The United Nations University (UNU) Institute for Water, Environment and Health released a report in 2026 warning that artificial intelligence’s environmental footprint—including electricity, water, and land use—will grow rapidly and could rival that of major countries by 2030 [1, 2, 3, 4, 5].
AI-related electricity consumption is projected to triple the combined annual power use of Pakistan, Bangladesh, and Nigeria, a population of about 650 million people [1]. The power demands of AI data centers are causing delays in grid connections, with bottlenecks and transformer availability issues lasting up to 1.5 to 2 years [6]. This shift prompts power markets to value supply reliability and capacity more than just electricity cost [6].
AI systems could generate as much as 400 million tons of CO2 annually by 2030, roughly matching current UK emissions [1]. Meanwhile, annual water consumption for AI infrastructure worldwide could reach approximately 9.3 trillion liters, more than the domestic water needs of 1.3 billion people in sub-Saharan Africa [1, 2, 3, 4, 5].
The water usage includes on-site server cooling, water used in generating electricity for data centers, and water in semiconductor chip manufacturing [2]. Training a single AI model such as GPT-5 may consume up to 1 billion liters of water, compared to 600 million liters for GPT-4 [2]. For individual users, ChatGPT consumes about 500 milliliters of water per prompt, Google’s Gemini about 0.26 milliliters per prompt, while AI video and image generation consume up to 4.1 liters and 29 milliliters per output, respectively [2].
Taiwan’s semiconductor fabs use a small fraction of local water compared to domestic use, but dry regions face growing water stress due to AI infrastructure expansion [3]. 「AI的伺服器它所需要的一些就是說生產的資源，上從晶圓代工開始到中游的像PCB板等等，這一些其實都需要耗費大量的水資源」, said 柴煥欣 of 雲報政經產業研究院 [3].
The land footprint of AI infrastructure—including data centers and energy facilities—is expected to exceed 14,500 square kilometers by 2030 [1]. Older hydroelectric dams are valued for their existing permits, grid interconnections, and dispatchable power crucial for AI data centers [6]. Google and Brookfield Renewable signed a $30 billion deal in July 2025 to jointly operate US hydroelectric plants with 670 MW capacity to provide renewable power to AI facilities [6].
Experts warn that without proper governance, AI growth could worsen global freshwater shortages and inequalities due to concentration of infrastructure in few countries [4, 5]. They urge technology firms to improve transparency on water and energy use and recommend users reduce unnecessary AI workloads to curb resource strain [4, 5].
On June 3, 2026, UNU published projections showing AI demand for electricity and water could more than double by 2030, with water use reaching 9.3 trillion liters annually [6]. Multiple reports on June 8, 2026, highlighted the global risks of AI’s water and energy consumption through 2030 [1, 2, 3, 4, 5].
