To position itself in the “AI factories,” NetApp has chosen its path: it will be pNFS.
This extension of the NFS specification emerged in the early 2000s. It favors parallel access by disaggregating the metadata server.
The metadata server can communicate to clients different layouts (data layouts) depending on the systems. NetApp has chosen the so‑called Flex Files, which allows aggregating NFS servers within a single namespace. A solution it deems ideal at the scale of AI factories to avoid the coordination burden tied to separate mount points.
The enterprise NAS option with a shared-everything architecture has been ruled out, essentially because it cannot perform this disaggregation.
Lustre supports it*… but the metadata server remains monolithic. The traditional response to scale—partitioning the namespace—failed to convince NetApp, which argues that it merely defers the bottleneck. It instead preferred a technology that has recently been grafted into the pNFS ecosystem: Lattice.
The PEAK:IO Acquisition Comes Into Focus
Open-source released with the support of the Linux Foundation, Lattice provides a user-space execution engine organized around a distributed authority model (a RonDB cluster). These properties ease upscaling—up to 72 nodes—and open the door to ephemeral metadata-server instances.
The orchestration with the storage infrastructure relies on direct RPC calls. The engine currently implements the NFSv4.2/pNFS personality. The goal is to make it a foundation for other distributed architectures.
With technical input from the Los Alamos National Laboratory, the project began evolving toward an open-core model. Modules (data restructuring, weighted distribution, tiering, …) emerged under the impulse of a startup: PEAK:IO. NetApp knows it well, having acquired it.
Novus is expected to deliver more than 100 TB/s of aggregated bandwidth, at approximately zettabyte scale, with “quasi‑linear” scaling. For the moment, data must reside on ONTAP shelves (there is no SDS option yet). The metadata server runs on x86 hardware.
To read in addition:
VAST Data takes on confidential GPU computing with NVIDIA
AI gateway: the next battlefield of AI infrastructures
Enterprise storage: hard drives in a revival phase
Microsoft aims to optimize every AI token
AI gigafactories: details of the bidding process
* Another limitation of Lustre: it requires maintaining a kernel tailored to each node. This requirement does not exist with pNFS, which is integrated natively in Linux.
