Hammerspace is moving from selling its software-defined data platform alone to offering a complete Hammerspace Data System that combines software, metadata servers and storage. The launch includes the Open Flash Platform (OFP), a 1U flash tray built around data processing units (DPUs) rather than conventional x86 storage servers. Hammerspace says a fully populated tray can deliver up to 390GB/s of writes and 15 million 4K random-read IOPS, with raw capacity of up to 9.83PB.
OFP is available for quotation and is scheduled to ship in the first half of 2027. Hammerspace also introduced a data-intelligence layer and described a role for its shared-storage architecture in providing access to GPU-local NVMe used for inference KV cache. The company has not published pricing for the hardware.
A packaged system around a global namespace
Hammerspace’s software creates a parallel global file system that presents existing NAS, object and cloud storage through one namespace. It separates metadata from the data path and supports pNFS, NFS, SMB and S3. Linux clients can connect through the standard NFSv4.2 client, without a proprietary agent, using the Flex Files layout that Hammerspace says it developed for the NFS specification.
The new system groups the offering around data unification, intelligence, orchestration, acceleration and storage. The change is primarily commercial as well as technical: customers can buy an integrated system from Hammerspace, while software-only deployments on customer hardware remain available. Hammerspace Ready Nodes from Dell, Supermicro and Hitachi Vantara, along with marketplace images for AWS, Azure, Google Cloud and Oracle Cloud, also continue.
System deployments pair Anvil metadata nodes—1U x86 servers available in Standard and Large M-Series configurations—with data-services nodes. Anvils are installed as a high-availability pair at each site and run the control plane for the namespace and placement policies. Sites can join a global file system: metadata replicates continuously, while data moves when a policy or access request requires it. Hammerspace says job schedulers can initiate movement through APIs and native connectors.
DPU-based flash density and performance
Each OFP sled contains eight U.2 or U.3 NVMe SSDs, arranged in two removable four-drive packs, and an Xsight Labs E1 DPU. The E1 has 64 Arm Neoverse N2 cores, two 400GbE ports, PCIe Gen5 connectivity and a typical power figure of 70W, according to the supplied specifications. Each sled runs standard Linux and appears to the Hammerspace file system as an NFS server.
A tray can hold one to five sleds, for a maximum of 40 drives, 320 Arm cores and ten 400GbE ports. Hammerspace rates maximum power at 360W per sled and 1,800W per tray. The system is 106.64cm deep and requires a 1,200mm rack. Its removable sleds and drive packs are designed for tool-less service; management supports IPMI and Redfish.
The capacity ceiling assumes eight 245.76TB QLC SSDs in each sled. On that basis, Hammerspace lists 1.97PB raw per sled and 9.83PB per tray. Customers can also use TLC drives, which the company positions for performance, or QLC for capacity. Its performance figures are company ratings based on NFSv4.2 with pNFS over TCP, a 400GbE fabric and enterprise Gen5 TLC SSDs. Hammerspace lists up to 80GB/s reads per sled, or 400GB/s per tray, and 78GB/s writes per sled without mirroring.
Protection is configured by file or directory policy. Two-way mirroring reduces the listed tray write rate to 195GB/s and three-way mirroring to 130GB/s; two-way mirroring also cuts usable capacity roughly in half. Hammerspace suggests leaving regenerable data, such as KV cache, unmirrored. The company has not published latency, client-count, IO500 or MLPerf Storage results for OFP, and the stated figures have not been independently audited.
Cataloging and GPU data workflows
The data-intelligence layer catalogs file and object metadata across systems in the namespace without relocating the data. An optional DeepScan capability examines file contents for personal, financial and protected health information, either on demand or on a schedule, and makes findings searchable as tags. Administrators can query using natural language through an LLM of their choice. Results can inform Hammerspace’s declarative placement rules, called objectives, which can be applied to shares, volumes, directories or individual files.
Hammerspace also positions NVMe inside GPU servers as a Tier 0 shared tier, with OFP behind it as Tier 1. The system exposes pooled GPU-local storage over NFS; inference software’s own cache manager can use it for KV cache. Hammerspace does not manage that cache. The company says the approach can let inference servers reload computed context from shared storage instead of recomputing tokens, but did not provide a test configuration for its claims about Tier 0 performance.
Hammerspace cited kaiko, a clinical AI company with offices in Amsterdam and Zurich, as a Data System customer. Kaiko uses a namespace spanning Microsoft Azure and the Nebul and Nebius GPU clouds, with GPU-node NVMe for training and checkpoints placed on Azure Blob Storage by policy. The case study does not include OFP hardware. Hammerspace says the new tray will be demonstrated at SC26; broader shipment is planned for the first half of 2027.