Models

NetApp has chosen its direction to invest in AI factories

October 4, 2026 · 2 min read

Also published in Português (Brasil), Italiano

macro photography of black circuit board
Alexandre Debiève / Unsplash

NetApp has announced its decision to invest in AI factories, selecting pNFS as its pathway. This extension of the NFS specification emerged in the early 2000s and promotes parallel access by disaggregating the metadata server. The server can communicate different layouts to clients based on the systems used. NetApp has opted for the Flex Files layout, which allows for the aggregation of NFS servers within the same namespace. This solution is deemed ideal for AI factories to avoid the coordination burden associated with separate mounting points.

The enterprise NAS option with a shared-everything architecture was dismissed, primarily due to its inability to operate this disaggregation. Although Lustre implements this, its metadata server is monolithic. The conventional response for scaling, partitioning the namespace, did not convince NetApp, which believes it only shifts the bottleneck. Instead, it has preferred a technology recently integrated into the pNFS ecosystem: Lattice.

The acquisition of PEAK:IO makes sense in this context. Released as open source with the support of the Linux Foundation, Lattice provides a user-space execution engine built around a shared authority plan (a RonDB cluster). These features facilitate scaling up to 72 nodes and pave the way for ephemeral metadata server instances.

Orchestration with the storage infrastructure relies on direct RPC calls. The engine currently implements the NFSv4.2/pNFS personality, aiming to serve as a foundation for other distributed architectures. With technical assistance from the Los Alamos laboratory, the project has begun evolving towards an open core model. Modules for data restructuring, weighted distribution, and prioritization are emerging, driven by a start-up: PEAK:IO, which NetApp has acquired.

The Novus system is expected to deliver over 100 To/s of aggregated bandwidth at a zettabyte scale, with nearly linear scaling. Currently, data must reside on ONTAP arrays (no SDS option yet). The metadata server operates on x86 hardware.