Source-linked AI summary
Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models
Pengfei Li, Jianyi Yang, Mohammad A. Islam, Shaolei Ren
TL;DR
AI’s water footprint has remained underreported despite rapidly increasing data-center demand and worsening freshwater scarcity. The paper presents a methodology covering operational and embodied water, evaluates GPT-3, and examines spatial-temporal runtime efficiency. It reports millions of liters for training and argues that water and carbon footprints require joint attention.
Problem
AI’s water footprint from cooling and electricity generation is less scrutinized and less transparently reported than its carbon footprint despite freshwater scarcity.
Method
The paper presents a principled methodology for estimating AI’s total water footprint, including operational and embodied water, and analyzes GPT-3 and WUE variation.
Results
GPT-3 training in Microsoft’s U.S. data centers can consume 5.4 million liters of water, including 700,000 liters of scope-1 on-site consumption.
Takeaways & Limitations
AI sustainability should address water footprint alongside carbon footprint, while training location and timing can reduce water use and transparency should improve.
Takeaways & Limitations
Scope-3 embodied water usage remains limited by insufficient data, especially for chip manufacturing.
Abstract
from arXiv · showhide
The growing carbon footprint of artificial intelligence (AI) has been undergoing public scrutiny. Nonetheless, the equally important water (withdrawal and consumption) footprint of AI has largely remained under the radar. For example, training the GPT-3 language model in Microsoft's state-of-the-art U.S. data centers can directly evaporate 700,000 liters of clean freshwater, but such information has been kept a secret. More critically, the global AI demand is projected to account for 4.2-6.6 billion cubic meters of water withdrawal in 2027, which is more than the total annual water withdrawal of 4-6 Denmark or half of the United Kingdom. This is concerning, as freshwater scarcity has become one of the most pressing challenges. To respond to the global water challenges, AI can, and also must, take social responsibility and lead by example by addressing its own water footprint. In this paper, we provide a principled methodology to estimate the water footprint of AI, and also discuss the unique spatial-temporal diversities of AI's runtime water efficiency. Finally, we highlight the necessity of holistically addressing water footprint along with carbon footprint to enable truly sustainable AI.
1 Introduction
AI’s expanding data-center workloads are driving a substantial but underreported water footprint amid growing freshwater stress. The paper proposes measuring this footprint holistically and improving efficiency, transparency, and sustainability practices.
- AI’s water footprint includes freshwater consumed for server cooling and electricity generation, but has received less scrutiny than its carbon footprint.
- 150–280 billion liters of annual on-site water consumption by U.S. data centers could occur in 2028, reaching two to four times the 2023 level.
- Water and carbon footprints are complementary, because improving carbon efficiency does not necessarily improve water efficiency and may worsen it.
- 5.4 million liters of water could be consumed when training GPT-3 in Microsoft’s U.S. data centers, including 700,000 liters of scope-1 on-site consumption.
- WUE varies spatially and temporally, so choosing when and where to train large AI models can significantly reduce water footprint.
2 Background
AI’s water footprint spans withdrawal and consumption across cooling, electricity generation, and server manufacturing. These components differ in environmental implications and efficiency trade-offs, while important scope-2 and scope-3 data remain limited.
- 2.1 Water Withdrawal vs. Water Consumption: Water withdrawal is freshwater taken from ground or surface sources, whereas water consumption is withdrawal minus discharge and reflects downstream availability.
- 2.2 AI Water Usage: AI’s water usage covers scope 1 for cooling, scope 2 for electricity generation, and scope 3 for server manufacturing.
- 2.2.1 Scope-1 Water Usage: Cooling towers withdraw added water for evaporation and discharge, with roughly 80% evaporated under good water quality.
- 2.2.1 Scope-1 Water Usage: 1–9 liters per kWh of server energy may be evaporated in data-center cooling, depending on weather and operational settings.
- 2.2.1 Scope-1 Water Usage: Dry coolers consume no on-site water year-round but typically increase cooling energy and potentially scope-2 water consumption.
- 2.2.2–2.2.3 Scope-2 and Scope-3 Water Usage: AI’s scope-2 water usage reflects electricity-generation water costs, while chip and server manufacturing add a substantial but poorly documented scope-3 footprint.
