NVIDIA Unveils Minerva to Bring End-to-End Observability for AI Factories

NVIDIA launched Minerva, an AIOps platform that pulls together telemetry from GPUs, networking, storage, schedulers and workloads, then uses machine learning to spot anomalies, predict failures and group alert storms into single incidents with a likely root cause. It comes with a built-in AI assistant and can connect to outside agentic tools and automation.

At the scale of hundreds of thousands of GPUs, one overheating chip or a slowing NVMe drive can drag down a whole training job. Minerva is aimed at what NVIDIA calls AI factory yield: GPU hours, goodput, uptime and tokens produced.

This is a sensible product for a real problem. When a training run spans enormous clusters, idle or degraded hardware is expensive and the signals are spread across teams that rarely share a view. A knowledge graph that maps jobs to the specific GPUs, SuperNICs, switches and cables they run on, so a slow job can be traced to the component causing it, is the kind of tooling operators have wanted. Fewer alerts, more meaningful ones, is a good goal.

We'd note what is and isn't in evidence. The launch claims longer time between interruptions, faster repair and less alert noise, but the announcement doesn't publish figures for any of those. IREN's CTO is quoted describing a shift toward proactive maintenance, and Aolani, Bitdeer AI, IBM, Jio, Lambda, Quake AI and Tesla plan to deploy it, with Dell, Lenovo and Supermicro offering it. That is strong adoption intent, not yet measured results.

The strategic read is the bigger one. NVIDIA already sells the chips, the networking and much of the software stack. Owning the operations layer as well tightens its grip on how AI data centers are run, and makes it harder for operators to mix in other vendors' hardware without losing visibility. That's good for integration, less good for competition.

The agentic piece also deserves care. Letting an AI assistant recommend fixes is useful. Wiring it to infrastructure APIs to apply them automatically can speed recovery, or spread a bad change quickly across a cluster. NVIDIA says teams can keep a person in the loop, and I'd expect most will, at least at first.

Minerva was announced alongside NVIDIA's European announcements around GTC Berlin. Traditional monitoring splits the data center into separate views for networking, compute, site reliability and machine learning teams; Minerva's pitch is a single correlated view using supervised and unsupervised machine learning over billions of data points per hour.