Michael JohnsonA failure in a large AI cluster can start with a subtle deviation: one GPU runs hotter than the others, storage slows under load, and a training job falls behind. On October 22, 2026, NVIDIA introduced Minerva, an AIOps platform that brings together telemetry from infrastructure and running AI workloads in a unified view.
According to NVIDIA’s description, the system correlates data from GPUs, networking equipment, storage, schedulers, and jobs. Its practical aim is to help operators link a job’s degraded performance to a likely problem with a specific component, rather than having to inspect disconnected monitoring dashboards.
From individual metrics to a map of relationships
Minerva builds a dynamic knowledge graph linking physical infrastructure—including GPUs, SuperNICs, switches, and cables—to running jobs. NVIDIA says the system processes billions of data points per hour, learns the cluster’s normal behavior, detects anomalies, and correlates related events based on topology.
For on-call teams, these connections matter when diagnosing cascading problems: an operator can trace how the degradation of one element affects a job using that part of the cluster. Minerva also groups related signals into an incident with a likely root cause, reducing the stream of individual alerts.
Recommendations remain part of the workflow
The platform includes a natural-language assistant that answers questions about cluster status using current telemetry and suggests ways to resolve problems. A team can apply a fix manually or connect automation to infrastructure APIs and external agent tools. An important practical question when deploying the platform is therefore which actions remain recommendations and which the operator authorizes it to perform automatically.
NVIDIA describes Minerva’s target metrics in terms of useful AI factory output: GPU hours, goodput, cluster availability, and token throughput. This shifts the focus of monitoring from alert counts to the relationship between infrastructure health and the compute work actually completed.
Announcement and deployment
NVIDIA says that IREN is among the companies already using Minerva to optimize AI factories. Aolani, Bitdeer AI, IBM, Jio, Lambda, Quake AI, and Tesla plan to deploy the platform; Dell Technologies, Lenovo, and Supermicro intend to offer it as part of their AI factory operations solution portfolios.
The October 22 announcement establishes the platform’s availability and the plans of the named partners, but the published material does not provide comparable independent measurements of how much Minerva reduces downtime or troubleshooting time. A useful starting point for operators is to check whether the system covers their telemetry sources, connect its incidents to their existing on-call process, and define the boundaries of automated actions separately. The approach’s core value is bringing infrastructure and workload signals together around a single task: understanding where useful compute capacity is being lost and why.