Start with the density math. One 1U tray: up to 9.83PB raw, 390GB/s of writes and 15M 4K random-read IOPS — the ratings Hammerspace brought to SC26 (2026-11-16 to 2026-11-19, Chicago) for its Open Flash Platform (OFP). Scale it to a full rack: 42U of OFP holds about 413PB and draws up to 75.6kW (calculated from vendor specs), or roughly 183W per raw petabyte. For drive density, Hammerspace claims 3.3 times a typical 2U 24-bay NVMe server: 160 SSDs in 4U against 48. The metric that matters in AI storage is shifting from "how fast is one array" to "how fast per U, and how efficient per PB."
The launch vehicle is the Hammerspace Data System: the parallel global file system specialist will, for the first time, sell software, metadata nodes and storage hardware as one complete system with unified support, licensing software by capacity under management. Software-only deployments, Ready Nodes from Dell/Supermicro/Hitachi Vantara, and marketplace images on AWS/Azure/Google Cloud/Oracle Cloud all continue. OFP is the new hardware inside it: available to quote now, shipping in 2027 H1, with pricing unpublished.
Architecture: the x86 storage server disappears, the DPU is the node

The building block is a sled: 1/5 U wide, 8 U.2/U.3 NVMe SSDs in two removable 4-drive packs, 2 400GbE QSFP-DD ports, and — where an x86 motherboard would sit — a 64-core Arm DPU, the Xsight Labs E1. Per StorageReview's on-site reporting, the E1 uses TSMC's 5nm process, 64 Arm Neoverse N2 cores, 2 400G MACs, PCIe Gen5 and typically 70W, running standard Linux; each sled appears to the Hammerspace file system as an NFS server. Xsight's official E1 product brief likewise confirms 64 N2 cores and 800Gbps of networking (the vendor site blocked our access; see the notes at the end).
A 1U tray holds 1 to 5 sleds: up to 40 SSDs, 320 Arm cores and 10 400GbE ports when full. Each sled tops out at 360W, the tray at 1800W, fed by dual 2160W Titanium power supplies; sleds and drive packs are tool-less to service, managed via IPMI/Redfish. Two hard constraints: the tray is 106.64cm deep and requires a 1200mm rack; cooling is air, with a 40°C maximum inlet temperature at sea level.
Capacity and performance ratings (vendor figures, Gen5 TLC SSDs, 400GbE, NFSv4.2+pNFS over TCP):
| Item | Per sled | Per 1U tray (5 sleds) |
|---|---|---|
| Raw capacity (8×245.76TB QLC, KIOXIA LC9 class) | 1.97PB | 9.83PB |
| Read throughput | 80GB/s | 400GB/s |
| Write throughput (no mirror) | 78GB/s | 390GB/s |
| Write throughput (2-way mirror) | 39GB/s | 195GB/s |
| Write throughput (3-way mirror) | 26GB/s | 130GB/s |
| 4K random read | 3M IOPS | 15M IOPS |
Protection is mirroring, set per file or directory by policy and changeable online: 2-way mirroring halves both write bandwidth and usable capacity (about 4.9PB usable on a full QLC tray), and 3-way mirroring leaves just 130GB/s of writes. Hammerspace suggests skipping mirrors for regenerable data such as KV cache. Unpack the read ceiling: 80GB/s per sled pushes the 2 400GbE ports close to line rate (about 40GB/s per port, essentially saturated once protocol overhead is counted) — the network ports, not the flash, set the read ceiling.
One discrepancy must be flagged: the official press release (2026-11-17) quotes 250GB/s read and 350GB/s write per U, claiming 37% greater read-throughput density than the alternatives compared, up to 4.4 times the write-throughput density, and 2 times the SSD density of the densest server configuration evaluated; the full-tray ratings StorageReview obtained at SC26 are 400GB/s read and 390GB/s write, alongside a claim of 2 to 5 times the per-U performance of an unnamed nearest competitor. The two sets of numbers use different comparison baselines and configurations — always cite them with their conditions, never mix them.
Route comparison: GDS owns the "last meter", OFP owns the supply chain

NVIDIA GPUDirect Storage (GDS) solves an intra-node problem: direct DMA from NVMe to GPU memory, bypassing the CPU bounce buffer, requiring matching drivers and a supported file system. Hammerspace is not on that road — it insists on the standard NFSv4.2 client (built into Linux distributions, no proprietary agent) and spreads the data path across nodes with the pNFS Flex Files layout. The two are less rivals than a division of labor: GDS optimizes the "last meter" inside a node, while OFP and Tier 0 optimize cluster-level data supply; putting the DPU in the storage node frees x86 cores from the data path.
Tier 0 is the more intriguing design in this launch: export the local NVMe already inside GPU servers over NFS and aggregate it into a cluster-wide shared tier, with no client installed on compute nodes; checkpoints are written locally first, then sink to object storage by policy. Hammerspace claims Tier 0 delivers 4.4 times the sequential bandwidth and 56% lower latency versus external shared storage (vendor figures, test configuration unpublished). The inference side works the same way: KV cache lives on the pooled Tier 0, and the inference software's own cache manager reloads already-computed context from shared POSIX storage instead of recomputing it on the GPU; Hammerspace does not act as the cache manager itself. Early hyperscale customers place 1 to 2 OFP trays at the top of each GPU rack, valuing rack and performance density over power savings; in NVIDIA's tier labels, the NVMe inside GPU servers is G3, and one TLC-filled OFP tray per rack is G4. This DPU-storage-node road also carries NVIDIA's BlueField-4 STX reference architecture introduced at GTC; Hammerspace's difference is that its file system also runs without its own hardware.
The real bottleneck: data supply, not capacity

