As organizations expand AI training and inference, they must manage far more than GPUs. Networks, storage systems, schedulers, software workloads and physical infrastructure all affect whether an AI job runs smoothly. A failure or slowdown in one part of the system can waste expensive computing time and delay results. NVIDIA Minerva aims to address this operational challenge by connecting infrastructure and workload data in a single AIOps platform.
NVIDIA announced Minerva on October 22, 2026, describing it as a system designed to help teams identify issues earlier and improve AI factory productivity. Its central idea is straightforward: give operators a connected view of the environment, use machine learning to identify meaningful patterns, and present those findings as actionable incidents rather than a flood of separate alerts. NVIDIA’s announcement explains the platform’s goals and its approach to observability across AI factories.
Why AI Factories Need a Broader View
An AI factory combines computing resources and software to train models or deliver AI services. At large scale, its components depend on one another. A training job may use many GPUs, move data across high-speed networks and rely on storage systems to supply information fast enough. A problem in any one of those areas can affect the overall job.
The challenge is not always a complete failure. A GPU may perform differently from its peers, a storage device may slow under pressure, or a network component may introduce a bottleneck. Each symptom can look small in isolation. Yet when these signals appear during a large, distributed workload, teams need to understand whether they share a cause and which jobs they affect.
Traditional monitoring can leave operations teams with fragmented views. One team may focus on networking, another on compute, while site reliability and machine-learning teams use separate tools. When data stays divided, it becomes harder to connect a change in hardware behavior to an impact on an active workload. Minerva’s stated purpose is to bring these signals together so operators can investigate the whole system rather than one component at a time.
How NVIDIA Minerva Connects Telemetry
Minerva gathers telemetry from AI infrastructure and workloads, including NVIDIA GPUs, networking, storage, schedulers and the jobs running across them. Telemetry is the operational information systems produce about their condition and activity. When teams can examine these signals together, they can look for relationships between equipment performance and workload behavior.
The platform uses this information to create a dynamic knowledge graph. In simple terms, the graph maps how infrastructure components relate to one another and to the jobs that use them. A GPU, switch, cable or storage resource does not appear as an isolated item; the graph provides context about its place in the environment.
That context can help operators trace a slow job beyond the first visible symptom. For example, a workload delay may appear in a dashboard as a compute issue, while the underlying cause may involve a connected component elsewhere in the system. A relationship map gives teams a way to follow those dependencies and narrow their investigation.
This approach supports GPU cluster observability by connecting workload-level events with infrastructure-level signals. Instead of asking only, “Which device is generating an alert?” an operator can also ask, “What is this device connected to, which jobs rely on it, and what else changed at the same time?”
Using Machine Learning to Find Meaningful Signals
According to NVIDIA, Minerva analyzes large volumes of operational data with supervised and unsupervised machine learning. These methods serve different purposes. Supervised models learn from examples that have been labeled, while unsupervised methods can identify patterns without relying on predefined labels for every possible event.
The platform is designed to learn typical operating behavior, detect anomalies, anticipate emerging problems and correlate related events using the factory’s topology. That combination matters because complex infrastructure often produces many symptoms at once. A burst of alerts may reflect one underlying issue rather than dozens of independent failures.
Minerva aims to group related signals into a more focused incident and point operators toward a likely root cause. This can help teams avoid spending time sorting through repetitive notifications before they begin diagnosis. However, an alert correlation or likely cause should be treated as a guide for investigation, not as proof that the system has identified the cause in every case. Operators still need to validate findings against their environment and operational procedures.
This is where an AIOps platform can be useful: it applies analytical techniques to operations data so teams can prioritize potential problems and respond with more context. In the case of Minerva, the goal is to turn infrastructure and workload telemetry into a more coherent operational picture.
A Conversational Assistant and Human Oversight
NVIDIA says Minerva includes a built-in AI assistant and can connect with external agentic AI tools. The assistant provides a natural-language way to ask questions about live telemetry. Instead of navigating multiple dashboards for every investigation, an operator may ask about cluster health or request help understanding an incident.
