As SC26 opens in Chicago under the theme “HPC Unites,” Hammerspace arrived with a notably different story than it told a year ago. The company is expanding from a software-defined data platform into what it calls the Hammerspace Data System, combining Data Unification, Data Intelligence, Data Orchestration, Data Acceleration and Data Storage. It is backing that shift with Fall 2026 IO500 results the company says show standards-based pNFS on the native Linux client outperforming purpose-built AI storage platforms in the 10-Node Research category, along with a standing-room-only inaugural Hammerspace User Group and production deployments such as the U.S. Army Research Laboratory’s Patriot AI supercomputer.
In this VMblog Q&A, Molly Presley, Senior Vice President of Global Marketing for Hammerspace, explains how the company is positioning itself for an era of tightening AI infrastructure budgets and rising hardware costs. Her message to attendees is to get more value from the infrastructure they already own before buying more. She walks through Tier 0 acceleration of server-local NVMe, the new DPU-based Open Flash Platform powered by the Xsight Labs E1 DPU, and the company’s emerging pitch as a standards-based alternative to Lustre in traditional HPC environments. She also addresses how KV cache, checkpointing and the broader AI memory hierarchy factor into keeping costly accelerators fed.
++
VMblog: If you were giving an SC26 attendee a quick overview of your company, what would you say? How has your company evolved since SC25, and what has been the biggest change in your business over the past twelve months?
Molly Presley: Our biggest evolution since SC25 is the expansion from a software-defined data platform into a complete, integrated Data System. At SC26, we’re introducing the Hammerspace Data System, combining five capabilities: Data Unification, Data Intelligence, Data Orchestration, Data Acceleration and Data Storage.
Over the past year, enterprises, research institutions and government agencies have been using Hammerspace to operationalize AI, accelerate scientific discovery and support mission-critical workloads using data distributed across existing infrastructure. These deployments demonstrate the strength of our standards-based architecture, combining Linux-native parallel I/O, global metadata and policy-driven orchestration to deliver HPC-class performance across storage systems, data centers and clouds without disruptive migrations or proprietary clients. The expanded Data System builds on this proven foundation, adding new capabilities without requiring customers to rebuild their infrastructure.
We’ve also dispelled the assumption that HPC-class performance requires proprietary file systems. Our Fall 2026 IO500 results demonstrate that standards-based pNFS, using the native Linux client, can outperform purpose-built AI storage platforms. In the 10-Node Research category, Hammerspace achieved a substantially higher overall score than competing AI storage vendors’ submissions, demonstrating that open standards and industry-leading performance are not mutually exclusive.
This differentiation is translating into significant business momentum. We’ve expanded our presence across neoclouds, quantitative trading, national laboratories, defense and enterprise AI. As customer deployments grow in scale and complexity, we’re expanding how we deliver our technology.
We’re doing for AI data what AI factories have done for compute: transforming complex, fragmented infrastructure into one unified system that’s simpler to deploy, operate and scale. Unlike proprietary storage architectures that create new silos, Hammerspace combines the simplicity of an integrated system with standards-based interoperability, providing a consistent data foundation as GPU generations, storage technologies and AI workloads evolve.
VMblog: Your company is participating in SC26 in Chicago. What does that presence look like, and if you were in St. Louis last year, how does it differ from your SC25 participation?
Molly Presley: Our presence at SC26 reflects how significantly Hammerspace has advanced over the past year. We’re moving beyond demonstrating what our technology can do to showcasing production deployments, major advances in benchmark performance and a growing community of organizations using Hammerspace for AI and HPC.
One highlight is our standing-room-only inaugural Hammerspace User Group, bringing together customers, partners and industry leaders to share experiences deploying Hammerspace across demanding computing environments. The level of engagement reflects a growing community putting our technology to work for AI and HPC.
At our booth, we’re showcasing real-world deployments, including the U.S. Army Research Laboratory’s Patriot AI supercomputer, which uses the Hammerspace parallel file system for high-performance data access. Patriot features 160 NVIDIA H200 GPUs, 48 NVIDIA RTX 6000 GPUs and a 400Gb storage network, supporting demanding AI, simulation and scientific research workloads.
We’ve also made substantial advances in performance since SC25, demonstrated by significant improvements in our IO500 benchmark results. These gains reinforce that standards-based pNFS, using the native Linux client, can deliver exceptional parallel file system performance without specialized client software. Our latest results challenge the assumption that HPC-class performance requires proprietary or non-standard file system interfaces.
We’re also beginning to address an important opportunity in traditional HPC environments: providing a standards-based alternative to Lustre. While Lustre is open source, it requires specialized client software rather than a standards-based file system interface native to Linux. Hammerspace delivers parallel I/O through Linux-native pNFS while adding global data unification and policy-driven orchestration. For organizations evaluating the next generation of HPC infrastructure, this provides a compelling alternative without sacrificing performance or interoperability.
The biggest difference from SC25 is the evidence. We’re bringing production customer deployments, dramatically improved benchmark performance and an expanding user community to Chicago. Hammerspace is demonstrating that standards-based data infrastructure can deliver the performance required for the most demanding AI and HPC environments.
VMblog: What key message will SC26 attendees hear from you this year? With AI infrastructure budgets under heavy scrutiny and hardware costs rising, what concrete value proposition can they take back to their CFO and CIO?
Molly Presley: Our message to SC26 attendees is simple: Before buying more infrastructure, get more value from the infrastructure you already own. AI infrastructure economics are changing rapidly. Rising hardware costs, power constraints and pressure to maximize GPU productivity mean organizations can no longer afford to solve every new performance or capacity requirement by purchasing another storage system.
