MIT Research Uses AI to Cut Data Center Energy Waste

The global surge in data center construction is stressing power grids and increasing reliance on fossil fuels. At MIT, newly tenured associate professor Christina Delimitrou is tackling this challenge by applying machine learning to make existing server fleets far more efficient.

Software bloat, not just hardware, drives waste

Delimitrou's group at CSAIL targets the "bloating" in cloud software stacks that forces operators to over-provision hardware. By redesigning resource schedulers and runtime systems with AI, they extract more compute from the same silicon — delaying the need for new builds and cutting both capital and operational expenditure.

Explainability is a core focus. The team builds interpretable models so engineers can trust and act on AI-driven tuning recommendations, from container placement to network traffic shaping.

Beyond infrastructure, the same techniques help developers detect and resolve performance regressions in production services like streaming and video conferencing, eliminating wasteful retry loops and idle cycles.