Nvidia Launches Minerva Platform for AI Factory Observability

Nvidia launched Minerva on October 22, 2026, an AIOps platform meant to give operators of large AI factories one place to watch infrastructure and workloads instead of several. The company says traditional monitoring tools split that view across networking, compute, site reliability and machine learning teams, so a slow training job can take time to trace back to the one overheating GPU or degraded NVMe drive behind it. Minerva builds what Nvidia calls a dynamic knowledge graph, logically mapping GPUs, SuperNICs, switches, cables and the jobs running across them, and applies supervised and unsupervised machine learning to learn normal behavior, flag anomalies it has not seen before and group related alerts into one incident with a likely root cause rather than a flood of separate warnings.

The platform also ships with a native AI assistant that connects to live telemetry, letting an operator ask about cluster health in plain language and get a recommended fix back, with a person deciding whether to apply it manually or let Minerva push it through automation and infrastructure APIs. Nvidia frames the whole platform around a single idea it calls AI factory yield, the useful output a cluster produces measured through GPU hours, goodput, uptime and token production, rather than through uptime alone.

The numbers

Nvidia did not publish a count of alerts suppressed or dollars saved, so the evidence here is about scale and partners rather than a verified figure. The company says Minerva can trace a single slow job back to its cause across a factory running hundreds of thousands of GPUs, and that its models process the telemetry at a rate of billions of data points every hour. On the customer side, IREN is already using Minerva in production, with Aolani, Bitdeer AI, IBM, Jio, Lambda, Quake AI and Tesla named as planning deployments. Dell Technologies, Lenovo and Supermicro will offer the platform through their own AI factory operations portfolios, which puts Minerva on the shelf for buyers who never deal with Nvidia directly.

IREN’s chief technology officer, Denis Skrinnikoff, said the platform gives his operations team one view across the network, GPUs and jobs together, framing it as a move from troubleshooting toward proactive maintenance. That is the pitch Nvidia is making to every operator buying GPU capacity at scale: the chips are the same chips, but the software layer watching them decides how much of that capacity actually ships as finished work.