Nvidia’s Minerva wants to turn an AI factory’s alert flood into one incident

Nvidia has launched Minerva, an AIOps platform that pulls telemetry from GPUs, networking, storage, schedulers and workloads into one knowledge graph, so operators can trace a slow job to the failing component. IREN is already using it.

Nvidia announced the platform on October 22, 2026, and describes its purpose as finding problems earlier so AI factories stay productive. The company frames the goal around what it calls AI factory yield: the useful work infrastructure produces, measured through GPU hours, goodput, cluster uptime and token production.

One graph from the cable to the training job

According to Nvidia, traditional monitoring splits a data center into separate views, with networking, compute, site reliability and machine learning teams each seeing only their slice. Minerva instead builds a dynamic knowledge graph of GPUs, SuperNICs, switches and cables, plus the jobs running on them. Nvidia says this lets an operator follow a degraded training job back to the responsible part, even in a factory with hundreds of thousands of GPUs.

The platform, Nvidia says, analyzes billions of data points per hour with supervised and unsupervised machine learning. It learns what normal looks like, flags anomalies it has not seen before, predicts failures and correlates events using the factory’s topology. The examples Nvidia gives are mundane: a GPU running hotter than its neighbors, or an NVMe drive losing performance under load. The output is meant to be one incident with a likely root cause rather than a stream of separate alerts.

Minerva also ships with a native AI assistant and can connect to external agentic AI tools. Operators can ask about cluster health in natural language, answered from live telemetry, and receive suggested fixes. Nvidia says teams can keep a person in the loop to apply changes by hand, or wire Minerva into automation through infrastructure APIs and external systems.

Nvidia names three outcomes: longer time between interruptions, faster time to repair, and fewer but more meaningful alerts.

Cloud providers lead the customer list

IREN is the only named company already running Minerva. Its chief technology officer, Denis Skrinnikoff, said:

NVIDIA Minerva gives our operations team one view across the network, the GPUs and the jobs running on them

Aolani, Bitdeer AI, IBM, Jio, Lambda, Quake AI and Tesla plan to deploy it. On the hardware side, Dell Technologies, Lenovo and Supermicro will include Minerva in their AI factory operations offerings.

The announcement leaves several practical questions open. It gives no pricing, no licensing model, no date for general availability and no list of supported hardware generations. It does not say whether Minerva works on clusters with non-Nvidia components, or how it is deployed. The reduction in repair time is described as significant, but no figures are given, and none of the performance claims have been independently verified.

Nvidia folds a patchwork of tools into one product

Observability for Nvidia clusters has so far been assembled from parts. In an August 2026 developer post on choosing full-stack observability, Nvidia listed eight separate tools covering different layers: DCGM for GPU telemetry, NVSM for system health, UFM for InfiniBand fabrics, NetQ for Ethernet, NMX for rack-scale NVLink, Base Command Manager for cluster aggregation, Run:ai for scheduling and NIM for inference services. The same post described pairing them with Prometheus and Grafana, and in the table Nvidia’s own coverage ratings show only partial support for some layers. Minerva reads as the first attempt to put correlation, prediction and a conversational layer on top of that stack under one name.

It also moves Nvidia closer to the partners that sell into this space. In August 2025, Virtana announced a partnership with Nvidia for AI factory observability, offering automated topology discovery, AI root cause analysis, predictive performance management and a generative AI assistant for plain-language queries. Those are close to the headline features of Minerva. Nvidia’s announcement does not mention Virtana or any other observability vendor, and does not say how Minerva will coexist with such products. Operators who already run third-party monitoring will have to judge for themselves whether a first-party platform that sees Nvidia’s networking and GPU telemetry natively is worth adding or swapping in.

The timing fits the scale Nvidia’s partners are now building to. Our recent report on Google Cloud’s Vera Rubin A5X instances shows how quickly new capacity is arriving. The more GPUs in a cluster, the more a single weak component costs, which is the argument Nvidia makes for Minerva: at that size, one degraded part can slow a whole job.

What readers can do with it today

For most readers, nothing yet. Minerva is aimed at operators of large GPU clusters: AI cloud providers, telecoms and enterprises running their own factories. Access appears to run through Nvidia and its server partners, since Dell, Lenovo and Supermicro will offer it, but the announcement gives no order process or timeline. Those interested can start with the product page Nvidia links from its blog post.

What to watch is whether the early adopters publish measured results. Fewer interruptions and quicker repairs are straightforward to count, and the stated customers run enough GPUs to produce real numbers. Until they do, the claims remain Nvidia’s own.