DDN sells two storage substrates and stops there. EXAScaler is DDN's build of open-source Lustre on AI400X appliances. It's the training-era workhorse, and its layouts, quotas, and movement tools run on any Lustre. Infinia is a proprietary object platform sold as a subscription on qualified servers, on DDN appliances, or in Google Cloud and OCI, and it's what DDN's inference pitch rests on. Where the I/O path meets the GPU, DDN is the furthest along in the storage column: its Infinia backend for NVIDIA's NIXL library is the only documented, generally available KV-cache tier among the storage vendors on the instrument. Above that, nothing. No retrieval, no serving, no agent plane, no applications on DDN's own paper.
Authority splits along the product line. EXAScaler's Lustre core reads Retained. Infinia's tenants, roles, erasure-code engine, and management surfaces read Ceded, and that capture is coupled and visible: the objects and the tenant model live in DDN's namespace. The KV tier is the exception that keeps the inference story honest. It sits behind NIXL, an interface with several other backends, and the state it holds is recomputable, so leaving costs a configuration change, not a migration. The forward capture is elsewhere. Everything DDN is building for inference (Infinia, Horizon, HyperPOD, the Aleria factory) sits on the proprietary half. The open half is the heritage.
The buyer gets an HPC-grade data path and a KV-cache tier that ships today. In exchange, they assemble retrieval, serving, and agent governance from someone else, and accept that most of DDN's product documentation sits behind a partner login, so much of the pitch (metadata search, Horizon's GPU allocation, the Aleria SKU) can't be checked against public documents yet. DDN is a storage choice, not a platform choice. Buy it for the data path, and plan the rest of the stack as a separate decision.
Layer-by-layer status: Layer 0 (Two Storage Substrates: EXAScaler, DDN's Build of Open-Source Lustre on AI400X Appliances, and Infinia, a Proprietary Object Platform on DDN Appliances, Qualified Servers, or GCP and OCI Instances; No Compute, No Fabric of Its Own), Layer 1A (Lustre's Quotas, ACLs, Changelogs, and Encryption Under EXAScaler; Infinia's S3 Object Store With Tenants and Role-Based Access; No Catalog, No Classification, No Lineage; Infinia's Metadata Search Undocumented), Layer 1B (No Retrieval Product: No Vector Index, No Search, No RAG Pipeline on DDN's Paper; Infinia Stores the Corpus (1A) and the KV Cache (1C)), Layer 1C (KV Cache Tiering to Infinia Through an Open-Source NIXL Backend for NIXL-Enabled Dynamo, vLLM, SGLang, and LMCache Builds; Lustre's HSM, Mirroring, Migration, and Client Caching Inside EXAScaler; No Pipeline Engine, No Infinia Replication; Hot Pools Undocumented), Layer 2A (The EXAScaler CSI Driver Presents Lustre to Kubernetes and OpenShift (Topology Pins, Snapshots, Automated Lustre Module Install) and an Infinia Block CSI Driver; Infinia's Cluster Management and Marketplace Terraform Deployment; No Compute or GPU Scheduling; Horizon Undocumented), Layer 2B (No Serving, No Agent Runtime: DDN Feeds Other Engines' KV Caches and Weights From Storage; Inferno and Horizon's Endpoint Provisioning Undocumented), Layer 2C (No Reasoning Plane: Tenant Isolation and Access Control Govern Data, Not Agents; Horizon's Tenant Controls Undocumented), Layer 3 (+1) (No Value Plane on DDN's Paper: Storage Beneath Others' Applications; the Aleria Sovereign AI Factory Is a Press-Release Reference Architecture; IndustrySync in Early Access).
Assessment framework: 4+1 Layer AI Infrastructure Model. Scoring model: Decision Authority Placement Model (DAPM) — Retained, Delegated, or Ceded. Published by The CTO Advisor LLC (DBA The Advisor Bench). Author: Keith Townsend. Date assessed: October 15, 2026. Version: v1.0 - New Row, Scored on Public Evidence.
DDN sells two storage substrates and stops there. EXAScaler is DDN's build of open-source Lustre on AI400X appliances. It's the training-era workhorse, and its layouts, quotas, and movement tools run on any Lustre. Infinia is a proprietary object platform sold as a subscription on qualified servers, on DDN appliances, or in Google Cloud and OCI, and it's what DDN's inference pitch rests on. Where the I/O path meets the GPU, DDN is the furthest along in the storage column: its Infinia backend for NVIDIA's NIXL library is the only documented, generally available KV-cache tier among the storage vendors on the instrument. Above that, nothing. No retrieval, no serving, no agent plane, no applications on DDN's own paper.
