Analysis
NVIDIA launched Minerva, an AIOps platform that combines telemetry from infrastructure and AI workloads and applies AI models to it. According to the announcement, it connects signals across GPUs, networking, storage, schedulers and workloads into a knowledge graph, analyzes billions of data points per hour, and can group related events into a single incident with a likely root cause. NVIDIA frames the goal as AI factory yield, described through GPU hours, goodput, cluster uptime and token production.
IREN is quoted as already using Minerva. Aolani, Bitdeer AI, IBM, Jio, Lambda, Quake AI and Tesla are named as planning deployments, and NVIDIA says Dell Technologies, Lenovo and Supermicro will offer it as part of their AI factory operations portfolios. The announcement promises longer time between interruptions, faster time to repair and fewer, consolidated alerts, but it does not report measured results.
Analysis: the useful shift is the unit of account. An operator who reports uptime alone can miss a cluster that is running but not finishing useful work. Measuring goodput forces a question that an alert queue does not: how much of the paid capacity produced completed jobs? Before adopting any platform of this kind, a team should agree how each yield measure is calculated, which fixes remain behind a human approval, and how predicted failures will be checked against what actually happened.
Source: NVIDIA Unveils Minerva to Bring End-to-End Observability for AI Factories · 2026-10-22
Why it matters
Define how yield is measured and which automated fixes need human approval before letting an observability platform act on a cluster.