Water & Environmental Impact

The environmental footprint of AI extends beyond electricity. Water consumption for cooling, land use for data centers and solar farms, the lifecycle impact of GPU manufacturing, and e-waste from hardware turnover are all part of the full picture. This page covers the non-energy environmental dimensions that are receiving increasing attention from regulators and researchers.

Water Consumption

Data centers consume water in two ways: for cooling (evaporative cooling towers lose water to evaporation) and sometimes to generate electricity (hydroelectric and thermoelectric power plants consume water). AI data centers are significant water consumers:

• Microsoft reported that its global data center water consumption increased 34% in 2022 (FY2022) - the year ChatGPT was being trained - reaching 6.4 million cubic meters.
• A 2023 study estimated that training GPT-3 consumed approximately 700,000 liters of fresh water for cooling.
• A single ChatGPT conversation of 20–50 questions is estimated to require approximately 500 mL of water for cooling - roughly a standard water bottle.
• At scale (100M users, 10 queries each per day): ~5 billion liters of water equivalent per day - comparable to the daily water use of a mid-sized city.

Water Metrics: WUE and WCF

• WUE (Water Usage Effectiveness) = total water consumed (liters) / IT equipment energy used (kWh). Industry average: 1.8 L/kWh. Best-in-class: <0.5 L/kWh.
• WCF (Water Consumption Factor) = liters per unit of compute output. Harder to standardize but increasingly reported.

Strategies to reduce water use:

• Liquid cooling (closed-loop, no evaporation) vs evaporative cooling towers
• Site selection in cooler climates where free cooling (outside air or groundwater) reduces or eliminates the need for evaporative cooling
• Using non-potable water (recycled water, seawater, reclaimed wastewater) for cooling - some facilities in coastal areas use seawater cooling loops

Land Use

A large AI data center (100MW+) typically occupies 50–200 acres directly, but the associated renewable energy infrastructure requires significantly more land:

• A 100 MW solar farm requires ~500–1,000 acres (at 10–20 MW per acre in a good solar location)
• Wind power: 1–2 MW per turbine; a 100 MW wind farm requires 20,000–50,000 acres of turbine spacing (though land between turbines can still be farmed)
• Concentration in certain regions (Phoenix, Northern Virginia, Singapore, Amsterdam, Dublin) is creating land supply constraints - Northern Virginia, the world's largest data center market, is approaching build-out limits

GPU Manufacturing and Hardware Lifecycle

The embodied carbon in AI hardware - the emissions from manufacturing GPUs, cooling systems, and servers - is often overlooked in discussions focused on operational energy:

• Manufacturing a single NVIDIA A100 GPU is estimated to emit approximately 50–150 kg CO₂ equivalent (semiconductor fabrication is energy and chemical intensive)
• An H100 GPU has an estimated manufacturing carbon footprint of 150–300 kg CO₂e
• A large training cluster with 10,000 H100s: manufacturing carbon of ~1.5–3 million kg CO₂e before the first computation runs
• Hardware upgrade cycles in AI are extremely rapid: 2–3 years from H100 to B200 to next generation. Each cycle potentially retires hardware that still functions, creating e-waste.

Sustainable hardware practices:

• Extending hardware lifetime - using GPUs for inference after training on newer hardware
• Responsible recycling programs (NVIDIA, Google, and Microsoft all have hardware recycling commitments)
• GPU marketplaces (Vast.ai, CoreWeave secondary markets) that give retired data center GPUs a second life for smaller operators

Environmental Reporting Requirements

Disclosure requirements for tech companies' environmental impacts are increasing:

• SEC Climate Disclosure Rule (US, 2024) - large public companies must disclose Scope 1 and 2 emissions; Scope 3 (value chain) disclosure for very large companies
• EU Corporate Sustainability Reporting Directive (CSRD) - broad sustainability reporting for large EU companies and non-EU companies with significant EU revenue, including AI infrastructure operators
• EU AI Act - requires transparency about energy and environmental impact of AI systems, especially for high-risk applications
• Major hyperscalers (Google, Microsoft, Amazon, Meta) now publish annual sustainability reports that include data center energy and water metrics