Authority splits along the product line. EXAScaler's Lustre core reads Retained. Infinia's tenants, roles, erasure-code engine, and management surfaces read Ceded, and that capture is coupled and visible: the objects and the tenant model live in DDN's namespace. The KV tier is the exception that keeps the inference story honest. It sits behind NIXL, an interface with several other backends, and the state it holds is recomputable, so leaving costs a configuration change, not a migration. The forward capture is elsewhere. Everything DDN is building for inference (Infinia, Horizon, HyperPOD, the Aleria factory) sits on the proprietary half. The open half is the heritage.
The buyer gets an HPC-grade data path and a KV-cache tier that ships today. In exchange, they assemble retrieval, serving, and agent governance from someone else, and accept that most of DDN's product documentation sits behind a partner login, so much of the pitch (metadata search, Horizon's GPU allocation, the Aleria SKU) can't be checked against public documents yet. DDN is a storage choice, not a platform choice. Buy it for the data path, and plan the rest of the stack as a separate decision.
Raw compute, networking, and acceleration fabric
The community parallel filesystem DDN builds EXAScaler on and maintains upstream. Retained: the enterprise can run community Lustre without DDN, and the opinions it accumulates in layouts, pools, and quotas run on any other Lustre implementation, the HPE DAOS reading.
DDN's integrated appliance and the software it adds above community Lustre. Ceded: the management, tiering policy, and controller opinions have no second implementation, and the appliance is lifted only as DDN deployed it.
DDN's proprietary key-value object platform, its client libraries, and its installers. Ceded, the VAST and WEKA reading; the cloud instances under it are scored on the clouds' rows.
Commodity x86 servers the enterprise buys and runs Infinia on. Retained, the WEKA and Nutanix hardware reading; the validation narrows the choice without owning it. Dell and Lenovo qualification is named, not scored.
Interfaces with independent implementations: S3 across object stores, the Lustre client across DDN, HPE, and the clouds' managed Lustre services. Delegated, the NetApp storage-networking reading. GPUDirect Storage is NVIDIA's cuFile stack and sits in the NVIDIA note; Infinia's native SDK is DDN's and sits in the Infinia chip.
EXAScaler has supported NVIDIA GPUDirect Storage since EXAScaler 6 (July 2021). DDN's NVIDIA reference architectures pair the AI400X2 with DGX SuperPOD (DGX B200 and GB200), and the AI400X3 and AI400X3I with NVIDIA Cloud Partner designs (HGX H200, B200, and B300; GB200 and GB300). The AI400X3M lists Spectrum-X networking at XDR 800G, NDR 400G, and 200GbE. The Enterprise AI HyperPOD rack carries RTX PRO 6000 Blackwell GPUs, BlueField-3, and Spectrum-X on Supermicro servers. GPU-initiated data access with NVIDIA (announced August 4, 2026) has no availability. The filesystem and the object platform run on AMD and Intel x86 servers; the NVIDIA dependency is the compute the enterprise brings.
DDN sells storage substrate in two shapes and nothing else at this layer. The first is EXAScaler, DDN's enterprise build of the open-source Lustre parallel filesystem (GPL v2; DDN has maintained the community distribution since buying Intel's Lustre business, Whamcloud, in 2018), sold on AI400X-class appliances: the AI400X2 Turbo, the AI400X3 (launched June 10, 2025, a 2U AMD Genoa PCIe Gen5 system paired with NVIDIA Cloud Partner reference architectures), and the AI400X3M (announced June 22, 2026 at ISC with 'general availability ... expected by the end of Q3 2026', which no public surface has since confirmed). The second is Infinia, a proprietary key-value-based object platform sold as 'a software subscription solution offered on qualified DDN platforms', on DDN's own Infinia appliances (the AI2200 datasheet), on servers the enterprise buys (Supermicro's DDN solution page: 'Validated with the Supermicro Storage A+ 1U server'), and on public-cloud instances the enterprise rents: DDN's public deployment guides for version 2.4.1 cover Google Cloud Marketplace on n2d-standard-32 with 24 local NVMe drives (seven-VM minimum, 63 TB and 28 GB/s, a DDN activation token required) and OCI on BM.DenseIO.E5.128 with a RoCEv2 back end. DDN sells no compute and no fabric: networking is InfiniBand or Ethernet from NVIDIA (Spectrum-X on the AI400X3M), and the GPUs are the enterprise's. The Enterprise AI HyperPOD (October 28, 2025) is a full rack with GPUs, BlueField-3, and Spectrum-X, but it is built on Supermicro servers and sold, integrated, and supported through Supermicro as a turnkey package; the compute in it is Supermicro's Layer 0 under the channel-assembled-design reading (NetApp FlexPod, Everpure AIRI), and the Infinia inside it is scored here. Google Cloud Managed Lustre runs on EXAScaler but is sold on Google's paper and scored on the GCP row. Calibration: moderate is the storage column's grade when the vendor sells a storage substrate without compute or fabric (VAST, WEKA, Everpure, NetApp), against Dell and HPE strong on compute, fabric, and storage together. DDN's publicly readable product documentation is thin: Infinia's documentation site is behind a customer and partner login except its version index (2.3.0 through 2.4.1) and the GCP and OCI guides, and EXAScaler's public surfaces are the community Lustre manual, DDN's CSI driver repository, product pages, and NVIDIA's reference architectures.
