The Cost of AI: Environmental Impact and Energy Consumption
The computational demands of training and running AI systems have made the technology one of the most significant contributors to data center energy growth globally. Understanding the actual scale of this impact — as distinct from the inflated claims found at both extremes of the debate — requires engaging with verified data on energy consumption, water usage, and the carbon footprint of specific AI workloads.
The International Energy Agency estimated that global data center electricity consumption reached approximately 460 terawatt-hours in 2022 and has grown substantially since, with AI workloads representing an increasingly large share. AI model training — the process of exposing a model to vast amounts of data to develop its capabilities — is the most energy-intensive phase. Training GPT-4, according to estimates published by researchers at the University of Washington and elsewhere, is estimated to have consumed the equivalent of more than 500 metric tons of CO2 equivalent, comparable to the annual emissions of approximately 100 passenger vehicles. These are estimates, not company-disclosed figures; OpenAI and most major AI companies do not publish detailed training energy data.
Inference — the process of running a trained model to answer a user query — is less energy-intensive per operation than training, but occurs billions of times per day across all deployed AI systems, and the aggregate inference energy consumption of all AI services globally has grown substantially alongside adoption. A ChatGPT query consumes approximately 10 times the energy of a standard Google search, according to estimates by Goldman Sachs Research — a figure that, when multiplied by hundreds of millions of daily queries, represents meaningful incremental energy demand.
Water consumption is a related environmental concern that receives less coverage than energy. Data centers use water for cooling, and some large AI facilities consume millions of gallons per day. Microsoft disclosed in its 2023 sustainability report that its global water consumption increased by 34% between 2021 and 2022, a period corresponding to significant AI infrastructure expansion.
The countervailing argument — made by AI companies and accepted by many environmental researchers — is that AI tools can reduce emissions in applications where they improve energy efficiency, accelerate clean energy development, or replace carbon-intensive human activity. Google’s DeepMind developed an AI system that reduced cooling energy consumption in its own data centers by approximately 40%, validated by independent Google engineering teams. AI-optimized grid management, AI-assisted materials discovery for battery technology, and AI weather modeling that improves renewable energy forecasting are documented examples of climate-positive AI applications.
The net environmental impact of AI depends on whether the efficiency gains enabled by AI applications exceed the energy costs of the AI infrastructure producing them — a calculation that varies by use case and is being actively studied by researchers at MIT, Carnegie Mellon, and environmental organizations. The honest answer in 2026 is that the industry’s energy footprint is real, growing, and inadequately disclosed, while the potential for AI to contribute positively to energy transition is also real and actively being pursued. Both can be true simultaneously, and neither justifies ignoring the other.
