DARPA bets on ‘energy calculus’ to tame AI’s power appetite

DARPA is spending real money on a problem most AI companies would rather not discuss: their models eat electricity at an alarming rate.

The Defense Advanced Research Projects Agency has backed a $1.83 million project, led by the University of Michigan, that aims to build a mathematical toolkit for squeezing waste out of every stage of AI. The research team believes the approach could potentially cut energy use by as much as 50%.

What DARPA is actually paying for

The flagship effort is called E-CALC. Its partners include the University of Illinois Urbana-Champaign and defense contractor BAE Systems, and the award was announced in late October 2026.

The core idea is something the team calls “energy calculus.” The framework is meant to track and optimize power consumption across the full AI lifecycle. That covers data processing, model training and eventual deployment.

A second award points in a related direction. Stony Brook University, working with Ohio State University, received $1 million from DARPA on October 20, 2026, under a separate initiative known as the ML2P program.

That team is building tools to forecast how much energy a machine learning model will consume, without sacrificing the model’s accuracy.

Why the grid is sweating

The numbers behind the push are blunt. Data center electricity consumption is projected to climb past 840 terawatt-hours by 2030, up from approximately 420 TWh in 2024.

Part of the problem is plain waste. Mismatches between software and the hardware it runs on could squander up to 30% of energy during AI training alone.

The military angle

DARPA’s broader strategy is to bake energy awareness into AI development, especially for settings where power is scarce. Military operations are the obvious case. Drones and field equipment cannot plug into a hyperscale data center, and every watt spent on computation is a watt not spent on flight time or mission endurance.

BAE Systems’ involvement in E-CALC fits that picture. Defense contractors have a direct interest in AI that performs well on constrained hardware far from any power plant.

What this means

For the AI industry, the grants are a small signal about a big shift. A combined $2.83 million is pocket change next to what is being poured into data center construction, but DARPA has a long track record of funding foundational ideas that later spread into commercial use.

The risks are worth stating plainly. The 50% figure is a research target, not a demonstrated result, and academic frameworks often struggle to survive contact with messy real-world infrastructure. Efficiency gains can also be swallowed by growth: cheaper computation tends to invite more computation.