Retained at the filesystem and the metal, Ceded at DDN's software and appliances. Lustre is open source the enterprise can run without DDN, and its layouts, pools, quotas, and striping move to any Lustre (community builds, HPE's ClusterStor, the clouds' managed Lustre services): Retained, the HPE DAOS reading. The Supermicro servers validated for Infinia are the enterprise's own purchase: Retained, the WEKA and Nutanix hardware reading. What DDN adds on top of Lustre (the EXAScaler management framework, Hot Pools policy, and the appliance's own storage controller software) and all of Infinia are DDN's opinions with no second implementation: Ceded, the VAST, WEKA, and NetApp reading. S3 and the Lustre client protocol are multi-vendor interfaces: Delegated. The runtime call at this layer is where data lands. On EXAScaler the enterprise can set a file's or directory's stripe layout and pool, pin it to specific storage targets, and migrate it after the fact (lfs setstripe and lfs migrate in the Lustre manual), a per-object placement the engine honors and a documented manual action that reverses it: vendor decides, visible, overridable, Delegated. Infinia places objects by erasure code across its nodes with no documented per-object override; that half alone would read Ceded.
Watch-listed with dates: AI400X3M ('general availability is expected by the end of Q3 2026', ISC press release June 22, 2026; not confirmed as of October 15, 2026). Named, not scored: GPU-initiated data access with NVIDIA ('availability will be shared as development progresses', August 4, 2026); BlueField-4 alignment (Computex, June 1, 2026); the A1400X3i and A12200 names from SC25 coverage; Infinia on qualified Dell and Lenovo servers, stated in press coverage of the October 2026 IT Press Tour but in no public document (Supermicro's validation is public and scored); EXAScaler Cloud on AWS and Azure Marketplace, described by DDN as self-help proof-of-concept listings. Evidence boundary: the Infinia documentation site (partner-docs.ddn.io) requires a customer or partner login for every page except the version index and two cloud guides; gated pages are not cited. Public evidence that moves the cell: DDN selling compute or fabric on its own paper.
Durable, governed data foundation — the Governance Catalog that Layer 2C queries
Governance mechanisms documented in the community Lustre manual and carried through DDN's CSI driver. Retained: the policies the enterprise writes run on any Lustre.
DDN's proprietary store and its tenancy, access, and protection model. Ceded, the VAST and WEKA reading.
Buckets and objects behind the S3 API, which independent stores implement. Delegated, the NetApp StorageGRID and WEKA S3 reading; the supported-operation list is not public.
Neither EXAScaler's data services nor Infinia's object store depends on NVIDIA software. The NVIDIA ties are at Layer 0 (GPUDirect Storage, reference architectures) and at the KV-cache path, scored where they perform their function.
DDN's governed data foundation is two stores with storage-grade governance and no catalog. EXAScaler carries what Lustre carries, documented in the community Lustre manual: user, group, and project quotas; POSIX ACLs; nodemaps that map client networks to identities, the mechanism behind Lustre multi-tenancy; changelogs that record every namespace change; and client-side encryption with fscrypt. DDN's EXAScaler CSI driver (v2.10.0, October 1, 2026; OpenShift certification listed for 4.13 to 4.15 and 4.19 to 4.22) marks encryption (not yet on OpenShift) and compression generally available for Kubernetes volumes; the driver itself is scored at Layer 2A with the peers' CSI doors. Infinia, per DDN's public deployment guides for 2.4.1, is an 'S3-compatible object store' with 'multi-tenancy with role-based access control' and four-way replication for metadata; its IO engine picks the erasure-code layout in real time ('erasure coding levels are not set at cluster build time', the Infinia datasheet), and the OCI guide sizes clusters against 8+3P and 16+3P; managed through a web interface, the redcli command line, and a REST API. Under rule 9 the grade turns on what the catalog tells a reasoning plane about the data, and DDN publishes no catalog: no classification, no sensitivity labels, no lineage. Infinia's marketing describes metadata stored in the same key-value engine as the data and 'searchable at write-commit', and the 2.0 launch named a metadata tagging system, but no public documentation describes a tagging or metadata-query interface, so the capability is named, not scored. Lustre changelogs record what changed, not what the data is. That is the WEKA, Everpure, and NetApp reading: real storage with real governance, moderate, below VAST's Catalog and Dell's MetadataIQ, which answer what the data is (NetApp's Metadata Engine, the third such catalog, is in private preview and unscored).
