Enterprises are investing heavily in AI, but most are only using a fraction of their data to power it. At NetApp INSIGHT in Las Vegas, we sat down with Jürgen Hofkens, the new CTO for EMEA at NetApp. We discussed the reasons for the gap, but also how NetApp new NOVUS architecture is engineered to close it.
Hofkens, who joined NetApp on May 1 as CTO for EMEA, arrived at a moment he describes as a “perfect storm” for data and storage. Customers across industries are investing in AI, but the results are falling short of expectations. The core issue, he argues, is a lack of data accessibility.
“Our customers are not even using 5% of the data they have to be successful in their AI projects,” Hofkens says. The reason is structural. Organizations build AI projects in silos, creating isolated copies of data for each specific use case. This approach works for a proof of concept, but it becomes a liability at enterprise scale.
The problem compounds on the hardware side. A cluster of 50,000 GPUs, the kind of scale seen in neocloud AI factories and large pharmaceutical or banking deployments, requires approximately 100 terabytes per second of storage throughput. Today, storage systems typically max out at around 5 terabytes per second. That is a 20x gap between what AI infrastructure demands and what storage can deliver.
NetApp Novus wants to solve the metadata bottleneck
NetApp’s answer to the throughput gap is Novus. This was the headline announcement at NetApp INSIGHT. The innovation behind Novus is deceptively straightforward: disaggregate the metadata service from the rest of the storage architecture.
With its earlier AFX platform, NetApp had already separated compute from storage, allowing each to scale independently. But scaling to 100-terabyte-per-second namespaces surfaces a new constraint. The metadata service, the lookup layer that tells clients where a file or its constituent pieces are located, became the bottleneck, not the compute or storage nodes themselves.
Novus moves the metadata service into its own independently scalable layer. This is not a new concept in high-performance computing, where similar approaches have been used for years. What is new is bringing it to enterprise storage. The result: namespaces that can scale to exabyte and zettabyte levels.
No proprietary protocols
What is interesting from a technology level is that Novus runs on ONTAP and uses parallel NFS (pNFS). The latter is an open standard. NetApp doesn’t use any proprietary protocols in Novus. This was not an accident, Hofkens tells us. NetApp wanted to ensure that enterprises with existing data on ONTAP could expose that data directly without copying it and without renegotiating governance models.
“It’s always easy to just design for speed,” Hofkens acknowledges. “It becomes a little bit more important if you have to do that and also take all of the enterprise constraints, regulatory environments, resilience, backup, cyber resilience into account.” Delivering both, raw throughput and enterprise-grade trus, is what Hofkens calls “the real engineering feat.”
When asked about it, Hofkens says that the performance delivered is also sustained, not only available in bursts. Hofkens says that the speeds NetApp claims for Novus are steady-state figures, not peak numbers that drop off as buffers fill.
Novus and neoclouds
Novus is designed with a specific ecosystem dynamic in mind, Hofkens says. Neoclouds, specialized GPU cloud providers that are rapidly expanding capacity across Europe and beyond, need enterprise customers. Enterprise customers need GPU access. The challenge is getting enterprise data, which typically lives on ONTAP in NetApp environments, to those GPUs without creating new data silos.
Novus solves this challenge by letting a neocloud provider offer ONTAP-native access to GPU clusters. Hofkens notes that ONTAP compatibility is already becoming a differentiator for neoclouds trying to attract large enterprises. “These new clouds want to go to our customers and basically say: you have the data, we will make sure that you can use that data on our GPUs.”
For European enterprises in particular, this matters. Sovereignty requirements mean that AI projects often need to run in specific jurisdictions with specific governance controls. Novus enables that, Hofkens says. Customers retain their encryption keys and governance models while flexibly moving workloads across locations.
Hardware flexibility: Supermicro and beyond
Novus launches on NetApp’s own A90 hardware. This is standard x86 technology, so nothing proprietary or exotic. In parallel, NetApp is working with ODMs including Supermicro to bring Novus capabilities as software deployable on commodity hardware. This is consistent with how NetApp has operated in the cloud for years, where ONTAP runs on hyperscaler infrastructure rather than proprietary hardware.
The CPU era and the role of AI agents
Looking further ahead, Hofkens addresses a growing trend: the increasing importance of CPUs alongside GPUs. Two forces are driving this. First, some inferencing workloads are migrating to CPUs as model optimization improves. Second, and more significant, AI agents run on CPUs. Agents may call large language models running on GPUs, but the orchestration layer operates on CPU.
One of the implications of this is that agents run on file systems. “An enterprise file system is important for an agent because it gives you the structure to monitor your agents,” Hofkens explains. “You are able to constrain them, understand what they’re doing. You are able to monitor them so that rogue agents can be detected.” Enterprise file systems, in other words, are not just storage. They are a governance layer for agentic AI.
Listen to and/or watch the full conversation to get many more details around NetApp Novus and the broader AI story around data.
Also read: 100 TB/s to fight GPU hunger: NetApp Novus is “fastest storage ever”