NVIDIA's Minerva maps GPUs, networks and jobs into one live graph so AI factories fix faults before jobs stall
Synopsis
NVIDIA released Minerva, an AIOps platform for end-to-end observability of AI factories built from large GPU fleets. Previously networking, compute, site reliability and machine learning teams each saw a separate slice of the data center, and alerts arrived disconnected. Minerva pools telemetry from GPUs, SuperNICs, switches, cables, storage, schedulers and workloads into a dynamic knowledge graph, then uses supervised and unsupervised learning to detect anomalies, predict failures and correlate events by topology, collapsing an alert storm into one incident with a likely root cause.
Interpretation
AI factory operations move from separate views to one view spanning infrastructure and jobs, so operators can trace a slow training job to the component affecting it along the topology. Traditional monitoring tools often split the data center into separate networking, compute, site reliability and machine learning views, so each team sees only part of the operation. Minerva connects signals across NVIDIA GPUs, networking, storage, schedulers and workloads and builds a dynamic knowledge graph from them, with scope extending to factories of hundreds of thousands of GPUs.
The platform turns billions of data points per hour into a small set of actionable incidents instead of sending every symptom to an operator. The prior condition is an operator facing a blizzard of separate alerts with no clear sign of which belong to the same fault. Minerva uses supervised and unsupervised machine learning to learn normal behavior, detect previously unseen anomalies, predict emerging failures and correlate related events by factory topology, grouping an event storm into a focused incident with a likely root cause.
Operators can query cluster health in natural language and receive remediation recommendations, with fixes either applied by a person or handed to automation. A native AI assistant and connections to external agentic AI tools give a natural-language interface grounded in live telemetry. Teams can keep a person in the loop to implement fixes manually or connect Minerva to automation through infrastructure application programming interfaces and external systems.
Perspective
The report covers NVIDIA's Minerva platform and the capabilities described for it: pooling telemetry from GPUs, networking, storage, schedulers and workloads, building a dynamic knowledge graph, using supervised and unsupervised learning for anomaly detection and failure prediction, and offering a natural-language assistant and automation interfaces. The intended users are AIOps, networking, compute and machine learning teams running large AI factories. IREN is using the platform to optimize its AI factories; Aolani, Bitdeer AI, IBM, Jio, Lambda, Quake AI and Tesla plan to deploy it, and Dell Technologies, Lenovo and Supermicro will offer Minerva as part of their AI factory operations portfolios.
The material describes benefits qualitatively, namely longer time between interruptions, faster time to repair and more focused alerts, without specific figures or measured comparison against existing monitoring tools. The platform's capabilities are framed around NVIDIA's own GPUs, SuperNICs, switches and cables, and behavior on other infrastructure combinations is not described. Among the partners, only IREN is using it, while the others plan to deploy, so operational results remain to be seen.