Reported by BigGo Finance from NetApp Korea’s 2026 Insight Press Briefing (2026-10-14). Only claims present in that report are asserted here; throughput and capacity figures are NetApp’s stated targets.
Product AI 2026-10-16What NetApp announced
NetApp introduced Novus, a next-generation storage architecture aimed at large-scale AI data centers, on 2026-10-14. The design target is the AI factory case: hundreds of thousands of GPUs requesting data at the same moment. NetApp says the architecture is built to exceed 100 TB/s of total throughput, with a single file system scaling to zettabyte capacity.
The technology was first shown at NetApp’s annual event in Las Vegas on 2026-09-29 and presented to Korean press at NetApp Korea’s “2026 Insight Press Briefing” at the Grand InterContinental Seoul Parnas. Yoo Jae-sung, CEO of NetApp Korea, framed the pitch around return on investment rather than raw capacity: “Companies are making massive investments in AI, but they don’t yet seem to be at the stage of realizing return on investment (ROI). We will help customers reduce the time and cost required to extract insights from data and efficiently utilize the infrastructure resources they have invested in.”
The bottleneck is metadata, not capacity
The rationale is structural. As training and inference workloads run at the same time, how fast data can be pulled from storage to feed GPUs sets the ceiling for the whole system — sufficient compute capacity does not help if storage cannot keep pace, because GPUs sit idle waiting for data.
Novus attacks that by separating the paths metadata and actual data travel, distributing the metadata processing load that appears when many GPUs simultaneously verify file locations and access permissions. Kang Yeon-sik, Executive Director at NetApp Korea: “In existing storage environments, even when data processing performance has headroom, if a bottleneck occurs in metadata, it becomes difficult to supply data to GPUs smoothly.”
NetApp describes the scaling model as splitting the tiers: scale the metadata tier to raise concurrent processing capacity, add data nodes to raise throughput, and expand capacity independently. Access is plain NFS — no dedicated software install on the GPU nodes. Teams already running NFS scale-out on ONTAP will recognise the pattern; the metadata-caching argument is the same one behind FlexCache, pushed to a different scale.
AI Data Engine: process data where it already lives
NetApp is also pitching a different way to feed AI, the AI Data Engine (AIDE): it uses data scattered across different storage systems in place, without consolidating it first. The scope is deliberately wider than ONTAP — metadata discovery is being extended to third-party storage — with the stated payoff of fewer replication cycles, less data staleness and less security exposure from copying data around.
Kang: “The NetApp AI Data Engine transforms dark unstructured data assets into governance-based context for AI. As AI agents increasingly access enterprise data, the need to manage not only data location but also access permissions and security is growing.” If agents start hitting these datasets directly, governance moves from an add-on to part of the data path — see our AI topic hub for how that is playing out across the portfolio.
Oracle Cloud joins the roster
Cloud support expanded with Oracle Cloud added alongside existing AWS, Azure and Google Cloud coverage. Management runs through NetApp Console under consistent policies, with AI-based autonomous storage operations covering failure prediction and capacity management. The on-prem-versus-cloud framing is unchanged from what we covered in on-prem vs cloud, but the console is now the single control plane for both.
Cyber resilience with Commvault
NetApp expanded its collaboration with Commvault, the backup and recovery specialist, to strengthen an integrated framework spanning ransomware detection, data isolation, verification and recovery. The reasoning given is that AI agents sharply increase data access and therefore the attack surface — the data repository itself has to become an active security entity. For the ONTAP-side mechanics, ransomware protection covers the snapshot, ARP and tamper-proof snapshot primitives that underpin that kind of stack.
What admins should watch
Treat the headline figures as architecture targets, not shipped SKUs. The briefing gave no version numbers, part numbers, pricing or general-availability dates for Novus, and no statement on how the metadata tier is delivered — software on existing ONTAP hardware, or new nodes. Three questions decide whether this matters to a given shop: when Novus becomes orderable, whether the metadata tier is a separate purchase, and whether in-place discovery in AIDE reaches the third-party arrays holding a legacy estate. Until those are answered, the practical action for ONTAP teams is unchanged: know where the metadata hot spots are in your NFS workloads, because that is the same measurement Novus is being sold against.
Sources: BigGo Finance — “NetApp Unveils ‘Novus’ Storage for AI Factories, Targeting 100TB/s Throughput” (2026-10-14). The briefing reported the material first shown at NetApp’s Las Vegas event on 2026-09-29.