# NVIDIA Minerva

> Source: https://aiwiki.ai/wiki/nvidia_minerva
> Updated: 2026-10-25
> Fact-checked: 2026-10-25
> Categories: AI Infrastructure, Data Centers, NVIDIA
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "NVIDIA Minerva." aiwiki.ai, 25 Oct 2026. https://aiwiki.ai/wiki/nvidia_minerva
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**NVIDIA Minerva** is an AIOps (artificial intelligence for IT operations) software platform from [NVIDIA](https://aiwiki.ai/wiki/nvidia) for monitoring and troubleshooting large GPU data centers, which the company calls [AI factories](https://aiwiki.ai/wiki/ai_factory). NVIDIA announced it on 22 October 2026 as part of its GTC Berlin announcements.[1][10] Minerva collects telemetry from GPUs, networking, storage, job schedulers and AI workloads, builds a live graph of how those components connect, and applies machine learning to detect anomalies, predict failures and group related alerts into a single incident with a likely root cause.[1][3] It includes a built-in AI assistant called Sherlock and a [Model Context Protocol](https://aiwiki.ai/wiki/model_context_protocol) (MCP) server for connecting outside AI agents.[2][3] NVIDIA positions Minerva as "the primary observability solution for NVIDIA AI factories" and as part of the full-stack [NVIDIA Vera Rubin](https://aiwiki.ai/wiki/nvidia_vera_rubin) platform.[2]

This article is about the NVIDIA software platform. For the 2022 Google Research quantitative reasoning language model, see [Minerva (language model)](https://aiwiki.ai/wiki/minerva).

| | |
| --- | --- |
| Developer | NVIDIA[1] |
| Type | AIOps and observability software platform[1][2] |
| Announced | 22 October 2026 (NVIDIA blog post, GTC Berlin announcements)[1] |
| Documented version | 1.0.1 (user guide pages dated 8-21 October 2026)[4][6] |
| Main components | NVIDIA Minerva Collector; NVIDIA Minerva Platform[3] |
| AI assistant | Sherlock (multi-agent assistant); MCP server[2][3] |
| Ingestion standards | OpenTelemetry, gRPC/gNMI, syslog, DCGM, Slurm[3] |
| Related NVIDIA platforms | Vera Rubin, Blackwell, Spectrum-X Ethernet, NVLink, DSX OS[2][3] |
| Early user named by NVIDIA | IREN[1] |
| Pricing | Not listed in launch materials; datasheet directs buyers to NVIDIA sales[3] |

## Background

### AIOps

AIOps is a research and industry field that applies machine learning and data analytics to running IT systems. A 2020 mapping study of the literature by Paolo Notaro, Jorge Cardoso and Michael Gerndt says AIOps "has been proposed to tackle modern IT administration challenges thanks to AI and Big Data" and found that 62% of the contributions it classified address failure-related tasks "such as anomaly detection and root cause analysis."[18] A 2024 survey by Lingzhe Zhang and colleagues compares approaches based on [large language models](https://aiwiki.ai/wiki/large_language_model) for AIOps failure management with traditional AIOps methods.[19] Minerva follows this pattern: anomaly detection, event correlation and root-cause analysis on infrastructure telemetry, combined with an LLM-based assistant.[3]

### Failures in large GPU clusters

Large synchronous training jobs are sensitive to individual component faults. In the [Llama 3](https://aiwiki.ai/wiki/llama_3) paper, Meta reported that "During a 54-day snapshot period of pre-training, we experienced a total of 466 job interruptions," of which 419 were unexpected, and that "Approximately 78% of the unexpected interruptions are attributed to confirmed hardware issues, such as GPU or host component failures, or suspected hardware-related issues," with GPU issues accounting for 58.7% of unexpected issues.[16] The same paper notes that "Even a single straggler can slow down thousands of other GPUs."[16]

The Minerva datasheet uses the term "gray failure" for "components that are degrading but have not yet failed," and calls this "the dominant loss mechanism in AI factories."[3] The term is the subject of a 2017 Microsoft Research paper presented at HotOS by Ryan Huang and colleagues, which argued that major availability breakdowns in cloud systems "tend to be caused by subtle underlying faults, i.e., gray failure rather than fail-stop failure."[17]

