As SC26 gets underway in Chicago, the high-performance computing community is wrestling with a familiar set of pressures: soaring hardware costs, constrained power budgets, and a customer base that is shifting from traditional supercomputing sites toward GPU clouds and AI factories. For VDURA, that shift has defined the past year. In this pre-show Q&A, CEO Ken Claffey explains how the company, built on 25 years of parallel file system engineering and now running more than 1,000 production deployments in over 50 countries, has reoriented around the multi-tenancy, API-driven automation, persistent inference context and lower cost per GPU-hour that neocloud and AI factory operators asked for. The result is VDURA Data Platform V12, which became generally available on September 30.
Claffey makes the case that storage should be treated as a GPU revenue driver rather than a line item, and he backs it with figures he says CFOs and CIOs can take back to their teams, including more than 60% lower total cost of ownership and more than twice the performance per watt. He also covers VDURA’s open hardware ecosystem, including validation on Supermicro and an AI reference architecture with AMD, along with strategies for stretching scarce DRAM and NAND, security in multi-tenant environments, sovereign AI deployments, and where he expects the conversation to be headed by SC27. Read on for the full discussion, and find VDURA at booth 1423 for live V12 demonstrations.
++
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?
Ken Claffey: VDURA is the high-performance data storage infrastructure for AI factories and Neoclouds. Our software-defined, mixed-fleet HYDRA architecture keeps GPUs fed, tenants isolated and clusters online, with hyperscale-class durability and economics. We build on 25 years of parallel file system engineering, from PanFS to VDURA, and today run more than 1,000 production deployments in over 50 countries, with namespaces exceeding 1,500 nodes.
The biggest change over the past year is who is buying. The center of gravity has shifted from single-tenant supercomputing sites to GPU clouds and AI factories that sell compute as a service. Those operators handed us a very specific list: multi-tenancy, API-driven automation, persistent context for inference and lower cost per GPU-hour. VDURA Data Platform V12, which went generally available on September 30, is that list, shipped. It turns our platform into a multi-tenant, API-driven storage service that scales from 8 to 100,000+ GPUs on one software stack, running on the open hardware customers already buy, including the recently announced validation on Supermicro.
The other constant is efficiency. With all-flash and HDD unified in a single namespace, VDURA delivers more performance per watt and per rack unit at the best cost per terabyte, so customers can maximize the efficiency of their AI infrastructure and put more of their power and budget into GPUs.
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?
Claffey: You’ll find us at booth 1423 in Chicago, where we’ll be showcasing VDURA Data Platform V12. The biggest difference from St. Louis is V12 is generally available, with multi-tenancy, API-driven automation and persistent context for inference, and validated on open hardware including Supermicro. Our conversations have changed with it. Alongside the HPC sites we’ve served for 25 years, we expect to spend much of the week with AI factory and AI cloud operators who want to see how one storage platform scales from their first GPUs to their largest clusters. Our executive and engineering teams will be on hand for customer, partner, press and analyst meetings throughout the week.
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?
Claffey: Storage is a GPU revenue driver, not a line item. Every GPU-hour spent waiting on data is revenue an AI factory or neocloud never earns, and every kilowatt and rack unit storage consumes is one that can’t go to more GPUs.
For the CFO, the numbers are concrete. V12 delivers more than 60% lower total cost of ownership than competitive architectures at the same feed rate, and more than twice the performance per watt. Context-Aware Tiering keeps around 90% of files on flash while about 90% of capacity sits on lower-cost HDD, so customers don’t need to buy a second storage system for cold data.
For the CIO, it’s one software stack from 8 GPUs to 100,000, clusters that grow online from three nodes to thousands, and no manual rebuilds or downtime windows.
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?
Claffey: HPC Unites is exactly how we think about the hardware layer. VDURA is software-defined on standard servers and standard drives, with nothing proprietary in the stack, so the whole is built from the best of what our partners ship. V12 was recently validated on Supermicro Building Block Solutions, and with AMD we’ve built an AI reference architecture that pairs VDURA storage on AMD EPYC processors with AMD Instinct GPUs, ready to deploy on Supermicro or other hardware providers’ EPYC-based systems. Both sit within a broader open ecosystem: server platforms from AIC, Dell and Supermicro are certified today, with more on the way, alongside flash and HDD from the leading drive vendors and networking from AMD, NVIDIA and Broadcom. The same software runs on all of them. All nodes connect over RDMA directly to GPU nodes, with no dedicated back-end storage fabric, and VDURA is listed in NVIDIA’s GPUDirect Storage support matrix. Customers get the software-defined, mixed-fleet model hyperscalers use, on systems they already know how to buy and operate.
