power-generation), projected 3-5% by 2030-2040 at 40-60 GW or 5-9% at 70-110 GW. Withdrawals, not consumption; the paper's own grid figures give ~0.96 gal/kWh consumed vs ~87 gal/kWh withdrawn for grid-connected data centres, which is the withdrawal-vs-consumption gap in one number. Knowable's "3 to 9 percent by 2040" paraphrase is the high-capacity band.
- **Operator responses** ([Microsoft, Dec 2024](https://www.microsoft.com/en-us/microsoft-cloud/blog/2024/12/09/sustainable-by-design-next-generation-datacenters-consume-zero-water-for-cooling/)): all new data centre designs since Aug 2024 use closed-loop chip-level cooling with zero evaporative water; pilots in Phoenix and Mt Pleasant, Wisconsin in 2026, sites online from late 2027; claimed avoidance >125M L/yr per data centre; fleet WUE 0.30 L/kWh in FY2024 vs 0.49 in 2021. Google targets replenishing 120% of freshwater consumed by 2030 (cited by Knowable; Google water stewardship portfolio page not re-read). Both added to the operator-response text.
- Not read: the Masanet et al. 2025 review "The water use of data center workloads" (ScienceDirect returned 400). Knowable's framing from it (siting, local energy mix, climate and design decide the number) matches the page's existing position.

**Update completed:** 2 September 2026