Split along the same seam as Layer 0. Lustre's quotas, ACLs, nodemaps, changelogs, and encryption are community code the enterprise can run on any Lustre: Retained, the HPE DAOS reading. Infinia's tenants, subtenants, datasets, and role model are DDN's with no second implementation: Ceded, the VAST and WEKA reading. Infinia's S3 interface is a multi-vendor standard and the buckets and objects behind it lift to any S3 store: Delegated, the NetApp StorageGRID and WEKA S3 reading. The runtime call at this layer is the store enforcing the quotas, tenant boundaries, and access rules the enterprise set, and on Infinia protecting data by its own erasure-code layout: vendor decides, visible (the Infinia web interface, redcli, and the REST API; Lustre's quota and changelog tools), not overridable, Ceded, the NetApp, VAST, WEKA, and Everpure reading.
Named, not scored (public surfaces only; Infinia's documentation beyond the version index and two cloud guides requires a partner or customer login): Infinia metadata tagging and search ('searchable at write-commit', DDN blog, about April 2026; product page 'high-speed metadata indexing'); Infinia 2.4 identity integration, quotas, and governance controls (press release, July 7, 2026); Infinia POSIX access ('Limited Availability' in the 2.4 press release); Infinia 2.5 (IT Press Tour coverage, October 14, 2026, and a NIXL pull request; no release notes); EXAScaler's ISC 2026 additions (bare-metal multi-tenancy, KMIP encryption, file pinning, VictoriaLogs integration; press release, June 22, 2026); DDN Insight and the CORE data plane (SC25 announcements, no documentation). Infinia's block CSI driver appears only in the Kubernetes CSI driver list. Watch-list items need a date: Infinia's 'Coming (2.4)' items (POSIX GA with ACLs, AD and LDAP RBAC, per-tenant KMIP, object lock) carry none confirmed. Public evidence that moves the cell to strong: documentation of a catalog a reasoning plane can query for classification, sensitivity, or lineage.
Low-latency retrieval for RAG — vector/hybrid search, context windows
DDN's RAG reference demonstration runs NVIDIA NIM microservices for embedding, reranking, and inference with Milvus over Infinia, and the Enterprise AI HyperPOD carries NVIDIA AI Enterprise with NIM and NeMo Retriever under the NVIDIA AI Data Platform design; both put the retrieval engine on NVIDIA's and Milvus's code and, for HyperPOD, on Supermicro's paper.
DDN sells no retrieval capability. There is no DDN vector index, no hybrid search, no embedding pipeline, and no RAG service on DDN's paper. What DDN publishes at this layer is enablement: a reference demonstration of NVIDIA NIM microservices and Milvus reading a corpus from Infinia (DDN blog, undated, 'to demonstrate' rather than a validated design), and the Enterprise AI HyperPOD's NVIDIA AI Data Platform design, which Supermicro sells, integrates, and supports and which is scored on Supermicro's row. Infinia stores the corpus the enterprise's retrieval stack reads, which is Layer 1A. Infinia's KV-cache backend for NVIDIA's NIXL transfer library holds attention state the serving engine computed, not documents a retriever selects; the instrument's Layer 1C purpose line names KV cache tiering, so the backend is scored there, where WEKA's Augmented Memory Grid and Everpure's KV Cache Accelerator are named. Calibration: NetApp reads moderate on a validated RAG design on its own paper (AIPod Mini), WEKA and Everpure moderate on retrieval products they ship (the AI Data Platform, Data Stream), VAST strong on a native engine; DDN ships neither a product nor a design on its own paper, so the layer is gap. Native Kafka and vector-database interfaces are reported for 'next year' (iTWire, October 14, 2026) and carry no date.
