AI Data Centers Are Driving Up Energy Use, But Transparency Is Lacking

AI is rapidly increasing energy demands, but the true scale remains unclear due to a lack of transparency from tech companies.

A recent analysis by Nature highlights the growing but largely opaque energy demands of AI. Nature reports how in Virginia’s Culpeper County, large data centers are transforming rural landscapes. These centers, essential for running generative AI models like ChatGPT, require massive amounts of electricity.

Each facility can consume as much power as tens of thousands of homes, potentially driving up costs and straining local grids. Virginia, already the world’s data-center capital, could see its electricity demand double in the next decade.

Nature notes that the issue extends beyond Virginia. AI-driven data centers are concentrated in clusters worldwide, significantly impacting local energy grids. Unlike steel mills or coal mines, data centers are built close together to share resources and optimize efficiency.

In Ireland, they account for over 20% of national electricity consumption, and in five U.S. states, they exceed 10%.

Despite AI’s rising power consumption, data on its energy use is scarce. Nature reports tha researchers struggle to obtain precise figures from companies, forcing them to estimate using indirect methods. One approach is supply-chain analysis.

In 2023, researcher Alex de Vries calculated that if Google integrated generative AI into all searches, it would require up to 500,000 NVIDIA A100 servers, consuming 23–29 terawatt hours (TWh) annually—up to 30 times more energy than a standard search, as reported by Nature.

Another method involves measuring the energy use of individual AI tasks. Researchers use tools like CodeCarbon to estimate consumption from AI-generated images or text.

These studies suggest that generating an image consumes about 0.5 watt-hours (Wh), while text generation requires slightly less. However, these estimates are conservative since they don’t account for cooling or proprietary chips like Google’s TPUs, as reported by Nature.

Training AI models is also energy-intensive, but the energy spent answering billions of user queries is even greater. Training a model like GPT-3 requires around one gigawatt-hour, whereas daily AI queries consume terawatt-hours annually, says Nature.

With increasing competition, companies have become more secretive about AI’s energy demands. Some, like Google and Microsoft, acknowledge rising carbon emissions due to data-center expansion but do not provide specific data, noted Nature.

Despite AI’s local impact, its global energy footprint remains relatively small, argues Nature. The International Energy Agency estimated that data centers used 240–340 TWh in 2022, about 1–1.3% of global electricity.

However, as AI adoption grows, demand could surge. Without better data-sharing, policymakers may struggle to manage the environmental consequences of AI’s energy-intensive future.