NVIDIA Minerva launched on October 22 at GTC Berlin as an AIOps platform that correlates telemetry across GPUs, networking, storage, schedulers and AI workloads. The goal is to help operators identify degraded components before they reduce the output of large training jobs, with the system analyzing billions of data points each hour and building a dynamic knowledge graph of factory infrastructure.
AI data centers operate at a scale where a minor fault can become expensive quickly. A GPU running hotter than its peers, a storage drive losing throughput under load or a network interface behaving inconsistently can interrupt a distributed workload. The impact may extend well beyond the affected machine if operators need time to determine which component deserves attention first.
Minerva is designed for that complexity. Rather than treating infrastructure monitoring, workload scheduling and application performance as separate disciplines, the platform links their signals within one operational view. IREN is already using it, while Dell Technologies, Lenovo and Supermicro plan to include it in their AI factory operations portfolios.
Why AI Factory Observability Needs a Different Approach
Large AI clusters differ from conventional data centers in several important ways. They combine tightly connected accelerators, high-speed interconnects, specialized storage and parallel applications that may consist of thousands of coordinated tasks. A fault can appear as a slowdown rather than a complete outage, making it harder to distinguish an underlying hardware problem from normal variation in workload demand.
Traditional monitoring often reinforces those blind spots. Networking engineers examine one segment of the operation, site reliability teams watch another, and machine learning specialists focus on job metrics without seeing every physical signal. Each team may have a valid perspective while still missing the connection between a marginal cable, an abnormal temperature and a job that is taking longer than expected.
That fragmentation creates an additional cost through alert volume. When one component generates symptoms across multiple systems, engineers can spend their time sorting warnings instead of resolving the cause. NVIDIA positions Minerva as a way to consolidate those signals and present a smaller set of relevant incidents, with an indication of the likely source.
The company measures the operational objective as AI factory yield. That concept covers useful output from infrastructure, including GPU hours, goodput, cluster uptime and token production. It reflects whether expensive compute capacity is being used effectively, rather than merely whether equipment remains powered on.
How NVIDIA Minerva Connects Infrastructure and Workloads
At the center of NVIDIA Minerva is a dynamic knowledge graph that maps both hardware and the jobs running on it. The graph can represent GPUs, SuperNICs, BlueField DPUs, switches, cables, storage systems, schedulers and AI workloads as connected elements. Relationships between those elements give the platform the context required to connect a physical anomaly to the specific workload affected by it.
That context matters in environments containing hundreds of thousands of GPUs. A temperature reading or network error rarely carries enough information on its own to explain a declining result. Linking the signal to topology and workload state can turn a broad performance issue into a focused maintenance target. The platform can therefore trace a slow training job toward a component that may be responsible.
Minerva also uses supervised and unsupervised machine learning to examine billions of data points per hour. The models learn normal behavior for a given configuration, identify anomalies that may not resemble known faults and predict emerging failures. Because these analytical methods account for the factory’s topology, they can group related events instead of sending every symptom to an operator as an isolated warning.
The result is a change in the operating workflow. Instead of responding to an alert storm after a job has already lost time, teams can investigate an incident while it is developing and schedule maintenance earlier. The knowledge graph also gives support personnel a shared reference point when responsibility is divided among networking, compute, storage and model teams.
Natural-Language Investigation and Agentic Automation
Minerva includes a native AI assistant, identified in supporting material as Sherlock, and can also connect with external agentic AI tools. The interface allows operators to ask questions about cluster health in ordinary language instead of writing queries for each monitoring system. Answers are grounded in live telemetry and the relationships represented in the knowledge graph.
That approach does not necessarily remove the operator from the process. A team can review a recommendation, have a person perform the repair or connect Minerva to automation through infrastructure application programming interfaces and external systems. This balance is relevant in production environments where an incorrect automated action could affect many jobs at once.
The assistant also supports an MCP server, which helps it work with agentic tools in a more structured way. In practical terms, the technology can help narrow an investigation, identify relevant evidence and suggest a next step. It does not replace the need for operational judgment, but it can reduce the time required to assemble evidence from many consoles.
Adoption Signals and the Competitive Monitoring Market
Named organizations connected to the launch include IREN, Aolani, Bitdeer AI, IBM, Jio, Lambda, Quake AI and Tesla. Their involvement indicates interest from organizations operating or building substantial GPU infrastructure. Dell Technologies, Lenovo and Supermicro’s plans to offer Minerva within AI factory operations portfolios could also broaden access beyond operators directly involved in the announcement.
For infrastructure vendors and data center operators, the launch raises questions about where monitoring software fits within broader stack strategies. A unified AIOps layer could become a control point for GPU fleets, while competing observability products may emphasize portability, specialized analysis or lower cost. The practical answer will depend on deployment results, integration requirements and how much operational time each approach saves.
Some early evidence comes from NVIDIA’s internal testing of Minerva, which reportedly produced longer intervals between interruptions, shorter repair times and more actionable insights through noise reduction. Such vendor-reported results need continued evaluation in varied customer environments, where workloads and hardware configurations differ considerably. That is the bet behind NVIDIA Minerva, and incident-response times will show whether operators trust it to act on its own.
IREN chief technology officer Denis Skrinnikoff said the unified view across networks, GPUs and jobs allows his operations team to focus more on proactive maintenance rather than troubleshooting. That statement captures the central value proposition of NVIDIA Minerva: connecting a changing collection of signals to the output that matters, then giving operators enough context to act before a small fault becomes a large productivity loss.