NetApp announced on November 10, 2026 that Nebul, a European AI-native cloud provider, is the first neocloud to adopt its Novus storage architecture for sovereign AI infrastructure. Nebul will use Novus in its sovereign Intelligent Cloud. The architecture is generally available and designed for up to 100 TB per second of aggregate throughput.
Nebul describes its Intelligent Cloud as a sovereign platform combining accelerated infrastructure, enterprise data services and production AI capabilities while preserving customer ownership and control. Beyond the adoption, the companies are exploring opportunities to accelerate sovereign-AI uptake by bringing together Nebul’s services and the NetApp Platform.
For Nebul’s multi-tenant environment, NetApp says Novus can scale performance, capacity and concurrency independently beneath one namespace as demand grows. Nebul founder and CEO Arnold Juffer said: “With this architecture, we can scale our AI Factory to serve growing customer demand without sacrificing the performance, security or compliance they depend on.”
How Novus Separates Metadata From Data
Novus separates metadata, the information that maps file locations, from movement of the data itself. NetApp’s technical explanation describes a client requesting a layout from Novus Data Director and then communicating directly with ONTAP-based storage. Data Director is an independently scalable, software-defined control layer running on standard x86 compute, allowing the metadata and data planes to scale separately. This addresses the conflict that arises when small-file metadata operations and large checkpoint writes share controllers.
The Novus product documentation says the system works with standard in-kernel NFSv4.2 and pNFS Flex Files clients using nconnect, plus NVIDIA GPUDirect Storage. Operators therefore do not need custom client software across thousands of GPU servers. Additional ONTAP clusters add throughput without remounts or fragmentation, while every GPU continues to see the same namespace.
The ONTAP foundation also carries the data-governance side of the announcement. NetApp says the design is intended to let organizations using its platform extend existing data estates into sovereign cloud environments while preserving consistent security, protection and governance, including current jurisdiction and access policies. Willem Hendrickx, NetApp’s senior vice president and general manager for EMEA and LATAM, identified government, healthcare and biotech as regulated industries where providers must address compliance, security and multi-tenancy.
From Product Launch to General Availability
NetApp introduced Novus on September 29, 2026, making the architecture orderable that day. The initial release combines Data Director software on qualified Supermicro infrastructure with ONTAP data services on AFF A90 systems. NetApp described a path toward software-defined deployments over time while maintaining the same namespace, operating model and data foundation.
GPU Utilization and Storage Scale
NetApp’s economic argument is that GPUs waiting on storage still depreciate without generating a return. At the September introduction, the company said utilization can fall below 30% under traditional architectures when storage cannot supply data quickly enough. The launch therefore framed storage throughput as a direct constraint on GPU-cluster economics.
NetApp technical fellow Arindam Banerjee’s explanation illustrates the bandwidth requirement with one GPU demanding as much as 2 GB per second and 50,000 GPUs requiring that rate simultaneously. It puts conventional storage arrays at 40–80 GB per second; reaching Novus’s aggregate target with such arrays would require more than 100 systems, each adding a namespace, failure domain and management overhead. The comparison connects the architecture’s single namespace to the operational overhead of scaling storage.
For Nebul, the adoption will test whether independently scaling metadata and data can support more tenants and GPUs without fragmenting data across storage systems, while jurisdiction, access and governance controls remain consistent. The operational challenge is to expand shared AI infrastructure without weakening the isolation and control that regulated customers require.
