NVIDIA (Nasdaq: NVDA) has introduced Minerva, an AIOps platform intended to help operators find infrastructure problems earlier and protect the useful output of large AI factories. The platform combines telemetry from AI workloads and infrastructure components, then applies purpose-built AI models to identify anomalies, correlate signals and recommend a response.
The product arrives as enterprise AI infrastructure grows more interdependent. A large training or inference environment can combine GPUs, high-speed networking, storage, schedulers and orchestration tools across a cluster. A degraded GPU, an overheating component or a storage device that slows under load can affect work far beyond the component itself, while the initial warning may be buried in a large volume of operational data.
NVIDIA frames the operational objective as AI-factory yield: the useful work an environment produces through GPU hours, goodput, cluster uptime and token production. That measurement is more specific than traditional server availability because a cluster can technically remain online while failing to deliver expected model-training or inference output. For IT leaders funding AI capacity, the distinction affects both the cost of idle compute and the reliability of services that depend on it.
Minerva builds a dynamic knowledge graph that maps GPUs, SuperNICs, switches, cables and the jobs running on the infrastructure. NVIDIA said that context helps operators trace a slow training job back to a contributing component, including in factories with hundreds of thousands of GPUs. The system analyzes billions of data points per hour, learns normal behavior, detects anomalies and correlates related events using the environment’s topology.
The goal is to avoid treating every symptom as an isolated alert. NVIDIA said Minerva can group an event storm into a focused incident with a likely root cause. This is a familiar AIOps ambition, but the AI-factory context raises the stakes: an issue that reduces network performance, accelerator health or storage throughput can translate directly into lower job completion rates or a longer time to deliver AI capacity to an internal business unit.
The platform includes a native AI assistant and connections to external agentic AI tools, giving operators a natural-language interface grounded in live telemetry. Teams can investigate cluster health and receive recommended remediation through that interface. NVIDIA said customers can keep a person in the loop to carry out changes manually, or connect the system to automation through infrastructure APIs and external systems.
That choice is important. Recommendations derived from telemetry may speed triage, but infrastructure teams still need change-management controls, escalation paths and a clear policy for which actions may run automatically. Observability software can narrow the investigation, yet it does not remove the responsibility to validate a diagnosis before a high-impact change.
NVIDIA said several organizations plan to deploy Minerva and that multiple systems vendors will include it in AI-factory operations portfolios. The development shows how AI infrastructure is shifting from a hardware procurement exercise to a full operational discipline. Compared with legacy data-center monitoring, the competitive advantage increasingly lies in linking physical components, workload behavior and remediation workflows quickly enough to preserve expensive AI capacity.