Nothing to borrow at this layer from DDN. The enterprise's retrieval stack (its own or NVIDIA's blueprint with Milvus, or a partner's design on another vendor's paper) reads the corpus from DDN storage and decides retrieval on its own terms; that judgment belongs to whoever ships the stack, not to DDN. Decision authority: absent, the reading every gap layer carries when the vendor offers nothing and absorbs nothing (WEKA, Everpure, and NetApp at Layer 2C).
Named, not scored: native vector-database and Kafka interfaces ('next year', press coverage October 14, 2026, no date); the Infinia, Milvus, and NIM RAG demonstration (DDN blog, reference only); natural-language queries over a video corpus (a presenter's own build shown at AI Field Day 9, October 15, 2026, not a product). Public evidence that moves the cell: a retrieval product or a validated RAG design sold and supported on DDN's paper.
Move/transform data — ETL/ELT, lineage, cost-aware movement, KV cache tiering
DDN's storage behind NVIDIA's open transfer interface, selectable from Dynamo, vLLM, SGLang, or LMCache. Delegated: the engine configuration targets NIXL, which other backends implement, and the cached state is recomputable; the Infinia SDK beneath it is DDN's.
Movement and caching documented in the community Lustre manual. Retained: the policies and scripts the enterprise writes run on any Lustre.
Infinia's KV-cache path is a backend plugin for NIXL, the open-source transfer library under NVIDIA Dynamo, released in NIXL v1.3.0 (June 15, 2026) and bundled in the prebuilt wheels from v1.3.1 through v1.5.0 (September 30, 2026). The serving engines that call it (Dynamo, vLLM, SGLang, LMCache) decide what to offload and when; transfers move host or GPU memory pre-registered for zero-copy. DDN's GPU-initiated data access work with NVIDIA (August 4, 2026) and a BlueField-4 KV-cache offering ('later this year', June 2026 coverage) are future.
DDN moves data in two places. The first is the KV-cache tier the canonical purpose line names. DDN contributed an Infinia backend to NIXL, the open-source transfer library (the Infinia backend files carry Apache-2.0 headers, copyright NVIDIA and DataDirect Networks) NVIDIA Dynamo and the major serving engines use to move attention state between GPU memory and slower tiers. It was merged June 5, 2026, released in NIXL v1.3.0, and is still bundled in v1.5.0. Its public README documents the build against Infinia's async SDK, per-backend parameters that select an Infinia cluster, tenant, subtenant, and dataset, and host and GPU memory pre-registered with Infinia for zero-copy transfers. Using it takes a NIXL-enabled engine build, the engine's offload connector, and DDN's client packages on the serving nodes; a DMA-BUF RDMA option exists only on the repository's main branch, unreleased. That makes Infinia a selectable G4 tier for KV blocks: the engine evicts to it and reloads from it rather than recomputing. The second is what Lustre does inside EXAScaler, documented in the community Lustre manual: hierarchical storage management (HSM) archive, release, and restore through copytools; file-level redundancy mirrors across pools; lfs migrate between pools and targets; and the persistent client cache on compute nodes (DDN's EXAScaler CSI driver lists 'Exascaler Hot Nodes' as GA). Against the layer's function this is movement, not pipelines. DDN ships no transformation engine, no ETL or ELT, no lineage, and no documented replication or data mover for Infinia (a replication engine and data mover sit on a 'Coming (2.4)' list that no public release note confirms). Hot Pools, DDN's policy-driven flash-to-disk tiering built on file-level redundancy, is described in a DDN blog without documentation and is named, not scored. Calibration: WEKA, Everpure, and NetApp read moderate on tiering and replication without a general pipeline engine; VAST reads strong on DataEngine's arbitrary functions. DDN's KV-cache backend is the one documented, GA KV tier in the storage column (WEKA's Augmented Memory Grid and Everpure's KV Cache Accelerator are named, undocumented), but it is a backend the engine drives, a fixed-function slice of the layer under rule 4, so the cell holds moderate.
