Research Notes

Is the Hammerspace Data System What Enterprise AI Pipelines Have Been Missing?

Research Finder

Find by Keyword

Is the Hammerspace Data System What Enterprise AI Pipelines Have Been Missing?

Hammerspace adds purpose-built flash to its AI Data Platform while giving enterprises a way to put existing data and infrastructure to work.

11/18/2026

Key Highlights

  • Hammerspace introduced Data System: an integrated system for AI and HPC alongside its existing software deployment options.
  • New data intelligence tools let administrators catalog, query, and act on data across the global namespace.
  • Hammerspace Open Flash Platform (OFP) adds a persistent flash tier alongside the shared Tier 0 NVMe capacity inside GPU servers.
  • The company will sell the software and provide first-call support. Partners handle hardware fulfillment.

The News

At Supercomputing 2026 (SC26) in Chicago, Hammerspace announced the Hammerspace Data System. The company is adding data intelligence capabilities and its own flash storage subsystem to the software platform it uses to manage data across existing infrastructure. Hammerspace said quoting would begin November 17, with general availability planned for the first half of 2027. For more information, see the official company press release.

Analyst Take

Hammerspace CEO David Flynn’s work at Linux Networx and Fusion-io helps explain how the team arrived at this design. At Linux Networx, Flynn designed supercomputers using Linux clusters, and as CEO at Fusion-io, he helped bring flash closer to applications by putting it inside the server. Flynn co-founded Hammerspace in 2018 to tackle a related problem. When access to data depends on the storage system, serving a new workload may involve moving the data first. Hammerspace is built on Linux and parallel NFS to separate file metadata from the underlying storage and give applications a common view of data across systems.

The OFP applies that approach to the flash subsystem. Linux runs on a data processing unit (DPU) that handles storage and networking in each sled, removing the need for a separate CPU-based storage server in that tier. Publicly identified participants in the initiative include Hammerspace, Los Alamos National Laboratory, Phison, ScaleFlux, SK hynix and its subsidiary Solidigm, Xsight Labs, and Linux community contributors. Hammerspace selected Xsight Labs’ E1 800G DPU for its OFP flash subsystem.

The HyperFRAME Research Lens: State of the Infrastructure and Operations (1H 2026) found that 52% of respondents rated unified management of compute, storage, and networking as critical or very important. In the State of the Enterprise AI Stack Lens (3Q 2026), only 27% of leaders reported reaching production while meeting their ROI objectives. Connecting the dots across these findings, we believe this covers two stages of the same challenge. Enterprises have to coordinate the infrastructure that gets AI into production and manage the cost of running it well enough to earn the expected return.

As detailed in the HyperFRAME Research Unifying Data for Enterprise AI Success research brief, Hammerspace can make data available to AI pipelines while existing applications continue to use it. Its Tier 0 architecture makes the NVMe drives already inside GPU servers available as shared storage and can stage selected files on the servers that need them. When a workload reads a file staged on its own server, the read can bypass the storage network. Enterprises can use capacity they already own and avoid a separate migration just to make a dataset available to AI.

Delivering the AI Data Platform as an integrated system adds the hardware component for customers who want to buy a complete, purpose-built offering. It carries forward Hammerspace’s use of Linux and open-source technologies while utilizing existing storage investments. Customers can add OFP where they need persistent flash performance and bring existing storage into the same namespace. Hammerspace told us it had won technical evaluations with its software but customers were requesting an integrated system.

In our view, Hammerspace has experience deploying the software at scale. Meta documented its work with Hammerspace on parallel NFS for its generative AI infrastructure, where engineers could make code changes visible across thousands of GPUs during a running job. Hammerspace also brought its AI Data Platform based on NVIDIA’s reference design to market in March 2026. The Data System builds on that experience and adds a flash subsystem designed around the same data path.

