NVIDIA has launched Minerva, an AIOps software platform designed to diagnose hardware failures and maintain stability across large AI data centers. The system tracks real time telemetry across GPUs, switches, and storage fabrics to consolidate chaotic alert streams into single root cause diagnoses. Major infrastructure operators including Tesla, IBM, and IREN have already committed to deploying the software across their clusters.
Modern AI training clusters often stumble when 1 component begins to degrade under heavy load. A slightly overheated GPU or a slow storage drive can stall an entire distributed training run. Minerva addresses this by building a live knowledge graph of the physical network, linking individual GPUs, cables, and SuperNICs directly to the software jobs running on them. The platform processes billions of telemetry points every hour to catch emerging hardware faults before a full system crash occurs.
Instead of inundating technicians with hundreds of simultaneous warnings during a network hiccup, Minerva isolates the underlying fault and suggests remediation paths. Operators can interact with the system using a natural language interface or hand off fixes directly to automated infrastructure APIs. Denis Skrinnikoff, chief technology officer at IREN, described the operational shift:
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.
NVIDIA designed the platform to boost total cluster yield, tracking operational uptime, token generation rates, and raw compute efficiency. Client adoption spans several cloud and enterprise providers, with deployments planned at Aolani, Bitdeer AI, Jio, Lambda, and Quake AI alongside Tesla and IBM. Server manufacturers Dell Technologies, Lenovo, and Supermicro will also bundle Minerva within their standard enterprise hardware catalogs.