Hammerspace fundamentally changes those economics in three ways.
First, unlock the storage capacity already deployed. Most organizations have significant investments in NAS, flash and HDD infrastructure, with capacity distributed across systems and locations. Hammerspace unifies that capacity into one Data System, making existing data and storage resources available to AI without disruptive migrations, duplicate datasets or additional storage silos. Organizations can defer new storage purchases while increasing utilization of investments already on their balance sheets.
Second, turn existing storage into high-performance AI infrastructure. Hammerspace delivers standards-based parallel file system performance across existing NAS storage, overcoming the performance limitations of traditional storage access architectures. With Tier 0 acceleration, server-local NVMe becomes a high-performance, shared data acceleration tier, dramatically increasing throughput and reducing data access bottlenecks without replacing underlying storage. Organizations can feed more GPUs, reduce idle time and accelerate training and inference using infrastructure they already own.
Third, radically improve the economics of new storage. For workloads that require additional high-performance flash, the new Hammerspace Open Flash Platform (OFP) introduces a high-density, DPU-based architecture that eliminates conventional x86 storage server overhead. OFP delivers substantially greater performance density with lower power consumption, rack space and infrastructure costs, changing the price/performance equation for AI training, inference and agentic workloads.
The value proposition for a CFO and CIO is straightforward: more usable capacity, more performance from existing infrastructure and dramatically better economics when new storage is required. Hammerspace enables organizations to scale AI by maximizing what they already have, accelerating it where needed and adding new capacity far more efficiently. The result is higher GPU productivity, lower infrastructure costs and a more sustainable path to scaling AI.
VMblog: This year’s SC theme is “HPC Unites.” Where does your company fit into that idea? Which partnerships, joint reference architectures, or open standards are you announcing or demonstrating that make the whole greater than the sum of its parts?
Molly Presley: “HPC Unites” aligns directly with our vision of bringing AI and HPC together on one unified, standards-based Data System. Rather than building separate infrastructure for every workload, Hammerspace unifies existing and new technologies to deliver HPC-class performance, maximize infrastructure investments and accelerate scientific discovery and AI.
Our partnerships with NVIDIA, Cisco and Supermicro bring compute, networking and data together in integrated AI architectures. Our new Hammerspace Open Flash Platform, powered by the Xsight Labs E1 DPU, connects high-density flash directly to standard networks without conventional x86 storage servers, improving performance density and energy efficiency.
Hammerspace unites AI and HPC with the simplicity of an integrated system and the flexibility of open standards, making existing and new infrastructure more powerful together than in isolated silos.
VMblog: Nearly every vendor now claims to be “AI-native” and “HPC-ready.” What are your provable differentiators, and which independent benchmarks or production customer results (IO500, Green500, MLPerf, real-world deployments) back them up?
Molly Presley: Our fundamental differentiation is architectural. With conventional AI storage systems, the system typically stops where the vendor’s storage ends. Hammerspace spans the entire data estate, unifying existing and new storage with integrated data intelligence, orchestration, and high-performance access.
We’re also focused on delivering performance more efficiently. Our new Open Flash Platform is designed to deliver 250 GB/s of read throughput and 350 GB/s of write throughput per rack unit. Based on published peak specifications and evaluated configurations, that represents 37% greater read throughput density and up to 4.4 times the write throughput density of the highest-performing alternatives compared.
Equally important is what customers are achieving in production environments. Organizations across AI and HPC, including research institutions and AI companies, are using Hammerspace to simplify access to distributed data and support demanding workloads.
As Daan Rademaker, Staff Data and AI Platform Engineer at kaiko.ai, explains: “Hammerspace has been crucial for us to move toward one single data system. No matter where we want to do our training, it can be started from the same place, and we know the data is going to be there.”
For us, the real measure of AI and HPC readiness isn’t simply peak storage performance. It’s whether organizations can consistently deliver data to compute, minimize unnecessary infrastructure costs, and scale without creating new silos.
VMblog: With memory scarce and expensive, and modern AI leaning heavily on KV cache, checkpointing, and massive datasets, what innovations in memory tiering, storage orchestration, or interconnect (Ethernet, InfiniBand, optical) are you unveiling to keep costly accelerators fed?
Molly Presley: As AI shifts toward inference and agentic workloads, the challenge is no longer simply moving large datasets into GPU clusters. It’s managing an increasingly complex memory and storage hierarchy, where training data, checkpoints and reusable inference context must be delivered at high speed without consuming expensive GPU memory or creating network bottlenecks.
Hammerspace Tier 0 changes the role of server-local NVMe. Instead of leaving flash isolated inside individual GPU and CPU servers, Tier 0 makes that capacity part of a shared, persistent, high-performance storage tier accessible across the cluster. Combined with Linux-native parallel I/O and LOCAL-IO optimizations, this enables high-throughput data access and faster checkpointing while reducing unnecessary external storage traffic.
For inference, this architecture also provides a foundation for shared context storage. Tier 0 can provide persistent, high-speed access to KV-cache data, enabling inference frameworks to retain and reuse computed context across requests and nodes when integrated with their cache management software. This reduces redundant computation and can improve time to first token and GPU efficiency.
Our new Hammerspace Open Flash Platform extends this approach beyond server-local flash. Its DPU-based architecture delivers high-density context storage directly over standard networks, providing a more power- and space-efficient way to scale capacity as inference workloads and context requirements grow.
The result is a more efficient AI memory hierarchy: use expensive GPU memory for active computation, server-local NVMe for high-speed shared access, and dense network-attached flash for scalable context capacity. Hammerspace enables organizations to optimize each layer without building a separate storage silo for every new AI workload.
##