Little is borrowed from DDN here, and what is can be left. The KV-cache tier sits behind NIXL, an interface with in-tree backends for POSIX, S3-compatible object stores, Azure Blob, GPUDirect Storage, 3FS, Redis, and Mooncake beside Infinia; the engine configuration the enterprise accumulates selects a backend by name and parameters, so switching tiers is a configuration change and the cached blocks are recomputable state, not opinions: Delegated, a proprietary implementation (Infinia and its SDK) behind a multi-vendor interface. Lustre's HSM, mirroring, migration, and client caching are community code the enterprise can run on any Lustre: Retained. Neither half of the layer exercises DDN judgment at runtime. The Infinia backend stores and returns the blocks the serving engine hands it; which blocks to offload and evict is the engine's call (Dynamo's KV block manager, LMCache), scored on the rows that ship those engines. Lustre moves data between pools and archive when the enterprise's commands or policy tooling tell it to (lfs hsm_archive, hsm_release, hsm_restore, mirror resync, migrate); the scored surfaces carry no automatic tiering of their own, since Hot Pools is unscored. The engine executes what the enterprise configured: code decides, visible, overridable, Retained, the VAST DataEngine reading at this layer.
Named, not scored: Hot Pools and its NAND-accelerated variant (DDN blog; ISC press release June 22, 2026); 'Distributed KV Cache acceleration' as a DDN product across Infinia and EXAScaler (ISC press release June 22, 2026; the scored surface is the NIXL backend); Infinia replication engine and data mover ('Coming (2.4)', unconfirmed); Infinia-to-EXAScaler preload and return (described to press January 5, 2026, no product); event-driven data movement (Infinia 2.0 launch, February 2025); shared KV cache on Google Cloud Managed Lustre (Google blog design pattern, Google's paper); IndustrySync pipelines (90-day early-adopter program ahead of GA). Future: BlueField-4 KV-cache offering ('later this year'); GPU-initiated data access ('as development progresses'); native Kafka interface ('next year'). Public evidence that moves the cell: a documented pipeline engine that runs enterprise code, or documented Infinia replication plus Hot Pools.
GPU scheduling, quotas, RBAC, fair-share scheduling, utilization optimization
DDN's driver presenting EXAScaler through the Container Storage Interface. Delegated: the enterprise's claims and storage classes target a multi-vendor standard, the WEKA CSI and NetApp Trident reading; the license is DDN's.
DDN's driver presenting Infinia block volumes through the Container Storage Interface, documented only to datasheet depth. Delegated, the same CSI reading.
DDN's own control surfaces and templates for standing up and running Infinia. Ceded, the WEKA Operator and Everpure Fusion reading.
Neither the CSI driver nor Infinia's cluster management depends on NVIDIA software; GPU scheduling on DDN-fed clusters belongs to whatever scheduler the enterprise runs (Kubernetes, Slurm, Run:ai).
DDN orchestrates its own storage on the enterprise's orchestrators and nothing above it. The EXAScaler CSI driver (v2.10.0, October 1, 2026, public on GitHub) presents EXAScaler to Kubernetes through the Container Storage Interface: static and dynamic provisioning, read-write and read-only modes, volume expansion, topology-aware placement, snapshots, compression, encryption, per-volume quota opt-out, and automated Lustre client-module installation, all marked GA in its feature matrix, with OpenShift certification listed for 4.13 to 4.15 and 4.19 to 4.22 (OpenShift needs the Kernel Module Management operator). The driver is source-visible but licensed under DDN's end-user license agreement, not an open-source license. Infinia manages itself as a cluster through a web interface, the redcli command line, and a REST API, and DDN's public GCP guide automates deployment from Google Cloud Marketplace through a Terraform configuration (the Deployment Manager API is a prerequisite), provisioning the number of VMs the requested capacity needs in the zone the enterprise picks. Infinia also presents block volumes to Kubernetes through its own CSI driver (block.csi.red.ddn.com, version 1.9.0 in the Kubernetes CSI driver registry; DDN's Infinia datasheet and solution brief state CSI support), with no public driver documentation beyond that. The CSI half carries the grade: without it, what remains is a storage system managing itself. That is the storage column's Layer 2A: WEKA (Operator plus CSI), Everpure (Fusion plus Portworx and CSI), and NetApp (Console plus Trident) read moderate on running their own storage under the enterprise's orchestrators, and none schedules GPUs. DDN's version is thinner: no Kubernetes operator, no fleet control plane with public documentation. The GPU-scheduling claim belongs to DDN Horizon, announced March 16, 2026 as an 'AI orchestration control plane' allocating GPUs and CPUs across tenants with quality of service, self-service provisioning, and chargeback, with no documentation, datasheet, or availability statement; it is named, not scored.
