NVIDIA's Minerva measures AI factory yield, not GPU uptime. Private AI clusters need the same metric
NVIDIA launched Minerva, an AIOps platform that maps GPUs, network, storage and jobs into one knowledge graph to find what slows AI workloads. The idea to keep: measure useful work, not utilization.
Listen to this article · 5 min
AI-generated narration of the full article.
A training job that should finish in two days takes three. Every dashboard is green: GPUs are busy, the network is up, storage responds. Somewhere, one GPU runs hotter than its peers and throttles, or one drive slows under load, and because distributed training moves at the speed of its slowest participant, the whole job pays for it.
That is the problem NVIDIA’s Minerva, launched on October 22, is built for. It is an AIOps platform that collects telemetry from GPUs, networking, storage, schedulers and the workloads themselves, and applies purpose-built models to find problems earlier.
One graph for the whole cluster
Minerva’s central idea is a dynamic knowledge graph that maps the infrastructure, GPUs, network cards, switches and cables, together with the jobs running on it. With that map, an operator can trace a slow training run to the component affecting it, NVIDIA says, even in a facility with hundreds of thousands of GPUs.
- Telemetry from GPUs, network, storage, schedulers, jobs
- Knowledge graph of components and workloads
- Models detect anomalies and predict failures
- Related alerts grouped into one incident
- Recommended fix, by assistant or API
- Person applies it, or automation does
The operator choosesThe new layer
It analyzes what NVIDIA describes as billions of data points per hour with supervised and unsupervised models, learns normal behavior, predicts emerging failures and groups an event storm into one incident with a likely root cause. A built-in assistant answers questions about cluster health in natural language, and Minerva can connect to external agentic tools. Teams can keep a person applying fixes or wire recommendations into automation through infrastructure APIs.
IREN, an AI cloud operator, is using it. NVIDIA lists Aolani, Bitdeer AI, IBM, Jio, Lambda, Quake AI and Tesla as planning deployments, and Dell, Lenovo and Supermicro as offering it with their AI factory portfolios. Those are plans, and NVIDIA’s claims of longer time between interruptions and faster repair are not quantified in the announcement.
Yield, not utilization
The idea worth taking from Minerva even if you never buy it is the metric NVIDIA puts at the center: AI factory yield, the useful work an infrastructure produces, measured through GPU hours, goodput, cluster uptime and token production.
Most enterprise reporting on AI infrastructure still measures utilization: how busy the GPUs were. Utilization is a weak signal. A GPU stalled on a slow neighbor is busy. A job that crashes at hour 40 and restarts from a checkpoint was busy for all 40 hours. A cluster can report high utilization and deliver half the completed work it should.
We have argued that inference should be priced per completed task over the hours you pay for, and that the next cost reductions may come from below the model, in scheduling and serving. Yield is the infrastructure version of the same discipline: count what finished, not what ran.
Does this matter outside the hyperscalers?
More than it seems. Enterprises buying private AI infrastructure, whether racks sized by workload or clusters for sovereign or regulated workloads, are now running small AI factories. They rarely have the specialized operations teams that large GPU clouds employ. A platform that correlates GPU, network and job telemetry is exactly the expertise they lack, and the fact that Dell, Lenovo and Supermicro plan to offer Minerva means it may arrive bundled with the hardware.
Two cautions for those buyers. First, check what Minerva needs to see: its value depends on telemetry from NVIDIA GPUs and networking, and its view of third-party storage or schedulers should be confirmed for your environment. Second, decide early what it may change on its own. A recommendation to drain a node is harmless; an automated drain during a regulated batch job is an operational change, and it should go through the same approval as any other.
Questions for your next infrastructure review
- What fraction of the GPU hours you paid for last month produced completed work?
- How long does it take, today, to trace a slow job to the component causing it?
- Which remediation steps may automation take without a person, and who approved that list?
- If you run inference and training on the same cluster, which one yields when they compete, and is that a written rule?
Minerva is a product for operators of very large clusters. The metric behind it belongs in every enterprise that owns GPUs.