## Announcement

NVIDIA introduced Minerva in a blog post by Brian Sparks published on 22 October 2026, listed among its "NVIDIA in Europe" announcements from GTC Berlin.[1] Verdict identified Sparks as a senior director of marketing in NVIDIA's InfiniBand and Ethernet networking unit.[14] The @NVIDIAAIInfra account posted the same day that Minerva "connects infrastructure and workload telemetry across GPUs, networking, storage, and schedulers, using AI to surface root causes and reduce alert noise across factories with hundreds of thousands of GPUs."[10] NVIDIA's product page sits in its networking software section, alongside products such as [DOCA](https://aiwiki.ai/wiki/nvidia_doca), Cumulus, DSX Air, UFM and NetQ.[2] A webinar video on how IREN uses Minerva was uploaded to YouTube on 20 October 2026, two days before the blog post.[22]

The blog frames the product around "AI factory yield," which NVIDIA defines as "a measure of the useful work infrastructure produces through GPU hours, goodput, cluster uptime and token production."[1] It argues that traditional monitoring splits the data center into separate views for networking, compute, site reliability and machine learning teams, and that Minerva connects those signals.[1]

## Architecture

### Collector and Platform

The Minerva stack has two parts. The **NVIDIA Minerva Collector** gathers telemetry using OpenTelemetry (OTEL), gRPC/gNMI, syslog, DCGM and Slurm. The **NVIDIA Minerva Platform** holds the topology graph, known and unknown anomaly detection, ML auto-correlation, automated workflows and an agentic framework.[3] According to the user guide, the Collector must run on premises because it needs direct network access to monitored devices, and it can forward data both to the Platform and to any OTLP-compatible receiver such as Prometheus or VictoriaMetrics.[5]

The user guide divides the system into six functional components:[5]

| Component | Role (per NVIDIA documentation) |
| --- | --- |
| Ingestion | Collects data from external sources such as OTEL, syslog and gNMI collectors and normalizes it |
| Object modeling | Builds a relationship graph from ingested data |
| Data enrichment | Adds context in real time, for example tagging a syslog with a device role such as top-of-rack |
| Machine learning | Applies ML to enriched data, for example NLP on syslogs for zero-day anomaly detection |
| Data storage | Persists data for historical and real-time use |
| Presentation | Exposes data to external systems, for example by creating ServiceNow tickets for anomalies |

Minerva is installed with the NVIDIA Minerva Installer on a Linux cluster and scales horizontally from a single host.[5] The documentation's minimum per-node requirements for a test cluster connected to a single rack or ten Spectrum switches are 64 virtual CPUs, 128 GB of memory and 1 TB of local disk, on x86 CPUs.[9]

### Telemetry sources

NVIDIA lists telemetry from GPUs, [Spectrum-X](https://aiwiki.ai/wiki/nvidia_spectrum_x) Ethernet (scale-out, scale-across and scale-in), [NVLink](https://aiwiki.ai/wiki/nvlink) scale-up switches, SuperNICs, DPUs, hosts, storage, [NCCL](https://aiwiki.ai/wiki/nccl) and schedulers.[3] The product page names [Slurm](https://aiwiki.ai/wiki/slurm) and NVIDIA [Run:ai](https://aiwiki.ai/wiki/run_ai) as supported schedulers.[2] The operational readiness guide lists data sources to onboard, including DOCA Telemetry on hosts, Cumulus OTLP and syslog, NMX-T, Slurm event collectors and Slurm logs.[21]

### Knowledge graph and machine learning

Minerva builds what NVIDIA calls a dynamic [knowledge graph](https://aiwiki.ai/wiki/knowledge_graph) of the infrastructure (GPUs, SuperNICs, switches and cables) and the jobs running on it, which NVIDIA says lets operators trace a slow training job to the component affecting it.[1] The documentation describes this graph as automatically discovered, with objects such as devices, interfaces and queues, their properties stored as "Dimensions" and their metrics as "Measures."[20]