3 Estimating AI’s Water Footprint
The paper estimates AI’s operational and embodied water footprint by combining server energy, data-center efficiency, grid water intensity, and server-manufacturing water across time and locations. A GPT-3 case study applies this framework using Microsoft’s reported efficiencies and specified workload assumptions.
- Operational water footprint: Operational water includes on-site scope-1 cooling water and off-site scope-2 water associated with electricity generation.
- On-site water efficiency: On-site WUE is the ratio of on-site water consumption to server energy consumption and varies with outside temperature and humidity.
- Off-site water efficiency: Off-site WUE measures electricity water intensity and varies over time and across regions as grid fuel mixes change.
- Operational water footprint: The operational footprint combines time-slotted AI energy use, data-center PUE, on-site WUE, and off-site WUE.Energy can be measured with power meters or server tools, while PUE accounts for non-IT energy overhead.
- Embodied water footprint: Embodied water is amortized over the server lifespan by multiplying the fraction of the assessment period by total manufacturing water.The total footprint adds this embodied component to operational water.
- GPT-3 case study: GPT-3’s case study estimates operational water using Microsoft’s location-wise PUE and WUE, an estimated training energy of 1287 MWh, and a 0.004 kWh per-request conversation workload.Embodied water is excluded from the case study because public scope-3 water data are unavailable; Table 1 summarizes the results, including projected values for data centers under construction.
4 Our Recommendations
The recommendations address AI’s water footprint through scheduling, transparency, and reporting across operational, off-site, and supply-chain impacts.
- More Transparency and Comprehensive Reporting: AI model cards and cloud dashboards should report water consumption to improve transparency and inform end users about usage impacts.Current model cards commonly report carbon emissions but omit water consumption.
- More Transparency and Comprehensive Reporting: Scope-2 water consumption should become standard reporting because visibility into off-site impacts can unlock demand-side flexibility.Existing cooling efforts primarily address scope-1 water use while overlooking scope-2 impacts.
- More Transparency and Comprehensive Reporting: Further research on embodied water from chip manufacturing is needed because data on scope-3 supply-chain water use remains limited.The paper links this gap to achieving a comprehensive understanding of AI’s overall water footprint.
- Scheduling: Dynamic scheduling can reduce water footprint by choosing when and where to train or run large AI models.Water efficiency varies with weather conditions and electricity-grid fuel mixes.
- Scheduling: Carbon-efficient and water-wise scheduling may conflict because carbon intensity and scope-2 water consumption can be misaligned.The paper contrasts following the sun for solar availability with unfollowing the sun to avoid high-temperature hours.
5 Conclusion
The conclusion frames AI water use as a critical sustainability concern and summarizes the paper’s methodology, findings on runtime efficiency, and reporting recommendations.
- AI water usage is presented as a critical concern for socially responsible and environmentally sustainable AI.
- The paper combines water-footprint estimation, spatial-temporal efficiency analysis, transparency recommendations, and holistic carbon-water accounting.
- AI’s water footprint should be addressed as a priority amid collective efforts to combat global water challenges.
Appendix: Operational Water for Global AI in 2027
The appendix estimates global AI’s 2027 operational water use from electricity demand and water-intensity factors, while noting uncertainty in future efficiency assumptions.
- Estimation basis: 85–134 TWh is the conservative projected global AI electricity demand used to estimate 2027 water usage.A more aggressive U.S.-only projection is also noted but not used for the global estimate.
- Estimation basis: 43.83 L/kWh and 3.14 L/kWh are the U.S. average electricity water withdrawal and consumption intensity factors used for scope-2 estimates.
- Scope-2 water usage: 4.10–6.46 billion cubic meters is the estimated scope-2 water withdrawal, while 0.29–0.46 billion cubic meters is the estimated scope-2 water consumption.These values use total electricity consumption of 93.5–147.4 TWh after applying PUE 1.1.
- Total water usage: 4.2–6.6 billion cubic meters of total water withdrawal and 0.38–0.60 billion cubic meters of total water consumption may be attributed to global AI in 2027.The withdrawal estimate exceeds the equivalent of annual withdrawals in 4–6 Denmark or approximately half of the United Kingdom.
- Uncertainty: The 2027 estimates are uncertain because future water efficiency may differ from the current values used.The paper nevertheless characterizes its estimates as conservative.