NVIDIA Unveils Minerva to Extend AI Factory Observability Into Full-Stack AIOps

What happened

NVIDIA introduced Minerva as an end-to-end AIOps platform for AI factories, combining infrastructure and AI-workload telemetry. It can collect telemetry from GPUs, Spectrum-X Ethernet, NVLink switches, SuperNICs, DPUs, schedulers and AI workloads, and provide anomaly detection, root-cause analysis and closed-loop remediation.

Minerva uses a dynamic knowledge graph to connect infrastructure with jobs, helping operators trace a hardware anomaly to the affected training job. NVIDIA said Minerva analyzes billions of data points per hour and can learn normal behavior, predict anomalies and correlate events.

Who is affected

NVIDIA said IREN is already using Minerva; Aolani, Bitdeer AI, IBM, Jio, Lambda, Quake AI and Tesla plan to deploy it. Dell Technologies, Lenovo and Supermicro will include Minerva in their AI-factory operations offerings.

Inference: Deployments by compute operators such as IREN and Bitdeer AI, together with support from Dell Technologies, Lenovo and Supermicro, give Minerva paths to spread through both end customers and system suppliers. Bitdeer AI said Minerva can be deployed as a standalone platform or integrated with NVIDIA DSX OS for AI-factory operations.

“Minerva gives our operations team one view across the network, the GPUs and the jobs, so we can focus more on proactive maintenance.” — Denis Skrinnikoff, CTO of IREN

The case for and against

The case for: Minerva puts infrastructure and AI workloads in one observability framework and can connect anomalies to affected jobs. NVIDIA said IREN is using it, while Bitdeer AI said it supports Minerva. In a single NVIDIA internal AI-factory test case reported by Bitdeer AI, an NVMe power fault caused a multi-day outage; Bitdeer AI reported that Minerva reduced remediation time in that case to six hours, combined 81 noisy events into one alert and saved 54 hours of cluster uptime. This was a partner-reported single case, not broadly validated remediation performance.

Inference: Because Minerva reads NVIDIA hardware, networking, schedulers and workload data and can integrate with DSX OS, adoption could embed observability in NVIDIA’s existing full-stack AI-factory management workflow.

The case against: Minerva itself provides a turnkey platform, while open APIs and Grafana integration preserve flexibility; deployment options range from DIY to full turnkey. Those choices may affect implementation, but the evidence provides no comparison of costs or performance.

Risks

NVIDIA said Minerva analyzes billions of data points per hour and predicts anomalies; those statements about capability and scale come primarily from NVIDIA.

As of October 21, 2026, NVDA closed at $221.97, 9.3% above its 200-day moving average but 2.2% below its 20-day moving average. The stock data provided no financial statements, valuation multiples or analyst estimates.

What to watch

Inference: If the partner-reported test results can be reproduced in more environments, Minerva’s value could be measured more by reduced downtime and higher effective GPU hours than by simply adding hardware, because it links hardware anomalies to affected jobs and aggregates events.

The evidence does not provide Minerva’s pricing, contract sizes or financial contribution to NVIDIA. It also provides no independent third-party validation of the six-hour repair time, 81 events or 54 hours of cluster uptime, and does not specify actual deployment timing, cluster scale or continued use by IREN and other planned customers. No cost, performance or market-share comparison with existing observability and AIOps vendors is provided.