NVIDIA says the platform "analyzes billions of data points per hour using supervised and unsupervised machine learning," learning normal behavior, detecting previously unseen anomalies, predicting emerging failures and correlating related events by topology.[1] The datasheet lists NLP-based log [anomaly detection](https://aiwiki.ai/wiki/anomaly_detection), prediction models and "a proprietary topology-aware auto-correlation engine."[3] Instead of forwarding each symptom, Minerva can group an "event storm" into one incident with a probable root cause.[1]

### Sherlock and MCP

Sherlock is described by NVIDIA as Minerva's "multi-agent AI factory assistant," which uses live telemetry to help operators triage problems in natural language.[2] The MCP server exposes Minerva data to outside agents; NVIDIA names [OpenClaw](https://aiwiki.ai/wiki/openclaw) and Claude as examples, and the documentation gives setup steps for [Claude Code](https://aiwiki.ai/wiki/claude_code).[2][3][7] As documented for version 1.0.1, the MCP server is in beta, gives "authenticated, read-only access" to observability data, offers tools for validated read-only Druid SQL queries and (when Neo4j is enabled) topology queries, and "does not remediate issues or modify configurations." The documentation also states that it "is not tenant-aware."[7]

For remediation, the datasheet lists "LLM-recommended remediation" and either human-in-the-loop or automated mitigation through scheduler, SuperNIC and switch APIs or external tools such as Ansible, including draining nodes through the job scheduler.[3]

### Integrations

NVIDIA lists notifications to Slack, Microsoft Teams and PagerDuty, ticketing through [ServiceNow](https://aiwiki.ai/wiki/servicenow), visualization through Grafana, and scheduler integration including Slurm. Listed enterprise features are RBAC, SSO, multi-tenancy and high availability.[3] The product includes preconfigured dashboards for Spectrum-X Ethernet, the NVLink fabric, hardware health and services, which NVIDIA says follow its reference architectures.[2]

## Use cases

NVIDIA's materials group Minerva's functions into five use cases:[2][3]

| Use case | NVIDIA's description (summarized) |
| --- | --- |
| Proactive maintenance | ML anomaly detection to catch degrading components before they stop jobs, with alerts that can trigger node draining |
| Agentic debugging | Natural-language root-cause analysis through Sherlock or external agents via MCP |
| Full-stack correlation | Linking a symptom in one layer (for example GPU utilization) to a cause in another (for example a network anomaly or scheduler event), filtered by job ID |
| Noise reduction | Topology-aware correlation that presents a cascading fault as one incident, delivered through Slack or ServiceNow |
| Network observability | Built-in dashboards and alerts for Spectrum-X and NVLink networks |

NVIDIA says the platform improves mean time between interrupts (MTBI) and reduces mean time to repair (MTTR).[2][3] These are vendor claims. The documentation's case studies describe examples such as a switch page fault in which Minerva "correlated 81 distinct anomalies into a single actionable incident report," and retransmit storms that stalled NCCL collective operations.[8] Bitdeer AI, citing "NVIDIA's own internal AI Factory testing," wrote that Minerva cut MTTR for an NVMe power fault "from a multi-day outage to six hours" and in another case "collapsed 81 noisy events into a single alert, saving 54 hours of cluster uptime."[11]

## Deployment and consumption models

NVIDIA's product page and datasheet describe three consumption models; the user guide's platform overview groups them into two, a DIY observability model and a fully turnkey one:[2][3][5]

| Model | Description |
| --- | --- |
| DIY observability | Minerva used only as a data collector, feeding external tools such as Grafana |
| AIOps with DIY observability | Minerva collection plus its ML anomaly detection and correlation, feeding external observability tools and agents |
| Turnkey | The full platform with collection, analytics, built-in agent and dashboards |

NVIDIA's sources differ on hosting. The product page and FAQ say Minerva "can be deployed on premises or consumed as SaaS from NVIDIA cloud," and the user guide says the Platform Stack can run on premises or "in the cloud as a SaaS offering."[2][5] The datasheet's specification table lists deployment as "Fully on-premises," and the version 1.0.1 release notes state: "Cloud/SaaS deployments are not supported."[3][6]

