NVIDIA Minerva Watches Billions of Telemetry Points an Hour So AI Factories Stop Losing GPU Time

NVIDIA introduced Minerva on Thursday, an AI-native AIOps platform built to give operations teams earlier warning of trouble inside the GPU-dense facilities the company calls AI factories. The pitch is straightforward: at a scale where hundreds of thousands of accelerators train and serve models across high-speed networks, one degraded component can quietly tax an entire job, and the warning signs — an accelerator running hotter than its peers, an NVMe drive losing performance under load — are easy to miss in a wall of dashboards.

Minerva is organized around what NVIDIA calls AI factory yield: the useful work the infrastructure actually produces, measured in GPU hours, goodput, cluster uptime and token production. Rather than showing networking, compute and machine learning teams three different pictures of the same data center, the platform pulls telemetry from NVIDIA GPUs, Spectrum-X Ethernet, NVLink switches, SuperNICs, DPUs, storage, schedulers such as Slurm and Run:ai, and the AI workloads themselves into a single correlated view.

From those signals it builds a dynamic knowledge graph that logically maps the physical plant — accelerators, network cards, switches, cables — together with the jobs running on top of it. NVIDIA says that context lets an operator trace a slow training run back to the specific component causing it, even inside a factory with hundreds of thousands of GPUs, which is precisely the kind of diagnosis that previously consumed days of cross-team archaeology.

The analytics layer is where the AI comes in. NVIDIA says Minerva ingests and analyzes billions of data points per hour using a mix of supervised and unsupervised machine learning: it learns what normal behavior looks like, flags anomalies it has never seen before, predicts failures as they emerge, and correlates related events based on factory topology. Instead of forwarding every symptom, it collapses an event storm into one focused incident with a likely root cause attached.

The platform also leans into the agent era it is meant to support. It ships with Sherlock, a native multi-agent assistant that lets operators interrogate cluster health in natural language grounded in live telemetry, and exposes an MCP server so customer-side agents can plug in. Teams can keep a person in the loop to approve fixes manually, or wire Minerva into automation through infrastructure APIs. It deploys on premises or as SaaS, scales to hundreds of thousands of GPUs, and integrates with the DSX OS stack through a Minerva Collector component that NVIDIA says provides one view from GPU to grid.

Adoption is starting with the cloud operators who feel the pain most. IREN is already using the platform; its chief technology officer Denis Skrinnikoff said the single view across network, GPUs and jobs shifts the team from troubleshooting toward proactive maintenance. Aolani, Bitdeer AI, IBM, Jio, Lambda, Quake AI and Tesla all plan to deploy Minerva, and Dell Technologies, Lenovo and Supermicro will bundle it into their own AI factory operations portfolios.

The launch lands two days after NVIDIA used GTC Berlin to introduce DSX POD, a containerized reference design that cuts AI data center deployment from roughly six months to as little as 14 days. The pairing is deliberate: hardware gets installed faster, and software now exists to keep it running. NVIDIA claims longer time between interruptions and faster repair times, but those are company statements and have not been independently verified.

The economics explain the urgency. Jensen Huang put a gigawatt-scale AI data center at $50 billion, with expected revenue in the tens of billions per year — his arithmetic at GTC Berlin was that a three-month delay burns $10 billion of revenue. Against that denominator, even a single-digit percentage of recovered GPU uptime justifies an entire observability product line. The catch for buyers is that the platform watching the factory is made by the same company that built it, one more layer of a stack that is already very difficult to leave.