6 segments · 58 companies · as of Sep 18, 2026
Storage Systems
Systems and software that keep AI data accessible, fast and recoverable
Where it sits
Fed by: Chip ManufacturingFeeds: Servers & Systems
Storage systems organize drives into reliable pools and move data to GPU clusters. Enterprise arrays, parallel file and object platforms, public-cloud storage, backup software and storage networks solve different parts of the data pipeline.
Why now
Long-context inference adds a flash or CXL tier for KV-cache reuse. STX/CMX announcements and early CXL systems are ahead of broad deployment; arrays, object storage and recovery services already earn revenue. Lenovo completed Infinidat and Everpure now trades as P.
Enterprise storage arrays
Enterprise arrays provide durable block and file storage with availability, replication and recovery controls. AI adds capacity and throughput demand, but most vendors also serve broad enterprise workloads; listed-parent ownership determines each investable entry.
PowerStore, PowerMax and PowerFlex provide enterprise block storage, replication and software-defined capacity
leaderOceanStor Dorado all-flash arrays provide enterprise data services and storage for AI data pipelines
leaderAFF/ASA all-flash arrays and ONTAP data services provide enterprise block and file storage
majorFlashArray all-flash block/file storage with Evergreen subscriptions; renamed from Pure Storage in 2026
majorAlletra Storage MP B10000 delivers disaggregated enterprise block/file storage with GreenLake management
majorFlashSystem arrays use FlashCore modules and Storage Virtualize for enterprise block storage and resilience
majorVia Hitachi Vantara, VSP One Block High End provides all-flash NVMe storage for enterprise and AI workloads
majorThinkSystem storage plus Infinidat InfiniBox/InfiniVerse after the acquisition closed on April 9, 2026
majorControlled H3C supplies UniStor all-flash arrays and ONEStor distributed storage; HPE exited its remaining stake in May
majorAS-series all-flash and hybrid arrays serve Chinese enterprise storage and AI infrastructure
majorEnterprise arrays and ParaStor distributed storage provide shared data capacity for Chinese HPC and AI deployments
majorFsas Technologies ETERNUS EP300 and DX/AF families provide enterprise all-flash and hybrid storage
majoriStorage enterprise block and file systems support business and AI data infrastructure in Japan
nicheXEDP unifies block, file and S3 storage on standard servers; XEUS supplies packaged enterprise storage appliances
challenger
Sources: idc.com, news.lenovo.com, everpuredata.com, h3c.com, xsky.com, netapp.com
AI file & object platforms
Parallel file systems and flash object stores feed GPUs, checkpoint models and share large unstructured datasets. Independent storage platforms compete with established system vendors on sustained throughput, metadata handling, data access and deployment flexibility.
DASE shared flash file/object storage disaggregates compute from data and feeds GPU clusters through NFS, SMB and S3
leaderEXAScaler parallel file and Infinia object platforms serve AI/HPC; AI400X3M appliance GA targeted for end-Q3 2026
leaderNeuralMesh parallel file/object storage serves GPU workloads; NeuralMesh 6 GA was announced for H2 2026
majorGlobal file namespace and data orchestration; NVIDIA-reference AI Data Platform became available in March 2026
majorFlashBlade EXA scale-out flash file/object storage supports large AI training and inference datasets
majorStorage Scale and Scale System provide parallel file storage, including NVIDIA AI infrastructure deployments
majorPowerScale file and ObjectScale object storage supply AI datasets, checkpoints and shared inference data
majorAFX disaggregated all-flash scale-out storage extends ONTAP into high-throughput enterprise AI data pipelines
majorAlletra X10000 object storage and Storage for AI target GPU data ingest and shared enterprise AI datasets
majorOceanStor A-series distributed file and Pacific object storage support Chinese AI training and inference
majorScale-out file storage and Cloud Native Qumulo span on-premises, AWS, Azure and Google Cloud deployments
challengerParallel file storage software, formerly Panasas, pools flash and disk for AI and HPC workflows
challengerMyriad flash file storage and ActiveScale object storage serve unstructured data and AI pipelines
nicheSoftware-defined parallel file/object storage with GPU-converged deployments for training and inference
niche
Sources: vastdata.com, weka.io, ddn.com, hammerspace.com, netapp.com, ibm.com
Cloud & software-defined storage
Cloud object services, software-defined storage and data-tiering tools hold training corpora, model files and cold data. This segment covers actual storage products rather than warehouses, databases or analytics applications, which belong to a separate theme.
AWS S3 object storage, S3 Express and FSx file systems hold AI datasets, checkpoints and training data
leaderAzure Blob, Azure Files and Azure NetApp Files provide object/file storage for AI pipelines
leaderGoogle Cloud Storage and Rapid Buckets serve AI training/checkpoints; Rapid Buckets became GA in April 2026
majorVia Alibaba Cloud, OSS supplies object storage and retention controls; AI VibeFlow remains whitelist-access in 2026
majorOCI Object Storage, Block Volumes and managed Lustre/File Storage serve AI datasets, checkpoints and shared files
majorCommercial AIStor supplies S3-compatible object storage for AI data lakes with enterprise support and erasure coding
majorHyperStore 8.2.6 offers NVIDIA-certified S3-compatible object storage for enterprise AI data
majorRING distributed object storage and ADI manage large datasets; ARTESCA focuses on immutable backup storage
majorUnified Storage combines file, object and block services across private-cloud and hybrid deployments
majorDisaggregated software-defined NVMe/TCP block storage runs on standard Ethernet and commodity servers
challengerB2 and B2 Overdrive provide S3-compatible object storage for AI datasets, backup and cloud workflows
challengerHot Cloud Storage provides S3-compatible capacity for AI data and backup; 2026 Lyve Cloud acquisition expands its footprint
challengerR2 S3-compatible object storage holds training data, model weights and checkpoints with no egress charges
nicheUnstructured-data analysis, transparent tiering and migration move cold files from NAS to object/cloud storage
nicheXEDP software pools commodity servers into a shared block, file and object-storage platform for private clouds
challengerManaged Iceberg storage became GA in June on AWS/Azure; Horizon exposes stored open-table files to external engines
niche
Sources: cloud.google.com, min.io, scality.com, cloudian.com, docs.snowflake.com, xsky.com
Backup & cyber recovery
AI datasets and model artifacts still need recovery copies, retention policies and isolated restoration paths. Backup software and purpose-built storage protect production data against deletion and ransomware; this is distinct from general cybersecurity or AI applications.