Mostly the enterprise's own. The claims, storage classes, and topology rules the enterprise writes target the CSI standard, which every storage vendor implements: Delegated, the WEKA CSI, NetApp Trident, and Everpure PX-CSI reading, whatever the driver's own license. The Infinia CSI driver reads the same way. Infinia's cluster management, redcli, REST API, and the Marketplace Terraform deployment are DDN's with no second implementation: Ceded, the WEKA Operator and Everpure Fusion reading. The runtime call that carries the grade is where Kubernetes volumes land: the driver provisions and the scheduler places, inside the allowedTopologies and zone constraints the enterprise writes into each StorageClass, a persisted per-object placement pin: vendor decides, visible, overridable, Delegated, the WEKA nodeSelector reading. Infinia's own half reads Ceded: its IO engine chooses the erasure-code layout in real time ('erasure coding levels are not set at cluster build time', the Infinia datasheet), with no documented per-object override.
Named, not scored: DDN Horizon (launched March 16, 2026; product page and press release; GPU and CPU allocation across tenants with QoS, self-service S3 and Lustre provisioning, notebooks, pipelines, NIM endpoints, chargeback, on-prem and air-gapped; no documentation, no named scheduler, no availability date); the EXAScaler Management Framework (product page only); DDN Insight (SC25 announcement); lfsccm, a Slurm burst-buffer plugin for Lustre's persistent client cache (a GitHub utility). Public evidence that moves the cell: Horizon documentation (a named scheduler, quotas, fair-share) would argue for a GPU-orchestration component and test strong against the peers; its absence leaves the cell at the thin end of moderate.
Model serving, agent execution, inference APIs, distributed inference
Serving on DDN storage runs on engines DDN doesn't ship: NVIDIA Dynamo, TensorRT-LLM, and NIM, or vLLM and SGLang, with LMCache as a KV-cache extension to those engines. DDN's contribution is the NIXL backend those engines call, scored at Layer 1C. The Enterprise AI HyperPOD bundles NVIDIA AI Enterprise on Supermicro's paper.
DDN serves no models and runs no agents. Its inference story is real but sits below the runtime: the Infinia backend for NIXL lets Dynamo, vLLM, SGLang, and LMCache offload KV cache to Infinia instead of recomputing it (scored at Layer 1C, the layer whose purpose line names KV cache tiering), and both stores hold the model weights engines load and reload. The engine choosing what to serve, batch, route, and evict is NVIDIA's or the open-source community's, chosen and operated by the enterprise. DDN Inferno, announced March 18, 2025 as a BlueField-based inference-acceleration appliance, never received a configuration, datasheet, or availability statement. DDN Horizon's product page claims NVIDIA NIM inference endpoints deployed from a self-service catalog (the March 16, 2026 launch release says only inference services), with no documentation or availability statement. The Enterprise AI HyperPOD carries NVIDIA AI Enterprise (NIM, NeMo Retriever), but it is Supermicro's paper and scored on Supermicro's row. Calibration: NetApp reads gap on the same shape (the runtime is NVIDIA's, co-sold through the channel); WEKA and Everpure read moderate on sealed RAG appliances that generate answers on their own paper (the appliance RAG ruling); DDN ships no such appliance, so the layer is gap.
Nothing to borrow from DDN at this layer. The enterprise runs the serving engine and the agent loop on its own choice of software, and the runtime judgment (which model answers, how requests batch, which KV blocks to keep) belongs to that software and to the enterprise's configuration of it. DDN absorbs nothing beneath a service of its own. Decision authority: absent, the WEKA and Everpure gap-layer reading at Layer 2C.
Named, not scored: DDN Inferno (press release March 18, 2025; no configuration or availability ever published); DDN Horizon's NIM endpoint deployment (product page; launched March 16, 2026; no documentation); the Enterprise AI HyperPOD's NVIDIA AI Enterprise stack (Supermicro's paper). Public evidence that moves the cell: a serving or agent runtime, or a generating appliance, sold and supported on DDN's paper with documentation.
Policy-driven placement and resource coordination — the Autonomy Layer
DDN's Computex 2026 announcement (GTC Taipei, June 1, 2026) aligns its platform with NVIDIA Vera BlueField-4 STX and the DOCA security framework. Its companion post assigns runtime visibility into agent behavior to DOCA Argus, tenant and agent isolation at line rate to DOCA Flow, and enforcement to BlueField-4 silicon, with 'DDN's focus' the data layer; it says DDN looks forward to joint deployments 'as they become available'. Unscored: no GA documentation, and the acting components are NVIDIA's.
