Source boundary Blocks & Files reported the supplier addition. Performance figures below are DDN and Nebul test results published by DDN, not independent benchmarks. Nebul's own catalog is the source for the three named storage roles; undisclosed topology and commercial terms are left unknown.
AI INFRASTRUCTURE NEOCLOUD KV CACHE 2026-10-22What changed at Nebul
Dutch sovereign-AI cloud provider Nebul is using DDN Infinia to store and reuse the key-value (KV) cache created during large-language-model inference. The new layer does not replace every storage system in Nebul's estate. Nebul's current storage catalog separately lists DDN ExaScaler for parallel file workloads, NetApp ONTAP for connecting enterprise datasets to private-AI, retrieval-augmented generation and analytics workloads, and VAST Data as a unified data platform for AI and analytics.
That makes this a workload-placement story, not a clean vendor bake-off. Model context, training data, enterprise files and general cloud volumes have different access patterns and governance requirements.
The benchmark claim, with its boundary intact
DDN says Nebul tested the same eight GPUs, model and workloads on both sides, across short prompts, long contexts and concurrent demand. It reports nearly 10× faster time to first token than Nebul's CPU-and-memory cache, up to 5× faster token responses than alternative solutions, 2.5× more input tokens per second from the same GPUs, and nearly 60% lower cost per million input tokens.
Those are supplier-published results from Nebul's environment. DDN does not publish the model name, prompt mix, sequence-length distribution, cache-hit ratio, software versions, full competing configurations or raw run data in the announcement. The numbers establish a reported result, not a generally reproducible ranking.
Why external KV cache can matter
An inference server builds a KV cache while processing prior tokens. Recomputing that context consumes GPU time; retaining it only on one host also ties the useful state to that server. DDN's described design stores KV blocks outside individual GPU servers so compatible context can be restored elsewhere in the fleet when its key remains available.
DDN says tenant, subtenant and dataset scope remain part of the key namespace, with eviction policies applied at dataset and tenant level as capacity fills. Those controls matter in a shared neocloud, but the announcement does not disclose retention defaults, encryption details, failure-domain layout or an independent tenant-isolation assessment.
Where NetApp fits
NetApp's role is the enterprise-data bridge. Nebul describes ONTAP as the way customers connect existing NetApp environments and expose datasets to private AI, RAG and analytics. NetApp's customer profile likewise says Nebul uses its data infrastructure for a sovereign-cloud model, while Nebul's service terms identify AFF A90 as an ONTAP workhorse. None of those sources says ONTAP provides the newly announced KV-cache tier.
This distinction is useful for architects: an enterprise system of record can feed an inference pipeline without also serving every transient cache access. Place each data class against its own latency, throughput, protection, retention and tenancy requirements. The site's Azure NetApp Files reference covers the Microsoft-managed service; Azure NetApp Files is not named in this Nebul announcement and should not be inferred from the NetApp relationship.
Questions buyers should take from the result
- Measure hit rate as well as restore speed. An external cache only avoids recomputation when reusable context is found under the right key.
- Test cold and warm paths. Time to first token should be reported separately for cache hits, misses and partial reuse.
- Model isolation and deletion. Verify how tenant keys, retention, eviction, encryption and secure deletion behave under capacity pressure.
- Exercise failures. A shared cache should be tested through GPU rescheduling, storage-node loss and network impairment, not only steady-state throughput.
- Keep datasets and cache state distinct. Training corpora, governed enterprise records and disposable inference state need different protection and recovery policies.
Bottom line
DDN joins a deliberately heterogeneous Nebul storage estate. Its Infinia result targets a narrow, valuable job—reusing inference context across GPU servers—while ONTAP connects governed enterprise data and VAST remains part of the broader AI data platform. The reported gains are promising; the missing test detail means buyers should treat them as a case-study result to reproduce on their own models and traffic.
Sources: Blocks & Files via Google News · DDN announcement · Nebul storage catalog · NetApp customer profile