MIT professor uses ML to squeeze bloat out of data centers—most large systems run at ~15% utilization
MIT News AI · rss · 2026-10-08
MIT News profiles newly tenured Christina Delimitrou, whose group applies machine learning to make large-scale data centers more efficient, secure, and reliable—coaxing more compute from existing hardware so fewer new facilities need to be built, easing strain on power grids.
Highlights:
- An early key finding with mentor Christos Kozyrakis: despite heavy demand, most large computing systems run at only 15% capacity.
- Seer uses deep learning to anticipate and prevent failures in web applications before they happen.
- She rebuilt ML resource-management systems for microservice-style applications that legacy servers weren't designed for.
- Because operators use inaccessible proprietary hardware and software, her group built Ditto, which clones a system's structure and performance characteristics for research.
- A current focus is adding explainability to these ML tools so developers can verify answers and gain design insight.
The piece also covers her path from Greece to Stanford and her teaching philosophy for MIT's 350-student computation structures course.
More from Infra
- PartyKit shuts down free hosted platform 2.5 years after Cloudflare acquisition — threepointone · 2026-10-09
- 3090+3060 local video gen: which model — Wan 2.2, MiniMax H3 or LTX-2? — itspacedev · 2026-10-09
- Samsung expects 780% quarterly operating profit jump on AI boom — chemist_slime · 2026-10-09
- TRL v1.15 ships fused LM head: 82% less peak VRAM, 7x longer sequences — QGallouedec · 2026-10-09
- Dev streams a 66GB unquantized 22B video model on an iGPU with only 15.6GB shared RAM — Business_Swordfish_5 · 2026-10-09
- Asia-Pacific Needs $280 Billion to Build 26.5 GW of Planned Data Centers, Johor Leads — shashib · 2026-10-09