AI Daily — October 13, 2026
2026-10-13
engineering.fb.com
OCP Summit 2026: Powering AI with Open Innovations in Scale-Up, Scale-Out, and Scale-Across Networks
Meta is open-sourcing several major AI infrastructure components at OCP Summit 2026, including IcePack (liquid-cooled network rack), IceCube (102.4T switch platform), and Flattened Planar Fabric (FPF), a multiplane backend fabric that rearchitects GPU-to-network connectivity. They are also releasing MetaRoCE, an open-source AI transport protocol designed for large-scale GPU cluster communication. These contributions represent a significant push to commoditize the networking stack underlying large AI training clusters.
engineering.fb.com
Turning Opportunistic AI Inference Capacity Into Better Ads Recommendations
Meta describes a production system that reclaims idle GPU inference capacity fleet-wide and routes it to ads recommendation workloads via two elasticity levers, improving model quality without degrading baseline serving performance. The system operates at Meta's full fleet scale and represents a practical approach to improving inference utilization in heterogeneous, latency-sensitive environments. This kind of opportunistic scheduling is increasingly relevant as inference compute costs dominate AI operational budgets.
www.technologyreview.com
AI labs are buying data from dead startups to train their models
AI labs are acquiring training data by purchasing digital archives from failed startups, including code bases and user-generated content, often for relatively modest sums. This practice raises unresolved questions around user consent, data provenance, and regulatory compliance, particularly when the original companies had privacy policies that didn't anticipate asset liquidation to AI firms. The trend signals growing pressure on labs to find novel data sources as conventional web-crawl datasets saturate.