FaultSense cuts MoE fault tests 20x

Microsoft Research’s FaultSense cuts diagnostic tests for Mixture-of-Experts serving clusters by up to 20×, using an application-layer probe to localize faulty GPUs or communication paths without host instrumentation. The September 2026 paper frames gray failures as slow components that inflate serving latency without throwing explicit errors. For operators, this is a reminder to budget fail-slow detection alongside GPU capacity; the archive’s analysis of hyperscaler AI capex shows why a small diagnostic tax matters at rack scale.