The platform can also recommend remediation steps. Teams can review those suggestions and carry them out themselves, keeping a person involved in the decision. Alternatively, organizations can connect Minerva to automation through infrastructure APIs and external systems. That option may support faster responses, but it also makes governance important: teams should define which actions can run automatically, which require approval and how they will review the results.
A conversational interface can make operational information easier to access, but it does not remove the need for good data, clear permissions or skilled operators. A useful recommendation depends on the quality and context of the telemetry available to the platform. Human review remains essential, especially when a proposed action could affect a production workload or a large group of connected systems.
Measuring AI Factory Yield
Minerva focuses on AI factory yield: the useful work that infrastructure produces. NVIDIA describes relevant measures such as GPU hours, goodput, cluster uptime and token production. These metrics provide a way to look beyond whether machines are powered on and ask whether the infrastructure is delivering productive work.
For example, uptime can show whether systems remain available, while goodput can help teams consider how much useful work a workload completes. Token production can be relevant to services that generate AI responses. The right measurements depend on the workload and the organization’s goals, so operators need to select metrics that reflect the outcomes they actually care about.
The platform is intended to improve several operational outcomes: longer periods between interruptions, faster repair when problems occur, and clearer insight into the alerts that matter. Those outcomes describe the intended benefits, not a guarantee that every deployment will achieve the same results. Teams should assess performance against their own baseline after implementation.
A practical predictive maintenance strategy also depends on follow-through. Detecting a potential issue earlier can help teams schedule investigation or maintenance, but organizations still need procedures for checking equipment, prioritizing risks and recording what they learn. Observability is most valuable when it supports an operational process, not when it simply adds another dashboard.
Who Is Planning to Deploy Minerva?
NVIDIA’s announcement names IREN as a partner using Minerva and says Aolani, Bitdeer AI, IBM, Jio, Lambda, Quake AI and Tesla plan to deploy it. It also says Dell Technologies, Lenovo and Supermicro will offer Minerva as part of their AI factory operations portfolios.
This list signals interest from organizations involved in operating or supplying AI infrastructure. It does not, by itself, establish that every named organization has completed a deployment or achieved a specific performance improvement. Buyers should distinguish announced plans from generally available deployments and ask vendors for details about compatibility, rollout, support and measurable outcomes.
For teams evaluating AI infrastructure monitoring, useful questions include: Which data sources can the platform access? How does it represent dependencies between hardware and jobs? Can operators control its permissions? What does integration with existing observability and automation systems require? How will the organization measure improvement in uptime, incident response or workload productivity?
What Minerva Could Mean for AI Operations
As AI workloads grow, infrastructure operations become more interconnected. A monitoring approach that treats compute, networking, storage and workloads as separate domains may make it difficult to diagnose issues quickly. NVIDIA Minerva’s proposed approach combines these perspectives and uses machine learning to identify patterns across them.
For operations teams, the potential value lies in reducing the distance between a symptom and a useful explanation. A connected view could help teams understand which resources are involved, which workloads are affected and whether multiple alerts share a common origin. A conversational assistant could make that information easier to query, while human review or carefully governed automation could support remediation.
The platform’s real value will depend on implementation details, data quality, integration effort and the accuracy of its operational insights in each environment. Organizations should evaluate it against their existing tools and define success with measurable indicators. The key question is not simply how many alerts a system can process, but whether it helps people protect productive work and resolve issues with better context.
NVIDIA positions Minerva as an AIOps platform for improving end-to-end observability in AI factories. Its combination of telemetry correlation, a topology-aware knowledge graph, machine-learning analysis and operator-facing assistance addresses a real challenge in large AI environments: understanding how infrastructure health affects workload performance. For teams scaling AI operations, that connected perspective may help make complex systems easier to monitor and manage.
