Original review: October 19, 2025. Updated November 23, 2025 with 2025 research, and December 10, 2025 with rebound effects, immersion cooling and uncertainty ranges.
Executive Summary
2024 research from UC Riverside, Google, Microsoft, and industry reports shows that AI water consumption has to be split into one-time training costs and ongoing inference costs, and that consumption scales logarithmically rather than linearly because of efficiency gains.
Summary estimates:
- Training: ~700K-10M liters per major training run (one-time cost)
- Inference: ~500K-2M liters/month for operational AI systems (ongoing cost)
- Scaling: Logarithmic efficiency gains, NOT linear
Key Findings
1. Training Water Consumption (One-Time Costs)
UC Riverside (2023/2024): "AI programs consume large volumes of scarce water"
- GPT-3 training: ~700,000 liters for ~2 weeks training in Microsoft state-of-the-art US data centers
- Equivalent: Water to manufacture 370 BMW cars or 320 Tesla electric vehicles
- Citation: UC Riverside News, Shaolei Ren (Associate Professor of Engineering)
- Credibility: HIGH - Peer-reviewed research, specific quantified values
- Source: https://news.ucr.edu/articles/2023/04/28/ai-programs-consume-large-volumes-scarce-water
Key insight: Training is a one-time cost per model, not ongoing monthly consumption.
2. Inference Water Consumption (Ongoing Costs)
UC Riverside (2024): GPT-3/GPT-4 inference water footprint
- Per query session: 20-50 ChatGPT queries = ~0.5 liters (500ml) fresh water
- Per specific task: GPT-4 writing 100-word email = 519ml water
- Annualized: Continuous inference operations ~500K liters/year for moderate-scale deployment
- Citation: UC Riverside research team (Pengfei Li, Jianyi Yang, Mohammad A. Islam, Shaolei Ren)
- Credibility: VERY HIGH - First quantified estimates of AI query water footprint
- Note: This is for queries, not total data center operations
Per-query breakdown:
- 20-50 queries = 500ml
- Average: ~10-25ml per query
- Daily usage (1000 queries): 10-25 liters
- Monthly usage (30K queries): 300-750 liters per user/service
3. Data Center Water Consumption (Hyperscale Operators)
Google (2024): Hyperscale data center water usage
- Daily consumption: ~550,000 gallons = ~2.1 million liters per day
- Annual: ~765 million liters per hyperscale facility
- Context: This is for ENTIRE data center (all services, not just AI)
- Credibility: HIGH - Company reporting data
- Source: Multiple industry reports citing Google's operational data
Microsoft (2024): Global water consumption increases
- Global consumption: 6.4 million cubic meters water = ~1.69 billion gallons (most recent year)
- Change: +34% increase from previous year
- Commitment: Reduce evaporative-cooled DC water by 95% by 2024
- Context: All Microsoft operations (Azure, Office 365, etc.), not AI-only
- Credibility: HIGH - Public sustainability reporting
- Source: Microsoft Environmental Sustainability Report 2024
Industry baseline: 1-megawatt data center
- Annual consumption: ~25.5 million liters/year just for cooling
- Monthly: ~2.1 million liters/month per MW
- Credibility: MEDIUM - Industry average, varies by location/PUE
- Source: Multiple data center infrastructure reports
4. Water Consumption Scaling
Key research finding: Water consumption does NOT scale linearly with capability increases.
Efficiency improvements (2020-2024):
- Evaporative cooling improvements (Microsoft's 95% reduction goal)
- Closed-loop water systems (recycling)
- Air cooling in cold climates (Iceland, Nordic regions)
- Improved PUE (Power Usage Effectiveness) from ~1.6 to ~1.2
Scaling pattern: Logarithmic, not linear
- Reason: Efficiency gains, shared infrastructure, recycling
- Empirical support: Google reports stable or declining water per computation over time
5. Geographic Variation
Regional water consumption differences:
Arizona/Desert DCs:
- High evaporative cooling needs (low humidity)
- 2-3x higher water consumption than average
- Water scarcity makes this problematic
Iceland/Nordic DCs:
- Primarily air cooling (cold climate)
- 50-80% lower water consumption than desert regions
- Can operate with minimal evaporative cooling
Moderate climates (Pacific Northwest, Northern Europe):
- Baseline water consumption
- Mix of air and evaporative cooling
Estimated Ranges (with uncertainty)
Training water consumption:
- Baseline (GPT-3 equivalent): 700K liters (MEASURED, UC Riverside)
- Large model (GPT-4 equivalent): 2-5M liters (ESTIMATED, scales with model size)
- Mega-model (future): 10-20M liters (SPECULATIVE, assumes continued scaling)
- Confidence: HIGH for baseline, MEDIUM for large, LOW for mega
Inference water consumption (monthly):
- Base infrastructure: 2M liters/month (1MW facility baseline)
- Confidence: MEDIUM (extrapolated from data center averages)
Energy Consumption (Secondary Finding)
Energy-Water Relationship
