Nvidia has introduced Minerva, an operations platform designed to find failing or degraded components inside large AI clusters before they waste substantial GPU time. It connects infrastructure telemetry with the training and inference jobs using that hardware.
The platform builds a knowledge graph of GPUs, network adapters, switches, cables, storage, schedulers and workloads. Machine-learning models analyze billions of data points per hour, according to Nvidia, looking for abnormal behavior and related alerts. That context can connect a slow job to a hot GPU or weakening storage device instead of leaving separate teams to reconcile independent dashboards.
A built-in assistant lets operators investigate cluster health in natural language and receive proposed remediation. Teams can keep execution manual or connect Minerva to infrastructure APIs for automation. Nvidia frames the main measure as AI-factory yield: useful output expressed through GPU hours, cluster uptime, completed work and token production.
IREN is already using the system, while IBM, Jio, Lambda, Tesla and several infrastructure operators plan deployments. Dell, Lenovo and Supermicro will offer it in their AI-factory operations portfolios. Nvidia did not publish pricing or independent measurements of failure prediction and repair-time improvements in the announcement.