NVIDIA Launches Minerva for AI-Factory Observability and Agentic Operations

NVIDIA launched Minerva on October 22

NVIDIA announced Minerva on October 22, 2026 as an end-to-end observability and operations platform for large AI training and inference environments. It is infrastructure software, not a foundation model or public inference API. The product’s job is to connect physical and logical signals that are usually split among networking, compute, storage, site-reliability and machine-learning teams, then relate them to the jobs consuming the hardware.

Minerva is available through a sales-led route rather than a posted self-service plan. NVIDIA has published a product page, datasheet and versioned user guide, and says IREN is using the product. It names Aolani, Bitdeer AI, IBM, Jio, Lambda, Quake AI and Tesla as planned deployers, while Dell, Lenovo and Supermicro plan to offer Minerva in their AI-factory operations portfolios. NVIDIA does not publish list price, license unit, support tier or a generally available trial on the reviewed pages.

The platform connects infrastructure signals to AI jobs

The datasheet separates the system into the Minerva Collector and Minerva Platform. The collector gathers metrics, logs and topology through OpenTelemetry, gRPC/gNMI, syslog, DCGM (NVIDIA Data Center GPU Manager) and Slurm. Named sources include GPUs, Spectrum-X Ethernet, NVLink switches, SuperNICs, DPUs, hosts, storage, NCCL, schedulers and AI workloads. The platform maps those signals into a dynamic graph of components and jobs across one or more data centers.

That job-to-infrastructure mapping is the relevant differentiator to test. A network pause, warming GPU or degrading NVMe device can appear as a workload slowdown rather than a clean hardware failure. NVIDIA says Minerva can connect an anomalous component to the affected job and collapse cascades of symptoms into one incident. A buyer should verify this against its own topology, scheduler, telemetry fidelity and failure history rather than assume every custom device or software layer is covered.

Machine learning drives anomaly detection and correlation

NVIDIA says Minerva processes billions of data points per hour with supervised and unsupervised models, NLP-based log analysis, prediction models and a proprietary topology-aware correlation engine. Its stated goals are to learn normal behavior, identify known and previously unseen anomalies, predict failures and reduce alert noise. The launch article frames outcomes in mean time between interruptions, mean time to repair, cluster uptime, goodput, effective GPU hours and token production.

Those mechanisms are documented as product capabilities, but the public launch material does not provide a benchmark corpus, false-positive and false-negative rates, incident denominator, confidence intervals or an independent comparison with an incumbent observability stack. “Previously unseen” should therefore be read as NVIDIA’s product claim, not a guarantee that novel failures will be caught. Trial scoring should include missed incidents and unnecessary interventions, not only alerts that Minerva successfully groups.

Sherlock and MCP add an agentic control surface

Minerva includes a native assistant named Sherlock for natural-language investigation grounded in live telemetry. NVIDIA also documents an MCP server that exposes platform capabilities to external agentic tools. Operators can query cluster health, investigate a suspected root cause and receive recommended remediation. The public user guide separately exposes MCP and REST API sections, which makes the agent interface an implementation surface rather than a launch-slide promise.

The authority boundary is the consequential part. The datasheet allows LLM-recommended remediation to remain human-in-the-loop or become automated through scheduler, SuperNIC and switch APIs plus systems such as Ansible. Treat query, recommendation and execution as separate permissions. A diagnostic assistant should not inherit node-drain, switch-change or workload-termination authority merely because it can see the relevant telemetry.

On-premises deployment still creates integration work

NVIDIA specifies a fully on-premises, horizontally scalable microservices deployment across collector and platform stacks. It claims RBAC, SSO, multi-tenancy and high availability, and lists notifications to Slack, Microsoft Teams and PagerDuty, ticketing through ServiceNow, visualization in Grafana and scheduler integration including Slurm. DSX OS customers can add Minerva, while the datasheet also says the platform can run standalone.

The datasheet describes a range from DIY use of the collector, through collector-plus-platform alongside external observability tools, to a turnkey deployment. That flexibility does not reveal the implementation cost. Buyers still need the node and service footprint, storage retention, telemetry cardinality limits, upgrade and rollback process, supported versions, disconnected-operation behavior, tenancy isolation, backup requirements and support response commitments for their environment.

The launch evidence is useful but vendor-controlled

NVIDIA reports IREN as a current user and quotes its CTO describing a unified view across networks, GPUs and jobs. The datasheet claims proven hyperscale operation across deployments with hundreds of thousands of GPUs. The launch page also says Minerva can increase time between interruptions, shorten repair time and surface fewer, more useful alerts. These statements establish intended scale and customer interest; they do not disclose a controlled before-and-after study or a complete incident ledger.

No reviewed source gives a public price, a total-cost example, a test license duration, deployment labor, minimum supported cluster, model-update policy or measured overhead on the monitored systems. Procurement should request those items in writing. The absence of a list price is not evidence that the product is unavailable, but it prevents an outside cost-per-effective-GPU-hour comparison.

What AI infrastructure teams should do next

Start with a bounded replay and shadow-mode trial. Feed Minerva a representative slice of the target topology and historical telemetry containing known GPU, network, storage, scheduler and job failures. Freeze the incident labels before scoring. Measure ingestion completeness, detection delay, root-cause rank, false alerts, missed incidents, operator time, platform overhead and whether a recommendation would have helped or harmed the affected workload.

Keep all remediation advisory-only until the replay and live shadow period clear explicit gates. Then allow one reversible action—such as draining an already unhealthy node—behind policy validation, a named approver, rate limits and rollback. Adopt when measured recovered GPU time and reduced operator effort exceed license, integration and operations cost. Wait when stack coverage or pricing remains unclear. Reject closed-loop deployment when Minerva or an attached agent can alter shared infrastructure without independent authorization and durable audit evidence.

Disclosures

AccessAllGPT received no NVIDIA software, hardware, license, pricing, account, support, briefing, demo, customer reference, review, payment or compensation for this article. NVIDIA and the named customers and partners did not sponsor, review or endorse it. AccessAllGPT did not test, score or rank Minerva. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with NVIDIA or organizations cited. Publication-wide relationships are listed on the disclosures page.