AI Novus ONTAP INSIGHT 2026 Oct 02, 2026
The release in one minute
The initial Novus configuration combines Novus Data Director metadata software on qualified Supermicro infrastructure with AFF A90 systems providing ONTAP data services. NetApp says it is orderable now. By taking metadata management out of the data path, the design lets operators add metadata concurrency, bandwidth and capacity on different curves while retaining a single namespace over industry-standard NFS.
That is the product change. The claimed destination — more than 100 TB/s sequential read throughput, dozens of exabytes of effective capacity and zettabyte-scale deployments — is not the same thing. Omdia observed near-linear scaling in NetApp’s test as ONTAP clusters were added, then modelled the much larger result. NetApp’s “fastest storage on the planet” line should therefore be read as a vendor claim, not as an independent production benchmark.
What changed versus the previous product model
Traditional scale-out NAS binds namespace work and data service to the storage-controller architecture. Parallel file systems can spread bandwidth, but often add a specialist client and a separate operating model. Novus takes a different route:
- Before: metadata and data scaling were coupled more tightly to the array/controller design. Scaling one pressure often meant scaling the whole storage system.
- Novus: a separate Data Director tier owns metadata while AFF A90 ONTAP systems serve data. The resources can scale independently beneath one namespace.
- Continuity: the first release still uses AFF A90 and ONTAP data services; it is not a replacement ONTAP release or a firmware feature unlocked on an arbitrary existing cluster.
- Future direction: NetApp describes a path to software-defined deployments, but its release warns that future features, timing and availability can change.
The distinction matters for procurement: an existing AFF A90 estate does not automatically become Novus after an ONTAP upgrade. The qualified Supermicro metadata tier, supported data-node configuration, client stack and network design are part of the system.
What the analyst reaction actually says
Seeking Alpha reported that NetApp shares edged higher after the event and summarized three constructive but guarded calls. Morgan Stanley analyst Erik Woodring said Novus extends NetApp beyond traditional enterprise storage toward a broader AI data-infrastructure platform, but the firm retained an Equal-weight rating and $191 target. Citi highlighted an AI-storage market it expects could grow from $20 billion to $40 billion over three years and kept Neutral. Oppenheimer liked the direction but said it wanted execution and customer feedback before becoming bullish, retaining Market Perform.
So “analysts react positively” does not mean ratings upgrades. It means they saw a credible expansion in addressable market while withholding a stronger call until NetApp shows delivery, adoption and economics. For infrastructure teams, that maps neatly to the technical evidence still missing: validated configurations, field reliability, mixed-workload results and operational recovery behaviour.
Upgrade and adoption considerations
There is no announced in-place upgrade path from conventional ONTAP NAS to Novus. Treat adoption as a new architecture project, not a routine software change. Before a purchase order, require:
- A supported bill of materials. Pin the qualified Data Director servers, AFF A90 data-node variant, ONTAP release, switches, NICs, firmware, Linux kernels and pNFS client settings.
- A workload-shaped proof. Test the real file-size distribution, metadata operations, checkpoint pattern, client count and failure cases. A sequential-read projection is not a proxy for small-file ingest or checkpoint recovery.
- A migration and rollback design. The announcement does not disclose a conversion workflow for an existing namespace. Document how data enters Novus, how access is cut over and how clients return to the previous path.
- A support boundary. Record who owns incidents across Supermicro compute, the Data Director tier, ONTAP data nodes, the client OS and the Ethernet fabric.
Known unknowns, not known defects
No Novus defect list or release notes were identified in the public announcement material. That absence should not be misread as a clean field record. NetApp has disclosed the basic architecture and orderability, but not the Data Director HA topology, upgrade sequencing, namespace recovery objectives, minimum supported scale, mixed-workload benchmark data, or behaviour during a partial metadata-tier outage. Those are acceptance-test questions, not confirmed product issues.
Operational consequence for admins
The main change is a second control and failure domain. ONTAP teams now have to correlate a client request across metadata servers, an Ethernet client fabric and AFF A90 data nodes. Monitoring that stops at cluster latency or CPU cannot tell whether a stalled training job is waiting on metadata, network transport or the data tier.
- Build separate health and saturation signals for metadata requests, client-fabric loss/congestion and ONTAP data service.
- Test loss of one component at a time, including Data Director failover and partial fabric isolation, while watching what mounted clients experience.
- Define where tenant identity, QoS, snapshots, replication and retention are enforced across the two tiers.
- Keep the AI client network separate from the GPU collective backend where the validated design requires it; map MTU, ECN/PFC and redundant paths before load testing.
Use the ONTAP network architecture reference to turn that fabric discussion into a port, VLAN, LIF, MTU and failure-domain diagram. For the deeper pNFS and Data Director mechanics, see our technical Novus architecture analysis.
Bottom line
Novus is a material architecture change: metadata scales outside the AFF A90 data path rather than forcing the whole storage stack to scale together. The first configuration is orderable, but the largest performance claims are modelled and the software-defined future is roadmap. The analysts’ neutral ratings are consistent with the admin verdict: technically promising, commercially broader, and still dependent on execution evidence.