What Was Announced

The Hammerspace Data System brings existing and new storage across on-premises and cloud environments into one global namespace, with pNFS, NFS, SMB, and S3 access. Customers can expose selected live data to AI pipelines without an upfront migration or interrupting the applications that already use it. Administrators can use placement policies to move data closer to GPUs when a workload needs it. In the integrated system, metadata nodes and flash data nodes scale independently. Customers can also deploy the software on certified third-party hardware or in the cloud.

The new data intelligence tools index file and object metadata, and Deep Scan inspects file contents for sensitive information. Administrators can query the catalog in natural language and use the results to set an orchestration policy. For example, they can identify files that have gone unused for 30 days and move them from flash to a capacity tier. Hammerspace planned to demonstrate these capabilities at SC26 and said it had begun engaging early-access customers.

Tier 0 pools NVMe drives inside GPU servers as shared storage across the cluster. Training jobs can write checkpoints to Tier 0 and then move them to persistent storage. Inference workloads can retain KV cache data on shared flash for reuse. The reference design uses TLC flash for the performance tier and can use QLC flash, hard drives, object storage, or cloud storage for the capacity tier. Customers can also use existing storage for the performance and capacity tiers.

In terms of specs, a Hammerspace OFP tray holds up to five sleds in 1U. Each sled has a 64-core Xsight Labs DPU running Linux, two 400 Gb/s network ports, and up to eight U.2 NVMe SSDs. The design supports TLC and QLC flash, tool-free servicing, and expansion one sled at a time. Hammerspace measured 250 GB/s of read throughput across five sleds over TCP on a 400GbE fabric. It specifies up to 9.83 PB of raw capacity per tray using 245.76 TB QLC drives.

Hammerspace will size the system and sell the software on a recurring basis. The distributor coordinates quotes and final configuration, and the reseller delivers the system. Hammerspace monitors the hardware and takes the first support call for the full system. The company said it would begin quoting the Data System during SC26 on November 17 and expects general availability in the first half of 2027.

Looking Ahead

Turning Flynn’s design into a system enterprises can buy and deploy at scale will require trusted suppliers that can secure components, build to a consistent standard, and keep replacement parts available. Hammerspace has measured read throughput and plans to expand reliability and failover testing in larger lab clusters through the first quarter of 2027. It also needs to grow its ecosystem of distributors, integrators, and service partners so customers can deploy the system alongside existing infrastructure and get it serviced after installation.

Hammerspace has built a sizing tool for its sales engineers, and it will need to train its sales teams and channel partners to assess the workload, determine what infrastructure the customer can reuse, and specify where OFP adds enough performance or density to justify the purchase. We believe the initial enterprise opportunity is with customers that want an integrated, supported system and can measure the benefit of an optimized configuration. Shorter checkpoints, faster time to first token, or more flash capacity in a constrained rack give those buyers concrete gains to weigh against the cost of deploying and operating the system.

In our view, Hammerspace’s equity funding and market momentum give it a realistic chance of making the transition. Time will tell whether it can secure components and consistently deliver reliable systems as customer demand grows. If Hammerspace can do that, the Data System should help enterprises put existing data and infrastructure to work sooner and get measurable returns from AI workloads at scale.

Author Information

Don Gentile | Analyst-in-Residence, Data Platforms & Resiliency

Don Gentile brings three decades of experience turning complex enterprise technologies into clear, differentiated narratives that drive competitive relevance and market leadership. He has helped shape iconic infrastructure platforms including IBM z16 and z17 mainframes, HPE ProLiant servers, and HPE GreenLake — guiding strategies that connect technology innovation with customer needs and fast-moving market dynamics. 

His current focus spans flash storage, storage area networking, hyperconverged infrastructure (HCI), software-defined storage (SDS), hybrid cloud storage, Ceph/open source, cyber resiliency, and emerging models for integrating AI workloads across storage and compute. By applying deep knowledge of infrastructure technologies with proven skills in positioning, content strategy, and thought leadership, Don helps vendors sharpen their story, differentiate their offerings, and achieve stronger competitive standing across business, media, and technical audiences.