NVIDIA Minerva: An AIOps Platform to Keep AI Factories From Silently Slowing Down

NVIDIA has launched Minerva, an AIOps platform meant to help operators of large GPU clusters find problems earlier and keep expensive hardware busy. According to NVIDIA's announcement, NVIDIA Unveils Minerva to Bring End-to-End Observability for AI Factories, it ties telemetry from infrastructure and AI workloads into one view and applies "purpose-built AI models" to turn that data into incidents with a likely root cause. It arrived during the GTC Berlin announcements.

There are no performance figures in the announcement, so this guide explains what Minerva claims to do, why the problem it targets is real, and what to ask NVIDIA before treating the claims as proven.

TL;DR

The problem: one slow part drags the whole job

Large training and inference jobs run across thousands of GPUs joined by fast networks, and in synchronized training the slowest participant sets the pace. NVIDIA's own examples show how subtle the failures can be: a GPU running hotter than its peers, or an NVMe drive losing performance under load. Neither necessarily throws an error, but either can slow an entire job. At the scale NVIDIA describes, with "hundreds of thousands of GPUs", finding that component by hand is slow.

Traditional monitoring, NVIDIA argues, divides the data center into separate views, so networking, compute, site reliability and machine learning teams each see only their part. Minerva's pitch is to connect those signals across NVIDIA GPUs, networking, storage, schedulers and workloads.

How Minerva works, as described

• Knowledge graph. It maps the infrastructure (GPUs, SuperNICs, switches, cables) and the jobs running on it, so operators can trace a slow training job to the component affecting it.
• Anomaly detection and prediction. It analyzes billions of data points per hour, learns normal behavior, detects previously unseen anomalies, predicts emerging failures and correlates related events using the factory topology.
• Incident grouping. Instead of sending each symptom to an operator, it can group an event storm into one focused incident with a probable root cause.
• Assistant and agent hooks. A native AI assistant offers a natural-language interface grounded in live telemetry, and Minerva connects to external agentic AI tools. Operators can ask about cluster health and receive recommended remediation.
• Human in the loop or automated. Teams can apply fixes manually or connect Minerva to automation through infrastructure APIs and external systems.

An eye above a row of bars with one green bar, illustrating AI factory observability dashboards

The metric NVIDIA wants you to watch: AI factory yield

The platform "focuses on AI factory yield," which NVIDIA defines as the useful work infrastructure produces through GPU hours, goodput, cluster uptime and token production. Goodput is the share of compute time that goes into productive progress rather than waiting, restarting from checkpoints or running degraded. For anyone paying for GPU time, this is the number that converts failures into dollars, which is why vendors increasingly sell reliability tooling next to chips. Our coverage of NVIDIA's DSX pod and the open-source KAI scheduler shows the same theme: NVIDIA is selling the whole factory, not just the GPU.

NVIDIA lists three outcomes: longer time between interruptions, faster time to repair, and greater insight by filtering the "flood of data" down to the few alerts that matter.

Who is behind it

IREN's chief technology officer, Denis Skrinnikoff, is quoted saying Minerva gives his operations team "one view across the network, the GPUs and the jobs running on them" so they can focus on proactive maintenance. NVIDIA says Aolani, Bitdeer AI, IBM, Jio, Lambda, Quake AI and Tesla plan to deploy Minerva, and that Dell Technologies, Lenovo and Supermicro will include it in AI factory operations offerings. "Plan to deploy" is not "in production," and only IREN is described as using it today.

What is claimed versus what is shown

None of this means the product does not work. It means the announcement is a launch, not a study. Questions worth asking NVIDIA or a reseller:

1. False positives and misses. What are the precision and recall on failure prediction, and how many incidents were caught before a job failed?
2. Time to repair. What was the measured reduction, on what baseline?
3. Non-NVIDIA gear. NVIDIA says it spans NVIDIA GPUs, networking, storage, schedulers and workloads; how well does it cover third-party storage, power and cooling systems?
4. Data handling. Telemetry about jobs may reveal customer workloads. What leaves the cluster, and who can see it? Compare NVIDIA's push on confidential computing for European AI clouds.
5. Agent permissions. If Minerva is wired to automation, what actions can it take and who approves them? Agentic remediation inside a data center is a high-stakes permission, the same issue we examined in NVIDIA's open agent safety platform.

Where it fits in the observability market

Cluster observability is already served by general tools and by cloud and GPU-vendor software, and developers are building observability for the agents that run on top, such as Perplexity's open-source numbat tool. Minerva sits lower in the stack, at the hardware and network layer, and its advantage is access to NVIDIA's own telemetry across GPUs, SuperNICs and switches. The trade-off is the usual one for vendor platforms: deep integration with one vendor's stack, less neutrality. Large operators with heterogeneous fleets will compare it with what they run today.

Why it matters beyond data centers

When a frontier model's training run stalls, or inference capacity is short, the cause is often not a lack of chips but time lost to failures and degraded nodes. Better detection translates into more tokens per dollar and per megawatt, a theme throughout this week's announcements, including the Google Cloud Vera Rubin general availability. If tools like Minerva work, they raise effective capacity without new construction, which matters in a world of power-limited sites. If they do not, operators will keep relying on in-house tooling and expert intuition.

A worked example of the incident-grouping idea

Imagine a training job that runs 8 percent slower than last week. Nothing has crashed, and each team's dashboard looks acceptable. A real cause might be one switch port with rising error counts, forcing retransmissions that stall a handful of GPUs, which in turn hold up every other GPU at the next synchronization step. A conventional alert system may fire on the port, on the GPUs' idle time, on job throughput and on storage latency, producing four separate tickets owned by four teams. Minerva's described approach is to use the topology graph to see that these symptoms share one upstream component and present a single incident with a likely root cause. Whether it gets the cause right often enough is exactly what the missing benchmarks would show.

The same logic applies to the thermal example in NVIDIA's post. A GPU that runs hotter than its peers may throttle slightly, which shows up only as a small throughput loss that averages out in coarse metrics. Learning each device's normal behavior, rather than applying one fixed threshold, is what the announcement means by supervised and unsupervised learning working together.

Companies named in the announcement

IREN operates AI cloud and data-center capacity (iren.com), and Lambda is a GPU cloud provider (lambda.ai); both sell GPU hours, so lost uptime hits revenue directly, which explains their interest. Tesla and Jio are examples of large operators running their own clusters. Dell, Lenovo and Supermicro are server makers that can bundle Minerva with hardware, which gives NVIDIA a distribution channel into enterprise AI factories.

What to do next

• If you operate a cluster, request a proof of concept with your own failure history and compare predictions against what actually happened.
• If you rent GPU capacity, ask your provider whether it uses an AIOps tool and how it reports goodput and interruptions.
• If you follow the sector, watch for measured results from IREN, Lambda or Tesla, which would turn this announcement into evidence.

Related reading

• NVIDIA DSX pod: containerized Vera Rubin AI factory
• NVIDIA Run:ai open-source KAI scheduler
• Google Cloud A5X Vera Rubin general availability
• NVIDIA confidential computing for European AI clouds
• NVIDIA open agent safety platform
• Perplexity numbat agent observability
• Official: NVIDIA Minerva announcement

Details are accurate as of October 22, 2026 and come from NVIDIA's announcement. No independent benchmarks were available.