Vijay Gadepally cut AI training's power bill without losing the best model.
Cambridge/Lexington, MA, USA (MIT Lincoln Laboratory); also co-founder, Bay Compute · Vijay Gadepally
Published July 13, 2026
Training an AI model usually means burning energy on dozens of configurations that never end up winning. An MIT Lincoln Laboratory researcher built a way to catch the losing runs early and shut them down, then co-founded a startup to bring the same discipline to whole data centers.
The story
The person and the place
Vijay Gadepally, a researcher at MIT Lincoln Laboratory and co-founder of the startup Bay Compute.
The problem
AI training runs burn substantial energy testing configurations that never turn out to be the best-performing model, and the data centers running them rarely manage their power as one system.
The moment he didn't wait
Rather than let every training run burn to completion, Gadepally built and deployed a power-capping approach that tracks a model's performance mid-run and stops the underperforming ones early. "What we can do is essentially limit the amount of power that's going to a processor. That has a huge impact in terms of the power, but very little impact in terms of the actual performance of the processor," he told EE Times.
"we were able to very quickly eliminate nearly 80% of the runs that we were doing... we did not lose the best performing model... but we did get rid of nearly 80% of the energy use for training that model," Gadepally said. He then took the idea beyond the lab: Bay Compute now applies it to entire facilities, an approach he compares to "a Nest thermostat... conceptually, the same idea, but here applied to these much larger systems," with power savings he reports at up to 20% in its first data-center deployments.
"we did not lose the best performing model...but we did get rid of nearly 80% of the energy use for training that model." — Vijay Gadepally, Energy vs Climate podcast
Verified sources
Sources
Every claim in this story is checked against primary sources. Verify it yourself.