NetApp's Novus splits data, metadata for AI factories

NetApp and Nvidia's partnership dates back more than a decade and now includes co-engineering around AI infrastructure. That long relationship matters because the storage problems Banerjee describes are not abstract: as AI moves from experimentation into what Nvidia's Jason Hardy called "phase 2" of mainstream production, the infrastructure has to scale without being redesigned for every new use case.

Banerjee's argument is that AI workloads create a specific conflict. Transactional metadata operations and heavy sequential workloads — checkpoints, for example — happen at the same time, and older architectures could only do one well. When metadata and data share the same resources, data operations queue behind smaller metadata operations, which leaves GPUs underutilized while still burning power. Novus's answer is to split those functions onto separate scaling axes.

Agents raise the stakes further, since thousands may hit data concurrently with temporary permissions, increasing metadata volume. The bet is that API-driven, agent-friendly consumption — rather than storage specialists managing every request — becomes the norm. Whether this matters for any given enterprise hinges on how quickly agentic production workloads arrive and whether the new control plane makes the layer invisible to the AI teams meant to use it.