At NetApp INSIGHT 2026, there was plenty about AI-ready data, sovereignty, and hybrid cloud, but to me, it was pretty clear that Novus was the headliner. You’ll find some of my thoughts in the official NetApp Novus press release, where I make a familiar point about AI infrastructure. As providers add GPUs, the economics increasingly depend on keeping those GPUs busy, and storage that can’t keep up leaves them idle. Novus is NetApp’s answer to that problem in the biggest AI factories, and it runs on ONTAP, NetApp’s storage operating system. NetApp pitched Novus on speed, which, of course, is the whole point. Still, I really think ONTAP is the key differentiator that will matter most to the neoclouds this announcement was aimed at.

Novus takes metadata off the data path so GPUs wait less on storage

Novus is a new data architecture for AI factories, specifically the huge GPU clusters that train and run AI models. NetApp aims Novus at the major public clouds, where NetApp’s first-party storage services already set it apart. Because Novus is hyper-focused on GPUs, though, the pitch centers on neoclouds, the newer cloud providers built around renting out GPUs.

For those providers, GPUs may wait before any data moves. A storage system has to find and open a file before a GPU server can read it, and that lookup is metadata work. NetApp’s launch blog points out that a conventional array handles those lookups on the same controllers that serve the data. Those shared controllers struggle when a training job saves a checkpoint, a huge write from every node that records the job’s progress. The checkpoint can land right behind a flood of small lookups. Each workload then slows the other, and the GPUs wait.

Novus splits metadata from data. New software called Novus Data Director handles metadata on its own servers, while ONTAP systems hold the data. A GPU server asks the Data Director where to find a file, then reads it directly from storage.

Some file systems built for supercomputers already keep metadata on separate servers, so the design has a track record. Novus pairs that design with ONTAP and the data services enterprises already run. NetApp is putting more money behind metadata, too. Days before INSIGHT, NetApp announced its intent to acquire PEAK:AIO, which it calls a pioneer in metadata architecture and parallel file systems. To me, the planned deal adds depth to the same metadata design Novus is built on.

Standard Linux clients spare providers a fleet-wide software rollout

The GPU servers themselves reach Novus via the standard NFS client built into supported Linux distributions. And with no proprietary client to install, a provider running thousands of GPU servers has one less layer of software to deploy and patch. I’d expect the benefit to grow in a neocloud, where customers may run their own server images, and a proprietary client would have to be installed on each. Think about it this way. Over the years that a cluster runs, operators keep updating Linux and adding more, newer GPUs. Something proprietary would need retesting after every update. Each retest might feel like a lightweight check, but across thousands of servers, retesting could slow how fast a provider gets customers back to actively using their GPUs.

ONTAP gives neoclouds the data services enterprise customers expect

Enterprises already count on ONTAP’s security and resiliency in their own data centers, and Novus carries both into AI factories. In July, I wrote that newer vendors built their storage for AI speed, while handling enterprise data safely over many years is a different discipline. Novus brings that NetApp discipline and proven enterprise readiness to the top end of AI infrastructure.

Novus also keeps ONTAP’s multitenancy, which matters far more in a cloud than in a single company’s data center. A bank renting GPUs from a neocloud, for example, would expect its data to stay as private as the GPUs themselves. With tenant-level encryption and security controls for both people and AI agents, NetApp is covering that need. I’d bet that bank cares more about its data sitting on ONTAP than about any throughput number.

The AI Data Engine widens discovery to other vendors’ storage

Back on the enterprise side, metadata also drove the AI Data Engine update. The AI Data Engine is NetApp’s service that helps transform scattered, unstructured enterprise data into AI-usable data. NetApp is now expanding its metadata discovery past NetApp’s own systems to other vendors’ file and object storage.

I like that the discovery pays off before any AI project starts. NetApp says it helps reduce security risk, optimize storage, and speed up data migrations. Those benefits give a storage team its own reason to turn the AI Data Engine on.

Oracle and Keystone Sovereign aim to widen NetApp’s hybrid reach

NetApp and Oracle plan a fully managed ONTAP storage service built into Oracle Cloud Infrastructure. NetApp’s first-party storage services already run on AWS, Microsoft Azure, and Google Cloud, so Oracle would bring the total to four. To me, a fourth cloud makes it even harder for anyone to match that first-party position.

NetApp is also adding sovereignty controls to Keystone, its storage subscription, for European customers. Keystone Sovereign is designed to help keep data and support within Europe for qualified customers, backed by contractual and legal terms. I think sovereignty has become a buying decision in its own right. So covering things like who can access the data, who supports it, and where it sits is music to European customers’ ears.

Novus puts ONTAP in the running for the largest GPU clusters

Coming back to Novus, I’d say the biggest AI factories have mostly bought storage from specialist vendors. Novus is NetApp’s most direct bid yet for that business, and ONTAP gives that bid its edge.

On the enterprise side, the AI Data Engine, the Oracle service, and Keystone Sovereign extend NetApp’s lead with hybrid buyers. With Novus on top, NetApp now sells ONTAP from the enterprise data center to the largest GPU clusters. Novus is orderable now, and I’m curious to hear what the first providers see regarding GPU utilization.