DARPA funds $1.83m project to measure AI energy consumption across workflow stages

The Defense Advanced Research Projects Agency has awarded a two-year grant of $1.83 million to a research consortium led by the University of Michigan Engineering. The project, designated E-CALC or Energy Calculus for ML Lifecycle Optimization, aims to develop software systems that accurately track energy consumption at every stage of artificial intelligence workflows. The initiative involves collaboration with the University of Illinois Urbana-Champaign and defense contractor BAE Systems to address growing concerns over the sector’s resource intensity.

The funding addresses a critical gap in current AI infrastructure, where insufficient measurement and reporting of resource use obscure the true environmental and economic costs of machine learning. Researchers indicate that misalignments between software and hardware can waste up to 30 percent of electricity during model training alone. The proposed system will allow developers to estimate energy usage from data processing and model training through to real-world deployment, replacing expensive and time-consuming trial-and-error testing across thousands of hardware and software configurations.

This development occurs against a backdrop of rapidly escalating energy demands driven by the global expansion of data centers. Industry projections suggest that data center energy consumption will more than double by 2030, rising from a 2024 baseline of approximately 420 terawatt-hours. While proponents argue that AI will eventually enhance efficiency across economic sectors, the immediate impact is a significant strain on global energy grids. The sector’s growth is currently driving the construction of natural gas-powered plants, extending the dominance of fossil fuels and complicating efforts to meet climate targets set by the Paris agreement.

Mosharaf Chowdhury, an associate professor of computer science and engineering at the University of Michigan and the principal investigator for the project, stated that AI is becoming integral to everyday life, business, science, and national security. He noted that the sector requires enormous and rapidly growing amounts of electricity, which, if unchecked, could increase costs, strain power infrastructure, and hinder the deployment of advanced AI systems. The project seeks to answer the practical question of how to achieve widespread AI benefits with minimum energy consumption.

The technology is expected to be particularly relevant in military and defense contexts, where field equipment and remote computing systems must operate complex AI algorithms under strict battery and power limits. By enhancing the ability to track energy use, the team aims to reduce overall electricity demand, lower the cost of AI, and extend battery life for AI-enabled devices. This approach represents a shift from waiting for AI to make the world more efficient to making AI itself more efficient through precise measurement and software design.