Michael JohnsonWhen model training slows down, the cause may lie beyond the GPU: in the network, storage, job scheduler, or another component on which the computing task depends. On October 22, 2026, NVIDIA introduced Minerva, an AIOps platform that connects infrastructure telemetry with data about running AI workloads. The company positions it as a tool for improving the availability and productivity of its AI factories.
Minerva collects signals from GPUs, networking equipment, storage, schedulers, and workloads. The platform then builds a dynamic graph of connections between components and jobs. According to NVIDIA, this context helps trace how a hardware anomaly affects a specific training task and group related events into a single incident with a probable root cause. The platform analyzes billions of telemetry data points per hour—a system capability claimed by NVIDIA.
What changes in incident investigation
A common monitoring problem in large clusters is that compute, networking, and operations are viewed separately. Minerva aims to connect these layers: a network interface anomaly can be considered alongside GPU status and the performance of the job using that route. The practical goal is to shorten the path from a collection of signals to locating a bottleneck.
The platform includes an AI assistant for natural-language questions and connects to external agent tools. Operators can investigate cluster status using live telemetry and receive recommendations for resolving issues. NVIDIA describes a workflow in which an employee applies a fix manually, as well as the option to connect automation through infrastructure APIs and external systems.
For operations teams, this separates diagnosis from action. Correlating signals can help identify where to look for the cause more quickly, while automatically changing infrastructure remains a separate team decision. During deployment, it is useful to define in advance which recommendations may be applied automatically, which require approval, and where the results are recorded. This is a practical conclusion drawn from NVIDIA’s described distinction between recommendations, manual fixes, and connected automation.
Who can access the platform
NVIDIA said that IREN is already using Minerva; eight other organizations, including IBM, Lambda, and Tesla, plan to deploy it. Dell Technologies, Lenovo, and Supermicro, according to the company, will offer Minerva as part of their AI factory solution portfolios. The announcement lists current and planned deployments, but provides no independent comparative measurements of the impact on operators.
NVIDIA’s product page lists deployment options ranging from collecting telemetry and sending data to external tools such as Grafana to deploying a full AIOps platform. The documentation also describes collecting signals from different components and integrating with operations tools. Before evaluating the product, teams should compare the supported-source list with their own infrastructure, check telemetry coverage, and record baseline metrics: cluster availability, GPU-hours lost, and time to diagnose and recover.
Minerva is aimed at organizations that need to investigate incidents spanning compute, networking, and AI tasks. Its practical purpose is to bring these signals together in a shared context; its value for a particular cluster can be assessed by how well it identifies the causes of real failures and reduces recovery time.