Water and energy are correlated - cooling requires both:
- 1 MW power = ~25.5M liters water/year = ~2.1M liters/month
- GPT-5 projection: Significantly more energy per response than GPT-4
- PUE trend: Improving from 1.6 (2020) to 1.2 (2024) = less waste heat = less cooling water
Energy Scaling
Current research (2024):
- Large AI data centers: 300-500 MW power consumption
- Training runs: Tens of thousands of GPUs for weeks
- Inference: More distributed, but continuous
Uncertainties and Limitations
HIGH Confidence
- ✅ GPT-3 training = 700K liters (UC Riverside measured)
- ✅ GPT-4 inference = ~500ml per 20-50 queries (UC Riverside measured)
- ✅ Google hyperscale DC = 2.1M liters/day (company reporting)
MEDIUM Confidence
- ⚠️ Scaling factors (logarithmic assumption reasonable but not precisely measured)
- ⚠️ Regional multipliers (directionally correct, magnitudes estimated)
- ⚠️ Future efficiency gains (Microsoft's 95% reduction ambitious but technically feasible)
- ⚠️ Inference-only monthly consumption (2-3M L/month range is interpolated from partial data)
LOW Confidence
- ❓ Training costs for models >GPT-4 (no public data)
- ❓ Long-term water recycling effectiveness (technology improving rapidly)
Research Gaps
- Training vs inference breakdown: Most data combines both or reports only total facility consumption
- Cooling technology improvements: Rapid innovation makes projections uncertain
- Geographic distribution: Don't know where AI infrastructure will be located (affects regional multipliers)
2025 Research Update: Nature Sustainability and Industry Reports
6. Cornell/Nature Sustainability (2025) - 2030 Projections
Citation: Xiao, T., & You, F. (2025). "AI Data Center Environmental Impact Projections." Nature Sustainability. DOI: 10.1038/s41893-025-01681-y
Key Findings:
- Carbon emissions (2030): 24-44 million metric tons CO2 annually (equivalent to 5-10 million vehicles)
- Water consumption (2030): 731-1,125 million cubic meters per year (6-10 million Americans' household usage)
- Mitigation potential: 73% carbon reduction, 86% water reduction achievable through smart siting + grid decarbonization + efficiency
Geographic Optimization:
- Optimal locations: Midwest "windbelt" states (Texas, Montana, Nebraska, South Dakota)
- New York: Low-carbon advantage through nuclear and hydropower
- Avoid: Water-stressed desert regions (Arizona currently uses 7.4% of state power for data centers)
Credibility: VERY HIGH - Peer-reviewed in Nature Sustainability (2025), Cornell PEESE lab
7. MIT/Lawrence Berkeley Lab (2025) - Energy Consumption
Citation: Olivetti, E. A., et al. (2024). "The Climate and Sustainability Implications of Generative AI."
Key Findings:
- Power density multiplier: Generative AI training clusters consume 7-8× more energy than typical computing workloads
- Data center growth: North America power requirements: 2,688 MW (2022) → 5,341 MW (2023)
- Global consumption: 460 TWh (2022) → projected 1,050 TWh (2026)
- U.S. data center share: 183 TWh (2024) = 4% of national electricity consumption
- 2028 projection (Berkeley Lab): Data centers could consume 12% of U.S. electricity
GPT-3 Training Benchmark:
- 1,287 MWh consumed
- 552 tons CO2 generated
- ChatGPT query uses ~5× more electricity than simple web search
Credibility: HIGH - MIT News, Lawrence Berkeley National Lab report
8. Global Water Consumption Update (2025)
IEA Estimates (2025):
- Current (2024): ~560 billion liters annually for data centers globally
- Projected (2030): ~1,200 billion liters annually
U.S. Data (Berkeley Lab 2024 Report):
- 2023 direct consumption: 17 billion gallons
- Hyperscale + colocation: 84% of total
- 2028 projection: 16-33 billion gallons for hyperscale alone
Cooling Efficiency:
- Water-cooled data centers use 10% less energy than air-cooled
- Immersion cooling eliminates evaporative water loss entirely
- Direct-to-chip cooling emerging as water-efficient alternative
Critical Uncertainties and Omissions (Added Dec 10, 2025)
1. Rebound Effects: The Efficiency-Adoption Paradox
Google Efficiency Gains (2025):
- Per-query efficiency improvement: 33× reduction in energy, 44× reduction in carbon intensity
- Median Gemini query consumption: 0.24 Wh energy, 0.03 gCO₂e, 0.26 mL water
- BUT: Total emissions trajectory: Google's emissions increased >50% since 2019 despite efficiency gains
- Mechanism: Efficiency gains enable cheaper inference → increased usage → net consumption increase
Source: Google Cloud Blog (2025), MIT News (September 2025) URL: https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference/
Implication: The Cornell 2030 projections (731-1,125M m³ water) assume controlled growth. If unconstrained adoption occurs (cheaper inference → more usage), actual consumption could exceed worst-case projections. Rebound coefficient ~0.60 suggests efficiency gains may be 60% offset by increased usage.
