AI能源使用的环境成本_碳足迹_水足迹与土地足迹_56页_26mb

Environmental Cost of AI's Energy Use: Carbon, Water and Land Footprints

Core Content

This report by the United Nations University Institute for Water, Environment and Health (UNU-INWEH) examines the environmental impact of artificial intelligence (AI), focusing on its carbon, water, and land footprints. It highlights the growing energy demand of AI systems, particularly in the context of training and inference phases, and explores the geopolitical and social implications of this expansion.

Main Points

AI's Rapid Growth

• The global AI market is projected to grow from USD 189 billion in 2023 to nearly USD 5 trillion by 2033, a 25-fold increase over a decade.
• In 2026, the global AI expenditure is expected to exceed USD 2.5 trillion.
• Generative AI accounts for over 20% of the global AI market in 2026 and is projected to reach 40% by 2030.
• Corporate AI investment in 2025 surpassed USD 580 billion, with generative AI alone attracting USD 34 billion in private investment.
• 78% of organizations use AI, and 40% of employers expect to reduce their workforce where AI can automate tasks.
• 60% of jobs in advanced economies incorporate AI, versus 26% in low-income countries.

Energy Demand and Environmental Footprints

• AI training is highly energy-intensive. For example, GPT-4 training consumed 50–70 GWh of electricity over 100 days, 40–55 times more than GPT-3 (1.287 GWh over 34 days).
• Training GPT-5 is expected to require 100 GWh, equivalent to the annual residential electricity use of 770,000 people in Sub-Saharan Africa.
• The carbon footprint of GPT-4 is 25,000 tonnes CO₂e, which would require 420,000 tree seedlings grown for 10 years to offset.
• The water footprint of GPT-4 is 600 million liters, enough to meet the minimum annual domestic water needs of 81,000 people in Sub-Saharan Africa.
• The land footprint of GPT-4 training is about 0.9 km² (equivalent to 126 football fields), and for GPT-5 it is 1.5 km² (about 210 football fields).

Inference and AI in Use

• Inference, the phase where AI models generate responses, accounts for 80–90% of total AI energy use.
• A ChatGPT-style text query is 200 times more energy-intensive than text classification (e.g., spam filtering).
• An AI image requires 2.9 Wh of electricity, 1,450 times more than text classification.
• A short AI video can draw as much electricity as 200,000 spam classifications or hundreds of AI-generated images.
• The energy for a typical AI image is enough to power a 10-watt LED bulb for 17 minutes, while a high-complexity AI video would run it for 42 hours.
• The water footprint of a single AI image is about 29 mL, but for a complex AI video, it jumps to 4.1 liters, equivalent to two days of drinking water for one person.

Data Centers as AI Infrastructure

• In 2025, global data centers consumed 448 TWh of electricity, equivalent to the annual residential electricity needs of 1.3 billion people in Sub-Saharan Africa for 2.6 years.
• The carbon footprint of this electricity was 189 million tonnes CO₂e, which would require 3.2 billion tree seedlings over 10 years to offset—roughly twice the number of trees in the UK.
• The water footprint of 2025 data centers is 4.5 trillion liters, enough to fill 1.8 million Olympic-sized pools or meet the annual domestic water needs of 600 million people in Sub-Saharan Africa.
• The land footprint of 2025 data centers is 6,900 km², 4.5 times the size of Greater London.
• By 2030, data centers could consume 945 TWh, which would be enough to power all 1.3 billion residents of Sub-Saharan Africa for 5.5 years, and would rank 6th globally by electricity consumption.
• The water footprint of 2030 data center electricity use is projected to be 9.3 trillion liters, equivalent to the annual domestic water needs of 1.3 billion people in Sub-Saharan Africa.
• The land footprint of 2030 data center electricity use is estimated at over 14,500 km², about twice the size of Jakarta.

Environmental and Social Impacts

• The physical lifecycle of AI hardware presents a growing crisis, with 2.5 million metric tons of e-waste potentially generated annually by 2030, equivalent to 250 Eiffel Towers.
• AI infrastructure relies on critical minerals extracted in regions with weak environmental oversight, contributing to water depletion, pollution, and health impacts.
• The concentration of AI infrastructure in a few countries (primarily the US and China) exacerbates global inequalities in access to AI and its benefits.

Key Findings

• AI is a transformative force that reshapes economies, labor markets, and global systems.
• The environmental footprint of AI is not uniform and is heavily influenced by electricity supply mixes.
• Training is a major component of AI's energy use, but inference accounts for 80–90% of total energy consumption.
• AI's energy, water, and land use can have significant local and global consequences, often shifting burdens to vulnerable communities.
• Efficiency improvements are important but may not be sufficient to counteract rebound effects where lower per-use impacts lead to higher usage volumes.

The Way Forward

The report proposes a responsible AI ecosystem based on six operational principles:

1. Transparency
2. Efficiency by design
3. Equity and environmental justice
4. Lifecycle responsibility
5. Global cooperation
6. Sustainable use

These principles aim to ensure governance and accountability across jurisdictions, and to measure and mitigate AI's environmental impact. The report calls for standardized environmental footprint reporting, environmental impact screening, and community safeguards in AI infrastructure development.

Conclusion

This report underscores the critical need to address AI's environmental footprint as part of its governance and sustainability planning. It emphasizes that AI is not "just code" but involves physical infrastructure and supply chains that have real-world environmental and social consequences. By adopting a comprehensive and inclusive approach, the international community can ensure that AI's growth remains environmentally manageable and equitably distributed.