How Much Energy Do Data Centers and AI Use?
Hannah Ritchie quantifies the global energy footprint of artificial intelligence and digital infrastructure
A landmark analysis by Dr. Hannah Ritchie at Our World in Data provides a rigorous, quantitative assessment of the global electricity consumption of data centers and artificial intelligence. The study reveals that data centers consume around 1.5 percent of global electricity, with AI-focused data centers specifically accounting for roughly 0.5 percent of the world's total power. This findings clarify a highly contested debate, highlighting that while the global footprint remains manageable, local grid concentration represents the true bottleneck for the industry's expansion.
Key details
According to data compiled from the International Energy Agency (IEA), global data centers consumed approximately 485 terawatt-hours (TWh) of electricity in 2025. This amount is equivalent to the entire annual electricity generation of Germany and represents 1.5 percent of the world's total generation (estimated at 31,800 TWh by energy think-tank Ember). Within this total, non-AI digital infrastructure consumed about 330 TWh (two-thirds), while AI-focused data centers consumed 155 TWh (one-third). Consequently, AI models and accelerators consumed approximately 0.5 percent of global electricity in 2025.
Alternative historical estimates from the Energy Institute and S&P Global place data center electricity consumption significantly higher, at 790 TWh for 2025, or 2.5 percent of the global total. Much of this 305 TWh divergence is explained by S&P Global's inclusion of cryptocurrency mining, which accounts for 150 to 200 TWh of demand. The rest is due to methodological differences; S&P utilizes a capacity-based bottom-up model, whereas the IEA estimates IT equipment and cooling overhead directly. Looking forward, the IEA's base-case projection predicts that total data center power demand will nearly double to 945 TWh by 2030, with AI-focused facilities growing to 465 TWh and completely closing the gap with non-AI infrastructure.
On an individual level, the analysis highlights that standard chatbot text queries have a relatively minor energy footprint. A median Google Gemini query is estimated to consume 0.24 watt-hours (Wh), while standard OpenAI ChatGPT queries consume roughly 0.3 to 0.34 Wh. However, agentic tasks and long-context processing consume far more power. A long text query (7,500 words of input) requires 2.5 Wh, while a maximum-context query (75,000 words) jumps to 40 Wh. Standard requests to an AI agent average 1.1 Wh, but agentic tasks requiring complex multi-step reasoning require up to 50 Wh per execution—equivalent to the power needed to run a television for over twenty minutes.
Why this matters
The Our World in Data report shifts the focus of the AI energy debate from global panic to local realities. While a 0.5 percent global electricity footprint indicates that AI is not currently on track to crash the world's energy systems, the extreme geographic concentration of data center infrastructure poses immediate challenges. In countries like Ireland, data centers already consume over 20 percent of national electricity, while in states like Virginia, they represent more than 25 percent of the total grid demand. These highly localized loads threaten clean energy goals and strain transmission infrastructure.
Context
The explosive growth of generative AI since late 2022 has triggered intense speculation regarding its environmental impact. Tech giants have faced criticism for rising emissions, with Google's 2026 sustainability report showing a 37 percent surge in power demand and Amazon reporting an 18 percent emissions increase due to data center expansion. These increases are primarily driven by inference—running the models—rather than training, which remains a minor fraction of overall compute energy. To maintain their carbon-free power commitments, hyperscalers are increasingly investing in modular cooling, nuclear contracts, and behind-the-meter natural gas generation.
Risks and open questions
The primary risk associated with the rapid growth of AI infrastructure is the carbon intensity of the grids supplying the power. An AI query processed in a data center connected to a coal-dominated grid has a significantly higher carbon footprint than one served by nuclear or renewable energy. Furthermore, the wide range of future projections creates massive uncertainty for utility planners. If efficiency gains in chip architectures, such as NVIDIA's newly produced Vera Rubin platform, fail to offset rising user demand, the power deficit could widen, forcing utilities to extend the lifespan of fossil-fuel generation.
What happens next
To address grid constraints and public pushback, the data center industry is shifting toward on-site generation and zero-water cooling designs. Companies are deploying closed-loop liquid systems to eliminate the millions of gallons of daily water use traditionally required for cooling. Additionally, policymakers are ramping up regulatory oversight. With the White House's Ratepayer Protection Pledge and several states proposing data center tax incentive repeals or moratoriums, AI operators will likely face stricter resource-reporting mandates and higher capital costs to fund direct grid upgrades.
Source: Our World in Data Published on AI Usage Global, author: AUG Bot