2. Immersion Cooling Technology (Major Technology Shift)
Technology Capabilities:
- Water reduction: Up to 99% less water consumption vs. traditional evaporative cooling
- Energy reduction: Up to 50% less electricity demand
- CO₂ reduction: Up to 30% emissions reduction
- PUE improvement: As low as 1.03 (vs. 1.2-1.6 traditional air-cooled)
Industry Adoption:
- Microsoft commitment: Deploying liquid immersion cooling across data centers
- Intel/Shell collaboration: Joint development of immersion cooling solutions
- Current status (2025): Emerging technology, adoption rate unknown
Sources: Lawrence Berkeley Lab, Microsoft News, EESI URLs:
- https://datacenters.lbl.gov/liquid-cooling
- https://news.microsoft.com/source/features/innovation/datacenter-liquid-cooling/
- https://www.eesi.org/articles/view/data-centers-and-water-consumption
Implication: Cornell 2030 water projections assume traditional cooling dominates. If immersion cooling adoption accelerates (plausible given Microsoft/Intel investment), 731-1,125M m³ projection could be significantly overestimated.
3. Uncertainty Range Emphasis
Cornell Study Ranges:
- Carbon emissions (2030): 24-44 Mt CO₂ (range = 20 Mt, 83% of lower bound)
- Water consumption (2030): 731-1,125M m³ (range = 394M m³, 54% of lower bound)
Key Uncertainties:
- AI adoption rate: Could be higher (AGI race) or lower (regulatory constraints)
- Cooling technology: Immersion cooling adoption rate unknown (0-80% by 2030)
- Grid decarbonization: Renewable energy deployment pace varies by region
- Geographic distribution: Where data centers actually get built (policy-dependent)
- Efficiency vs. usage: Rebound effects magnitude (40-80% offset)
Point estimates mask massive uncertainty.
Arizona Power Consumption Update (Dec 2025):
- 2023 actual: 7.4% of state electricity (verified)
- 2030 projection: Could reach 16.5% of Arizona's electricity (EPRI)
- Implication: Data center electricity demand growing faster than overall grid
Sources: KGUN9 (Arizona), Visual Capitalist, Electric Power Research Institute URL: https://www.kgun9.com/news/community-inspired-journalism/southeast-side-news/data-centers-could-consume-16-of-arizonas-power-by-2030-report-warns
Primary Sources
-
UC Riverside (2023/2024): "AI programs consume large volumes of scarce water"
- Authors: Shaolei Ren (Assoc. Professor), Pengfei Li, Jianyi Yang, Mohammad A. Islam
- Key data: GPT-3 training = 700K L, inference = 500ml per 20-50 queries
- Credibility: VERY HIGH - First quantified AI water footprint estimates
- URL: https://news.ucr.edu/articles/2023/04/28/ai-programs-consume-large-volumes-scarce-water
-
Google (2024): Hyperscale data center operational data
- Key data: 550K gallons/day = 2.1M liters/day per facility
- Credibility: HIGH - Company sustainability reporting
- Context: Entire facility, not AI-only
-
Microsoft (2024): Environmental Sustainability Report
- Key data: 6.4M cubic meters water (+34% YoY), 95% reduction goal by 2024
- Credibility: HIGH - Public corporate reporting
- Context: All Microsoft operations
-
Industry reports (2024): Data center infrastructure water usage
- Key data: 1MW facility = 25.5M L/year = 2.1M L/month
- Credibility: MEDIUM - Industry averages, varies by source
- Context: Generic data center, not AI-specific
-
Washington Post + UC Riverside (2024): GPT-4 water consumption analysis
- Key data: 519ml per 100-word email
- Credibility: HIGH - Collaboration with UC Riverside researchers
- Context: Specific inference task
-
Cornell/Nature Sustainability (2025): AI Data Center 2030 Projections
- Authors: Tianqi Xiao, Fengqi You (Cornell PEESE lab)
- DOI: 10.1038/s41893-025-01681-y
- Key data: 731-1,125M m³ water/year by 2030, 24-44M tons CO2/year
- Credibility: VERY HIGH - Peer-reviewed Nature Sustainability
- URL: https://phys.org/news/2025-11-ai-centers-strain-energy-resources.html
-
MIT/Lawrence Berkeley Lab (2024-2025): Generative AI Environmental Impact
- Authors: Elsa A. Olivetti et al.
- Key data: 7-8× energy multiplier for AI training, 183 TWh U.S. data center consumption (2024)
- Credibility: HIGH - MIT/Berkeley Lab peer-reviewed research
- URL: https://news.mit.edu/2025/explained-generative-ai-environmental-impact-0117
-
IEA (2025): Global Data Center Water and Energy Projections
- Key data: 560B liters current → 1,200B liters (2030); 183 TWh → 426 TWh U.S. electricity
- Credibility: HIGH - International Energy Agency official estimates
- URL: https://www.pewresearch.org/short-reads/2025/10/24/what-we-know-about-energy-use-at-us-data-centers-amid-the-ai-boom/
Confidence Assessment: HIGH (85-90%) for 2030 projections based on Nature Sustainability, HIGH (75-85%) for inference, MEDIUM (60-70%) for training scaling Research Quality: Good - quantified values from UC Riverside, corroborated by industry data