NVIDIA unveils Minerva: end-to-end observability for AI factories that compresses an alert storm into one root cause

Source first. This is a product announcement on NVIDIA's official blog dated October 22, 2026, by Brian Sparks, titled NVIDIA Unveils Minerva to Bring End-to-End Observability for AI Factories, filed under GTC Berlin 2026. The post says the product launched today, and the product page is at nvidia.com/networking/minerva. Everything below about capabilities comes from that announcement and is vendor description; this article separates what is an existing customer statement from what is only a plan.

The pain point is stated plainly: AI factories are among the most complex operations in the world, with vast numbers of GPUs running across high-speed networks to train models and serve applications; at that scale a single degraded component can slow an entire job. And the warning signs are easy to miss — a GPU running hotter than its peers, or an NVMe drive losing performance under load. What makes these hard is that nothing errors out; things just get slower.

Minerva's first move is stitching split views back together. Traditional monitoring divides the data centre by team: networking, compute, site reliability and machine learning each see a different slice. Minerva connects signals across NVIDIA GPUs, networking, storage, schedulers and workloads, building a dynamic knowledge graph that logically maps GPUs, SuperNICs, switches and cables along with the jobs running on them. NVIDIA says this lets operators trace a slow training job to the component affecting it, even across a factory with hundreds of thousands of GPUs.

The second move is noise reduction. The platform analyses billions of data points per hour using both supervised and unsupervised machine learning: it learns what normal looks like, detects anomalies it has not seen before, predicts emerging failures, and correlates related events according to the factory's topology. The design choice that matters is that it does not forward every symptom to an operator; it groups an event storm into one focused incident with a likely root cause. Anyone who has been on call knows how large that difference is.

The interaction layer is addressed too. Minerva ships with a native AI assistant and connects to external agentic AI tools, offering a natural-language interface grounded in live telemetry: operators can ask about cluster health and get recommended remediation in conversation. There are two paths to acting on it: keep a human in the loop to implement fixes manually, or connect Minerva to automation through infrastructure APIs and external systems. That distinction matters, because the first is what most production environments will actually do.

NVIDIA lists exactly three benefits, all about operational tempo: longer intervals between interruptions because trouble is spotted before it lands, faster time to repair because the problem is pinpointed, and greater insight because only the alerts that matter survive the filter. It also names the metric it optimises, AI factory yield, defined through GPU hours, goodput, cluster uptime and token production. When buying, those four are the numbers to hold the system to — not the look of the dashboard.

Separate the customer states. IREN is using it, according to CTO Denis Skrinnikoff, who said that as AI factories scale, the key to serving customers well is maximising GPU productivity, and that Minerva gives the operations team one view across the network, the GPUs and the jobs running on them, shifting focus from troubleshooting to proactive maintenance. Those planning to deploy are Aolani, Bitdeer AI, IBM, Jio, Lambda, Quake AI and Tesla — the wording is plan to deploy, not live. On hardware, Dell, Lenovo and Supermicro will offer Minerva within their AI factory operations portfolios; the wording is will offer.

Boundaries worth stating. First, this is a vendor announcement with no third-party benchmark, no published false-positive or false-negative rates, and no detail on which cluster sizes were used to validate behaviour at the hundreds-of-thousands-of-GPU scale; this article does not fill those numbers in. Second, it is a stack-first design — GPUs, SuperNICs and switches all come from one vendor — and the announcement says nothing about coverage in heterogeneous environments with other vendors' accelerators or white-box networking. Third, connecting external agentic tools means putting detection and remediation on the same chain, which runs straight into the agent-permission questions the industry has been arguing about for weeks: decide which actions may be automated before wiring it up. Fourth, no pricing was disclosed.