The 1.0.1 release notes also list known issues, including that backend APIs "are unauthenticated by default" unless an API key is configured, and that tenant isolation applies only to certain data types.[6]

### Relation to DSX OS

Minerva is tied to [NVIDIA DSX](https://aiwiki.ai/wiki/nvidia_dsx), NVIDIA's AI factory design and operations platform. The datasheet describes DSX OS as "the operating layer for running and scaling multi-tenant AI factories" and says the Minerva Collector is part of DSX OS, letting operators correlate infrastructure telemetry with facility-level signals "for a single coherent view from GPU to grid." DSX OS customers can add the full Minerva platform as an option, and Minerva can also run standalone.[3] NVIDIA's FAQ says "key NVIDIA Minerva capabilities, starting with NVIDIA Minerva Collector, will be open sourced in DSX OS."[2] The user guide includes instructions for simulating a Minerva deployment in DSX Air.[4]

## Partners and customers

NVIDIA named IREN as "among the NVIDIA partners using Minerva."[1] [IREN](https://aiwiki.ai/wiki/iren) chief technology officer Denis Skrinnikoff said: "NVIDIA Minerva gives our operations team one view across the network, the GPUs and the jobs running on them, so we focus more on proactive maintenance rather than troubleshooting."[1]

| Company | Role stated in the announcement |
| --- | --- |
| IREN | Using Minerva[1] |
| Aolani | Plans to deploy[1] |
| Bitdeer AI | Plans to deploy; Bitdeer said on 22 October 2026 that it "now support[s]" Minerva[1][11] |
| IBM | Plans to deploy[1] |
| Jio | Plans to deploy[1] |
| [Lambda](https://aiwiki.ai/wiki/lambda_labs) | Plans to deploy[1] |
| Quake AI | Plans to deploy[1] |
| Tesla | Plans to deploy[1] |
| Dell Technologies | Named with Lenovo and Supermicro as offering Minerva in AI factory operations portfolios[1][14] |
| [Lenovo](https://aiwiki.ai/wiki/lenovo) | Will offer Minerva in its AI factory operations portfolio[1][13] |
| Supermicro | Will offer Minerva, integrated into its SuperCloud Software Suite[1][12] |

The NVIDIA blog's wording is "In addition to Dell Technologies, Lenovo and Supermicro will offer Minerva as part of their AI factory operations portfolios."[1] Supermicro said it is integrating Minerva into the SuperCloud Software Suite, the management layer of its Data Center Building Block Solutions.[12] Lenovo's GTC Berlin post said Minerva "can help teams manage AI infrastructure availability and performance as environments scale."[13] The NVIDIA announcement does not give deployment sizes or dates for the companies that plan to deploy Minerva.

## Reception

Verdict reported the launch on 23 October 2026, noting that Minerva forms part of NVIDIA's Vera Rubin offering.[14] Developer Tech's Ryan Daws wrote that the claimed gains in uptime and repair time were not "independently verified yet," while noting that IREN runs AI factories and had adopted the platform.[15]

## See also

- [NVIDIA DSX](https://aiwiki.ai/wiki/nvidia_dsx)
- [NVIDIA Vera Rubin](https://aiwiki.ai/wiki/nvidia_vera_rubin)
- [NVIDIA Spectrum-X](https://aiwiki.ai/wiki/nvidia_spectrum_x)
- [OpenTelemetry](https://aiwiki.ai/wiki/opentelemetry)
- [MLOps](https://aiwiki.ai/wiki/mlops)
- [AI infrastructure](https://aiwiki.ai/wiki/ai_infrastructure)

