NVIDIA launches Minerva observability platform for AI factories

NVIDIA launched Minerva, an AIOps platform that brings infrastructure and workload telemetry together to detect anomalies, correlate incidents and support remediation across large AI computing clusters.

Operations teams get a shared view across hardware, networks, storage, schedulers and running jobs. The monitoring and remediation features are documented. NVIDIA’s claims that Minerva will improve uptime, repair times and token production are not independently measured in the available evidence.

NVIDIA’s product materials say Minerva draws signals from compute, networks, storage, schedulers, NCCL and AI workloads. They list OpenTelemetry, gNMI, syslog, DCGM, Slurm and JSON among the supported standards and data sources.

How the cross-stack view works

Minerva builds a dynamic knowledge graph that maps physical components, including GPUs, SuperNICs and switches, to the jobs that depend on them. NVIDIA says the model can connect a symptom in one layer, such as a network-interface anomaly, with its effect on GPU utilization or job throughput elsewhere in the stack.

According to NVIDIA, purpose-built machine-learning models learn normal behavior, flag anomalies, predict emerging failures and use topology-aware correlation to group related alerts into fewer incidents. The company says Minerva analyzes billions of data points per hour and can scale to environments with hundreds of thousands of GPUs. Those scale figures are vendor claims; the available sources do not include an independent benchmark.

NVIDIA also describes a native assistant called Sherlock for natural-language investigation and a Model Context Protocol server for connecting external agents. The launch post says operators can receive remediation recommendations, keep a person in the loop or connect Minerva to automation through infrastructure APIs and external systems. Those are described capabilities; any operational improvement depends on a customer’s configuration and environment.

Deployment and consumption options

NVIDIA markets Minerva for on-premises deployment or consumption as SaaS from NVIDIA cloud. The opened product page does not identify which Minerva releases support the SaaS option. NVIDIA also offers consumption models ranging from DIY observability to a turnkey AIOps platform. In the DIY path, Minerva can act as a data collector for external observability software such as Grafana. Another configuration adds Minerva’s anomaly-detection and correlation output to customer-managed observability and agent tools.

The product page names Slack, Microsoft Teams, PagerDuty, ServiceNow, Grafana and Ansible among its integrations. It says Minerva Collector is part of NVIDIA DSX OS and that Minerva can add AI and machine-learning insights within DSX OS or operate as a standalone product. NVIDIA’s DSX documentation describes DSX OS as modular software for operating multi-tenant AI-factory infrastructure and lists observability, health, resilience, workload orchestration and infrastructure operations among its functions.

The opened materials do not provide pricing, licensing terms or firm delivery details for each consumption model. They also do not establish which integrations are included by default rather than requiring customer configuration or additional products.

Capabilities versus reported benefits

NVIDIA says Minerva can lengthen the time between interruptions, reduce repair time, prevent hundreds of lost GPU hours and increase token production. The available sources contain no controlled benchmark or independently audited before-and-after results for those claims.

The company says IREN is using Minerva, while Aolani, Bitdeer AI, IBM, Jio, Lambda, Quake AI and Tesla plan to deploy it. NVIDIA also says Dell Technologies, Lenovo and Supermicro will offer Minerva in their AI-factory operations portfolios. The announcement gives no deployment dates, fleet sizes, measured customer results, pricing or delivery terms.