DDN offers nothing that acts on agents. Infinia's tenants, subtenants, and role-based access and Lustre's nodemaps decide which identity reaches which data, which is data governance at Layer 1A; nothing gives an agent an identity, gates its requests, registers it, orchestrates it across other agents, or places its inference by policy. DDN's Computex 2026 language about governance for agentic AI attaches to storage multi-tenancy and observability. DDN Horizon's tenant isolation, policy enforcement, and chargeback (the March 16, 2026 release) and 'automated QoS' for GPU allocation (its product page) would govern infrastructure tenants, not agents, and it is undocumented. DDN's Computex 2026 security story puts agent-behavior visibility and per-access enforcement in NVIDIA's BlueField-4 and DOCA, with DDN supplying the data layer beneath it. The DDN presenter at AI Field Day 9 framed KV cache placement as a multivariate cost decision (retrieve or recompute, by reuse probability and movement cost), which is the shape of Infrastructure-2C, but the decision is made by the serving engine and no DDN surface exposes it as policy. Calibration: WEKA, Everpure, VAST, and NetApp all read gap at 2C on the same ground (data guardrails are not agent governance), and the live-placement gap is universal across the instrument.
Nothing offered as a plane over agents. The enterprise that wants agent identity, a gateway, a registry, or orchestration brings its own; DDN's permissions travel with the data, not with the agent. Absent.
Named, not scored: DDN Horizon's tenant controls (March 16, 2026; undocumented, and infrastructure tenancy rather than agent governance even if documented). Public evidence that moves the cell: a DDN component that authorizes, blocks, routes, or orchestrates an agent's action at runtime.
AI-powered business capabilities — business logic, workflow automation
DDN ships no applications; the business capabilities built on DDN storage are the enterprise's or its ISVs'.
DDN sells no business application. IndustrySync, pre-integrated industry pipelines for financial services and life sciences on the Enterprise AI HyperPOD, is a 90-day early-adopter program ahead of general availability, which keeps it unscored under the GA-gate; if it ships, the question is whether it carries DDN-supplied business logic (the financial-services pipeline claims continuous expected-shortfall calculation) or only accelerates the customer's own models and workflows. Tintri, DDN's subsidiary, sells VMstore storage for virtual machines, not an AI application. The video-corpus search a DDN presenter built over AI Field Day recordings is a demonstration. Calibration: WEKA, Everpure, and NetApp read gap at Layer 3 on the same shape: storage beneath others' applications.
Nothing offered, nothing inherited. Absent.
Named, not scored: IndustrySync (early-adopter program ahead of GA; on shipping, judge function fit against the value plane). The DDN and Aleria Sovereign AI Factory (March 16, 2026 release: 'available now as a reference architecture', 'one SKU, one subscription, purchased directly from DDN', with Aleria's Legal AI, HR Intelligence, Finance AI, Board Advisory, and compliance applications on DDN storage and NVIDIA Vera Rubin): a press-release availability statement with no DDN documentation, price list, or ordering guide, so named, not scored; DDN's Aleria success story (about May 2026) describes First Abu Dhabi Bank and Aldar running Aleria on DDN without stating that DDN sold it. Public evidence that moves the cell: a first-party business application on DDN's paper.
Public surfaces only; DDN's Infinia documentation (partner-docs.ddn.io) requires a partner or customer login beyond its version index (2.3.0 through 2.4.1) and the Google Cloud and OCI deployment guides, which are cited, and gated pages are not. Infinia GCP and OCI deployment guides (v2.4.1), the Infinia datasheet and solution brief, and the Infinia software subscription brief; ai-dynamo/nixl (the Infinia backend plugin, README, and release notes v1.3.0 through v1.5.0); DDNStorage/exa-csi-driver (v2.10.0 README, feature matrix, OpenShift certification, license); the community Lustre Operations Manual; DDN's NVIDIA reference architecture index; Supermicro's DDN solution page; DDN product pages (AI400X3, AI400X3M, EXAScaler, Infinia, Enterprise AI HyperPOD, Horizon, IndustrySync) and press releases (AI400X3, ISC 2025; Inferno, March 18, 2025; HyperPOD, October 28, 2025; Horizon and the DSX Factory with the Aleria Sovereign AI Factory, March 16, 2026; Computex, June 1, 2026; ISC 2026, June 22, 2026; Infinia 2.4, July 7, 2026; GPU-initiated data access, August 4, 2026); DDN blogs on Infinia, Hot Pools, KV cache management, RAG, and the data path as a security control plane; the Aleria success story; Google Cloud Managed Lustre documentation; Blocks & Files, iTWire, and StorageNewsletter coverage (pointers only). AI Field Day 9 presentation (October 15, 2026) used as a pointer to public sources, not as evidence. Each cell was attacked by an outside reviewer and every claim ruled before lock; ledgers in reviews/ddn-*-chatgpt.json.