On October 22, 2026, NVIDIA launched Minerva, described as "an AIOps platform built to help teams find problems earlier and keep AI factories productive." It is aimed at operators of large GPU clusters, not at model developers, and its framing is worth noting: the target metric is AI factory yield, "a measure of the useful work infrastructure produces through GPU hours, goodput, cluster uptime and token production."
What it does
NVIDIA's description has four parts.
- One telemetry plane. Signals from GPUs, networking, storage, schedulers and workloads are joined, rather than kept in separate dashboards for the network, compute, SRE and ML teams.
- A topology graph. Minerva "builds a dynamic knowledge graph that logically maps the infrastructure — GPUs, SuperNICs, switches and cables — and the jobs running on it", so that a slow training job can be traced to the component causing it.
- Learned baselines. It "analyzes billions of data points per hour using supervised and unsupervised machine learning", learns normal behaviour, predicts emerging failures and groups "an event storm into a focused incident with a likely root cause."
- An agentic front end. A built-in assistant answers questions about cluster health from live telemetry and recommends fixes. Teams can keep a person in the loop or connect Minerva to automation through infrastructure APIs.
The examples NVIDIA gives are gray failures rather than outages: "a GPU running hotter than its peers or an NVMe drive losing performance under load." Those faults leave every node up while slowing every synchronized step of a training job.
Who says they will use it
IREN is named as a current user. Its CTO, Denis Skrinnikoff, says Minerva gives his 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, and Dell, Lenovo and Supermicro will offer it in their AI factory operations portfolios. Jio's inclusion is the India angle: as Indian operators add GPU capacity, the tooling that turns it into useful work becomes a purchasing decision too.
The missing numbers
The launch lists three benefits: longer time between interruptions, faster repair and fewer alerts that matter. It gives no figure for any of them, no pricing, no deployment model and no detail on which non-NVIDIA components it can read. This contrasts with Meta's October 19 Prometheus post, covered here on October 21, which published interruption rates, mean time between interruptions and spare-capacity ratios for its own NVL72 fleet. Those are the kinds of figures any yield claim should be measured against.
What remains uncertain. It is not yet clear whether Minerva works outside NVIDIA networking, how its graph handles multi-vendor storage and schedulers, how its remediation actions are permissioned, or whether the planned adopters have committed to production use.
Analysis: Defining yield as useful work per GPU-hour is the right move. It makes operations a finance conversation, and it exposes failures that utilization dashboards hide. Before buying, operators should ask for before-and-after MTBI and mean-time-to-repair on a comparable fleet. They should also decide which remediations an assistant may run without a human, for example draining a node, and which it may only recommend, for example reflashing firmware. An assistant that can act on infrastructure APIs is an agent with production credentials and needs the same controls as any other.
NVIDIA's Minerva (October 22) is an AIOps platform that joins GPU, network, storage, scheduler and job telemetry into a topology graph, learns baselines to predict failures, groups alert storms into incidents, and adds an assistant that can trigger fixes; IREN uses it and Jio, IBM, Lambda, Tesla and others plan to, but NVIDIA has published no MTBI, repair-time or pricing figures.