## References

1. Sparks, Brian. "NVIDIA Unveils Minerva to Bring End-to-End Observability for AI Factories." NVIDIA Blog, 22 October 2026. https://blogs.nvidia.com/blog/minerva-platform-infrastructure/
2. "NVIDIA Minerva: AI-Native AIOps for AI Factories." NVIDIA product page, accessed 25 October 2026. https://www.nvidia.com/en-us/networking/products/software/minerva/
3. "NVIDIA Minerva Datasheet." NVIDIA, October 2026. https://dam-cdn.nvd.orangelogic.com/AssetLink/1pedat80a16our7i026o1q38e3f8q2g8.pdf
4. "NVIDIA Minerva User Guide" (version 1.0.1). NVIDIA networking documentation, last updated 8 October 2026. https://networking-docs.nvidia.com/minerva/1.0.1
5. "Platform Overview." NVIDIA Minerva User Guide 1.0.1, last updated 13 October 2026. https://networking-docs.nvidia.com/minerva/1.0.1/platform-overview
6. "Release Notes." NVIDIA Minerva User Guide 1.0.1, last updated 13 October 2026. https://networking-docs.nvidia.com/minerva/1.0.1/release-notes
7. "NVIDIA Minerva MCP." NVIDIA Minerva User Guide 1.0.1, last updated 15 October 2026. https://networking-docs.nvidia.com/minerva/1.0.1/nvidia-minerva-mcp
8. "Case Studies." NVIDIA Minerva User Guide 1.0.1, last updated 14 October 2026. https://networking-docs.nvidia.com/minerva/1.0.1/case-studies
9. "Resource Requirements." NVIDIA Minerva User Guide 1.0.1, last updated 13 October 2026. https://networking-docs.nvidia.com/minerva/1.0.1/resource-requirements
10. NVIDIA AI Infrastructure (@NVIDIAAIInfra). Post on X, 22 October 2026. https://x.com/NVIDIAAIInfra/status/2113313944966177176
11. Xiong, Evelyn. "Bitdeer AI Brings NVIDIA Minerva's AI-Native AIOps For Maximizing AI Factory Yield." Bitdeer AI blog, 22 October 2026. https://www.bitdeer.ai/en/blog/nvidia-minerva/
12. "Supermicro Integrates NVIDIA Minerva into SuperCloud Suite to Maximize AI Factory Yield." Supermicro Data Center Stories, 22 October 2026. https://learn-more.supermicro.com/data-center-stories/supermicro-nvidia-minerva-supercloud-ai-factory
13. "Europe's AI moment: turning investment into impact." Lenovo StoryHub, 22 October 2026. https://news.lenovo.com/europes-ai-moment-gtc-berlin-26
14. Swagath, BV. "Nvidia launches Minerva AIOps platform for AI factory observability." Verdict, 23 October 2026. https://www.verdict.co.uk/nvidia-minerva-aiops-platform/
15. Daws, Ryan. "NVIDIA Minerva targets AI factory observability." Developer Tech, 23 October 2026. https://www.developer-tech.com/news/nvidia-minerva-targets-ai-factory-observability/
16. Llama Team, AI @ Meta. "The Llama 3 Herd of Models." arXiv:2407.21783, 2024. https://arxiv.org/abs/2407.21783
17. Huang, Ryan; Guo, Chuanxiong; Zhou, Lidong; Lorch, Jay; Dang, Yingnong; Chintalapati, Murali; Yao, Randolph. "Gray Failure: The Achilles' Heel of Cloud-Scale Systems." Proceedings of the ACM Workshop on Hot Topics in Operating Systems (HotOS), May 2017. https://www.microsoft.com/en-us/research/publication/gray-failure-achilles-heel-cloud-scale-systems/
18. Notaro, Paolo; Cardoso, Jorge; Gerndt, Michael. "A Systematic Mapping Study in AIOps." arXiv:2012.09108, December 2020. https://arxiv.org/abs/2012.09108
19. Zhang, Lingzhe; et al. "A Survey of AIOps for Failure Management in the Era of Large Language Models." arXiv:2406.11213, June 2024. https://arxiv.org/abs/2406.11213
20. "Network Model." NVIDIA Minerva User Guide 1.0.1, last updated 21 October 2026. https://networking-docs.nvidia.com/minerva/1.0.1/network-model
21. "AI Factory Operational Readiness Review." NVIDIA Minerva User Guide 1.0.1. https://networking-docs.nvidia.com/minerva/1.0.1/ai-factory-operational-readiness-review
22. "NVIDIA Minerva: The AI-Native AIOps Platform for Maximizing AI Factory Yield." NVIDIA, YouTube, 20 October 2026. https://youtu.be/AYGSUv-EAfY