NVIDIA has launched Minerva, an AIOps platform for AI factory operations designed to help teams spot faults earlier and maintain output across large AI infrastructure estates.
Minerva brings together telemetry from infrastructure and AI workloads, including GPUs, networking, storage, schedulers, and the jobs running across them. The platform uses AI models to analyse operational data and turn it into alerts, incident groupings, and remediation suggestions.
The release reflects the growing operational challenge facing operators of large AI clusters, where a single underperforming component can affect a wider training or inference workload. AI factories can involve hundreds of thousands of GPUs connected across high-speed networks, making it difficult for operations, networking, reliability, and machine learning teams to work from a single view of events.
Single view
According to NVIDIA, Minerva creates a dynamic knowledge graph that maps physical infrastructure, such as GPUs, SuperNICs, switches, and cables, alongside the jobs using those resources. The model is intended to help operators trace a slowdown in a training run or service back to the infrastructure affecting it.
The system analyses billions of data points an hour using supervised and unsupervised machine learning. It is designed to learn what normal behaviour looks like, identify anomalies, predict failures, and correlate related events based on a site's topology.
Rather than presenting operations teams with large volumes of separate alerts, Minerva groups related signals into a smaller number of incidents with a likely root cause. It also includes a native AI assistant and can connect to external agentic AI tools, giving staff a natural-language interface to query cluster health and review proposed actions.
Teams can either keep human operators responsible for implementing fixes or connect the platform to automation through infrastructure application programming interfaces and external systems.
Yield focus
The platform is centred on what NVIDIA calls AI factory yield, which it defines through measures such as GPU hours, goodput, cluster uptime, and token production. The aim is to help operators extend the time between interruptions, shorten repair times, and reduce the volume of low-value alerts that can obscure more serious issues.
These features target a key commercial issue for companies building AI data centres: keeping expensive computing assets running at useful output for as long as possible. In large training environments, symptoms such as a hotter-than-normal GPU or a storage drive losing performance under load may seem minor in isolation, but can still affect a much larger job.
According to NVIDIA, IREN is already using the software, while Aolani, Bitdeer AI, IBM, Jio, Lambda, Quake AI, and Tesla plan to deploy it. Dell Technologies, Lenovo, and Supermicro will also offer Minerva within their AI factory operations portfolios.
IREN provided a customer example. "As we scale our AI factories, the key to providing the best service for our customers is maximizing the productivity of our GPUs," said Denis Skrinnikoff, Chief Technology Officer at IREN. "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. That visibility strengthens our operating model for the infrastructure we're building next."
The launch adds to a broader push by infrastructure suppliers to move beyond selling chips and servers into the software used to run AI estates day to day. Operators are increasingly looking for tools that combine hardware telemetry with workload data, particularly as AI systems grow in size and as downtime or degraded performance becomes more costly.
For NVIDIA, Minerva also extends its reach deeper into the operational layer of AI data centres by linking its GPUs and networking products with software that monitors how those systems behave in production. The platform is intended to turn event storms into a focused incident with a probable root cause.