"Simply adding more storage isn't the answer." Founder and CEO David Flynn frames the launch as the consolidation AI factories brought to compute — now applied to data. The Data System's skeleton is still the separation of metadata from the data path: each site runs a high-availability pair of Anvil metadata nodes (1U x86, M-Series, 2 100GbE ports) for the control plane; metadata replicates continuously across sites with about 1 second of lag to distant locations, while the data itself moves only when a policy or an access request triggers it.
The new data intelligence layer does three things: catalog file and object metadata across every system without moving data; DeepScan opens files on demand or on a schedule to find personal, financial and protected health information, turning findings into searchable tags; and natural-language queries through a customer-chosen LLM feed results straight into the policy engine — declarative rules such as "move data untouched for 30 days" or "keep 3 copies at 3 sites", with tags inheriting down the directory tree and following files across systems and clouds.
Reference customer kaiko (a clinical AI company in Amsterdam/Zurich) runs Azure, Nebul and Nebius — three clouds — as one namespace, growing from about 128 GPUs to 1000+ on Nebius, using GPU-local NVMe as Tier 0 for training and placing checkpoints on Azure Blob by policy — note that this case does not involve OFP hardware. Another, unnamed hyperscale AI lab replaced FSx for Lustre on AWS with a 30PB Hammerspace cluster (after a year in which Lustre was its largest source of support tickets), reports large-file operations about 3 times faster (vendor figure), and is extending the namespace to an on-premises data center.
What this means for Pearl River Delta AI machine rooms
Drop OFP into the Guangzhou/Shenzhen AI data center context, and three things decide success earlier than rated bandwidth. First, racks: the 106.64cm depth and 1200mm rack requirement rule out many shallow cabinets — measure before you retrofit. Second, power and cooling: 1800W per tray means 1.8kW per U, and a full rack's 75.6kW far exceeds what most existing machine rooms can feed a single rack (calculated comparison); the 40°C air-cooling inlet ceiling needs measured cold-aisle temperatures in Guangdong summers before anyone commits. Third, the timeline: the OFP initiative started in 2025 (with Xsight Labs and Los Alamos National Laboratory among the participants), 2025-10 promised production systems in early 2026, and the plan is now 2027 H1 — a slip of about 1 year, so leave margin in procurement schedules.
Watch the evidence gaps, too: Hammerspace has not published OFP latency, client counts or IO500/MLPerf Storage results, and StorageReview did not run or audit these tests. For teams planning GPU clusters, the questions worth putting to suppliers in writing: real write bandwidth under your mirroring policy, 400GbE fabric behavior under congestion, how existing NetApp/PowerScale/object storage is taken over, and a re-measured "watts per PB" on your actual workload — not just the single number 390GB/s.
Sources & verification notes (accessed 2026-11-22)
- StorageReview (Brian Beeler, 2026-11-18): https://www.storagereview.com/news/hammerspace-data-system-open-flash-platform-ofp-dpu-flash-xsight-e1 (fetched and verified directly)
- Hammerspace official press release (2026-11-17): https://hammerspace.com/hammerspace-unveils-data-system/ (JS-rendered page; body verified via the StorageNewsletter syndication: https://www.storagenewsletter.com/2026/11/17/sc26-hammerspace-unveils-the-data-system-for-the-ai-era/ )
- Xsight Labs official site: https://www.xsightlabs.com/ (returned a 429 bot checkpoint at access time, not directly fetchable; the E1's 64 N2 cores/800Gbps cross-verified via a search snapshot of the official E1 product brief (https://xsightlabs.com/wp-content/uploads/2024/10/E1-PB-Public.pdf ) and Tech Field Day coverage (https://techfieldday.com/video/the-architecture-argument-with-xsight-labs/ ); 5nm/70W appears only in StorageReview's reporting and is cited as media-sourced)
- Calculated: 42U≈413PB, 75.6kW, 183W/PB, about 40GB/s per port, 1.8kW per U, the full-rack power comparison. Vendor ratings/unaudited: 390/400GB/s, 15M/3M IOPS, 1.97/9.83PB, mirror derating and 4.9PB, 4.4x/56% (Tier 0), 3x (FSx replacement), 3.3x density, the 2 to 5x competitor claim, and the press release's 250/350GB/s with its 37%/4.4x/2x comparison basis. SC26 booth 4238. Shipping in 2027 H1 is a vendor plan.
