The Rising Energy Demand of AI: A Closer Look
As artificial intelligence (AI) becomes increasingly integrated into our daily lives, its energy demands have sparked a global conversation. This rapid adoption of AI raises questions about its impact on energy consumption and the broader implications for the environment and local communities.
AI’s energy consumption concerns can be categorized into three main areas: environmental impact due to potential increases in carbon emissions, local community effects from strained electricity supplies and rising energy prices, and the possibility of energy consumption acting as a bottleneck for AI expansion.
Understanding AI’s energy usage involves examining both the total electricity consumption by AI and the energy required for individual AI queries.
AI’s energy consumption includes the electricity used for both training AI models and running them (inference). Estimates suggest that most of the energy is consumed during inference rather than training. These estimates account for the electricity used by servers and additional energy needed for cooling and lighting in data centers but exclude the energy used by devices accessing AI, as well as cryptocurrency mining.
Globally, data centers consumed approximately 485 terawatt-hours (TWh) of electricity, equivalent to around 1.5% of the world’s electricity generation, according to the International Energy Agency (IEA). To put this into perspective, it matches Germany’s annual electricity production. However, another estimate by the Energy Institute, which includes cryptocurrency mining, suggests data center demand could reach 790 TWh by 2025, making up 2.5% of global electricity generation.
Data centers support a wide range of digital services, not just AI. AI-focused data centers currently consume about one-third of the total electricity used by data centers. This means AI accounted for approximately 0.5% of global electricity consumption in 2025, translating to 0.1% to 0.2% of the world’s primary energy consumption.
Despite AI’s rapid growth, AI-focused data centers still consume less electricity overall compared to non-AI data centers. However, projections indicate this gap may close by 2030, with AI data centers potentially consuming as much as non-AI ones.
The geographical concentration of data center demand poses a significant challenge. In the United States, data centers consume about 5% of electricity, and this figure can be much higher in certain states. In Europe, Ireland’s data centers account for more than 20% of electricity consumption.
Individual AI queries, such as those made to language models like ChatGPT, consume relatively small amounts of energy. Google estimated that a text-based query to its Gemini model uses around 0.24 watt-hours (Wh) of electricity. OpenAI’s CEO, Sam Altman, reported a similar figure for ChatGPT. More complex or lengthy queries consume more energy, but even these remain small compared to average daily electricity consumption in high-income countries.
As AI continues to evolve, its future energy demands remain uncertain, influenced by factors like user demand growth and hardware efficiency improvements. The IEA has projected various scenarios, with differing growth rates for user demand and efficiency gains.
Ultimately, the impact of AI on energy consumption and climate change depends not only on its energy usage but also on how that energy is generated. The environmental implications of AI hinge on whether data centers are powered by renewable energy or fossil fuels.
Original Story at ourworldindata.org