📅 October 31, 2026 • University of Michigan
DARPA has awarded a 1.83 million dollar, two-year contract for a project called E-CALC, short for Energy Calculus for ML Lifecycle Optimization. The effort is led by the University of Michigan together with the University of Illinois Urbana-Champaign and BAE Systems. Its goal is to measure and reduce the energy consumed by AI across the entire machine learning lifecycle, from data processing through training to deployment.
Background
Energy use has moved from a side concern to a central constraint in AI development. Training runs and inference at scale translate directly into electricity bills and grid load, and hardware improvements alone have not kept pace with the growth in demand. E-CALC approaches the problem from the software side, asking how much of the consumption is determined by design choices rather than by raw hardware capability.
What will be measured and how
The team will build open-source tools that combine physical sensors with software analysis to estimate energy use at every stage of an AI workflow. This covers data processing, model training and deployment, giving engineers a breakdown of consumption rather than a single aggregate number. The motivation comes from a known inefficiency: when software is not aligned with the hardware it runs on, up to 30 percent of the electricity used in AI model training alone can be wasted. By profiling where energy actually goes, the tools are intended to show developers and operators exactly which parts of a pipeline are inefficient and where redesign would pay off.
Why it matters
The researchers estimate that better software design could reduce AI energy use by up to 50 percent, a large margin for what is essentially an engineering and measurement problem. The broader context is demand growth: the International Energy Agency (IEA) expects data-center electricity consumption to more than double by 2030, rising from about 415 TWh in 2024. Against that backdrop, measurement tools that span the whole lifecycle give data-center operators, cloud providers and model developers a concrete way to understand and shrink their share of consumption.
Impact
If the project delivers, energy profiling could become a routine part of machine learning engineering, similar to how accuracy and latency benchmarks are standard today. Open-source tooling would lower the barrier for smaller teams that cannot run their own instrumented hardware, and the involvement of BAE Systems suggests the results are also intended for defense and industrial applications. The DARPA funding also signals that energy efficiency is increasingly treated as a first-class research problem alongside model capability.
Sources: news.engin.umich.edu, IEA
Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky
Top comments (0)