Data Platform v13.1 and Data Cloud Vault provide backup, immutable cloud copies and archive retention
leaderDataProtect and acquired Veritas NetBackup protect hybrid workloads with backup and cyber-recovery controls
leaderCloud platform unifies backup and cyber recovery; Cleanroom Recovery restores data into isolated environments
majorSecurity Cloud provides immutable backup, threat analysis and recovery across enterprise and cloud workloads
majorManaged SaaS backup stores isolated immutable copies of cloud, SaaS and data-center workloads
challengerR-Cloud protects SaaS, public-cloud and on-premises workloads with application-aware backup and recovery
challengerPowerProtect Data Domain and Cyber Recovery supply deduplicated backup appliances and isolated recovery vaults
majorStoreOnce backup appliances and GreenLake recovery services provide deduplicated protection for enterprise data
majorDXi backup appliances and Active Vault tape isolation protect secondary copies and recovery workflows
nicheARTESCA immutable S3 storage supplies protected backup targets for Veeam and other backup platforms
nicheiStorage HS deduplicated backup systems; September 2026 new models are scheduled to ship in October
nicheData Protector backs up enterprise workloads to disk, tape and cloud; Cloud Workloads extends virtual/cloud recovery
major
Sources: veeam.com, cohesity.com, rubrik.com, commvault.com, druva.com, hycu.com
Storage networks & servers
Fibre Channel and Ethernet fabrics connect storage to compute, while storage servers and JBOF/JBOD enclosures package drives at scale. DPU-based designs move protocol handling off host CPUs; prototype STX systems are distinguished from shipping platforms.
Brocade 128G G820 Fibre Channel switches and Emulex HBAs connect enterprise storage fabrics
leaderMDS 9000 64G Fibre Channel SAN switches support all-flash arrays and NVMe over Fibre Channel
majorQLogic 2870 64G Fibre Channel HBAs support FCP and NVMe/FC; QLogic is distinct from Broadcom Emulex
majorBlueField DPUs offload storage protocols; Spectrum-X and ConnectX carry RDMA traffic for GPU storage
majorEthernet switches provide loss-aware RDMA storage fabrics linking shared flash systems and GPU clusters
majorPetascale BlueField-3 JBOFs and dense storage servers; BlueField-4 CMX server shown as a March 2026 prototype
majorStorage server and SD6300 dense SAS-4 JBOD platforms support hyperscale and enterprise storage deployments
majorSV7000G4/ST7000G4 storage servers/JBODs and ST7200/ST5100 all-flash NVMe JBOFs serve cloud operators
majorQCT storage platforms and 24Gb SAS JBOD designs provide high-density drive infrastructure
majorF2026 Gen5 NVMe JBOF, HA2026 dual-node storage and high-density SAS JBOD enclosures
nicheViking Edge AI computational-storage appliances place processing near stored data for edge pipelines
niche
Sources: docs.broadcom.com, cisco.com, marvell.com, supermicro.com, celestica.com, wiwynn.com
Inference context storage
Long-context inference can offload and reuse KV cache beyond GPU HBM. CXL memory appliances, shared flash software and emerging DPU storage tiers solve different latency and capacity problems; published benchmarks do not establish universal performance gains.
STX/CMX BlueField-4 context tier plus Dynamo/NIXL orchestrate KV-cache movement; partner systems target H2 2026
leaderCommercial Augmented Memory Grid offloads persistent KV cache to NeuralMesh flash through RDMA and GPUDirect
leaderMemoryAI production-ready KV-cache server combines 3TB main DRAM and up to 8TB CXL memory
challengerDistributed KV-cache acceleration launched in June 2026 across Infinia/EXAScaler with NVIDIA Dynamo integration
majorShared flash stores reusable KV-cache blocks through Dynamo integration; STX/CMX adds a planned context tier
majorFlashBlade EXA is being aligned with NVIDIA STX for context storage; no blanket GA claim for that integration
challengerObjectScale NIXL object backend offloads KV cache; integration is previewed in 4.3, with GA planned for 4.4
challengerSupremeRAID KV-cache solutions protect and accelerate NVMe flash offload across local and rack deployments
nicheOceanStor AI Data Platform pairs a KV-cache store with Unified Cache Manager for inference context reuse
majorPliops assets acquired Feb 2026: XDP LightningAI/XDP-PRO tier offloads GPU KV cache to petabyte NVMe via GPU-initiated RDMA
emerging
Sources: nvidianews.nvidia.com, ir.penguinsolutions.com, weka.io, ddn.com, vastdata.com, graidtech.com
Glossary
- Object storage
- Stores data as objects with identifiers and metadata, usually accessed through an S3-compatible API.
- KV cache
- Stored attention keys and values let inference reuse earlier token computations.
- NVMe-oF
- NVMe over Fabrics connects remote flash storage across a network rather than a local PCIe bus.
Heat, spend, exposure and shares are editorial estimates. How to read the map ↗