We’ve also leaned hard into open interfaces so storage slots into the rest of the stack instead of standing apart from it: REST APIs, Kubernetes CSI, infrastructure-as-code provisioning, S3 and KMIP key management. In V12, S3 objects are files within the same volume, so ingest, training, inference and archive all work from one copy of the data.
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?
Claffey: Our proof is in production, across more than 1,000 deployments. New Mexico State University, a Carnegie R1 institution, went live on VDURA in August with NVMe flash and HDD in a single namespace. V12 was named AI Data Management Solution of the Year in the 2026 AI Breakthrough Awards.
On V12 specifically: a single platform at roughly 20 PB usable spans a 35x performance range, from a capacity-optimized mixed fleet with 2% flash up to all-flash at 1,000 MB/s per TB. Our Elastic Metadata Engine, VeLO, delivers up to a 20x metadata improvement.
The benchmark operators care about most is GPU utilization. That’s the number we design for.
VMblog: How can attendees find you at SC26? What hands-on demos, live benchmarks, or interactive experiences will you have at your booth, and what’s the one thing you’d tell someone not to miss?
Claffey: Come to booth 1423 to see V12 running live. We’ll be demonstrating multi-tenancy with per-tenant isolation on a shared fleet, Context-Aware Tiering keeping small, hot data on flash while large, cold data settles on HDD with no stubs or rehydration, and the simplicity of expansion.
We’ll also be highlighting VDURA Sentinel, which monitors the fleet proactively and opens service requests with diagnostics attached before customers notice a fault. And we’ll have our GPU and Storage Optimizer Tool running live at the booth, so attendees can model their own environment and see how storage choices affect GPU utilization, power and cost.
The one thing not to miss is seeing a single cluster span capacity-optimized to all-flash performance on one software stack. It’s the clearest way to understand why AI factories don’t need a separate storage system for every tier.
VMblog: The memory and accelerator supply crunch has pushed up prices and stretched lead times for DRAM, HBM, and NAND. How are you helping customers plan and get more from constrained hardware, whether through flexible configurations, software efficiency, disaggregation, or extending the life of existing systems?
Claffey: Storage software can do a lot to stretch constrained hardware, and V12 attacks the problem at three points.
First, host memory. Our DirectFlow parallel client uses about 191 MB of DRAM and zero cores on the GPU node, so scarce memory and CPU stay with the workload rather than the storage stack.
Second, flash. With NAND prices volatile, Context-Aware Tiering automatically keeps small, hot data on flash and lets large, cold data settle on HDD, with no manual tuning. Customers buy flash for the data that needs it, not for everything. And because the platform runs on standard servers and drives from multiple vendors, customers can buy what ships, from whoever can ship it.
And we protect existing investments: V12 is available for all V5000-class systems and as an upgrade for V11 customers.
VMblog: Rack densities are climbing fast, and 800 VDC power and liquid cooling are now on nearly every roadmap. What are you showcasing for power, cooling, and performance per watt, and how do you help customers retrofit existing facilities rather than only build new ones?
Claffey: In power-constrained data centers, storage efficiency is the easiest retrofit lever available. Every watt storage doesn’t consume is a watt available for GPUs.
V12 delivers more than twice the performance per watt. A roughly 20 PB capacity-optimized mixed fleet runs at 18 kW. Because our nodes connect over RDMA directly to the GPU nodes, there’s no dedicated back-end storage fabric to power, cool or house, and SMR HDD support adds 25 to 30% capacity per rack. For operators retrofitting existing facilities, that combination frees power budget and rack space for accelerators without a rebuild.
VMblog: Simulation, AI training, and inference are increasingly expected to share one platform. How does your architecture balance the needs of traditional FP64 science against low-precision AI, and what do you tell customers who can’t afford to buy separate systems for each?
Claffey: From a data perspective, simulation and AI look very different. Traditional HPC is dominated by large, sequential I/O and massive checkpoints. AI training and inference bring millions of small files, heavy metadata load and latency-sensitive reads. Most storage systems are tuned for one and tolerate the other.
VDURA comes from 25 years of parallel file systems serving scientific computing, and V12 extends that to AI. VeLO handles the metadata storm AI creates, Context-Aware Tiering adapts data placement as access patterns change, and instant, space-efficient snapshots make checkpointing and recovery fast. With per-tenant isolation and namespaces, a simulation group and an AI group can share one fleet with guaranteed capacity and performance. Our advice to customers who can’t afford separate systems is simple: don’t build separate storage.
VMblog: With a Chinese system taking the No. 1 spot on the TOP500 and nations investing in domestic compute stacks, sovereignty and export controls are shaping procurement. How do sovereign AI, data residency, and geopolitical risk influence your product design and go-to-market, and how do you support on-prem, cloud, neocloud, and hybrid deployments?
Claffey: Sovereignty requirements favor software-defined architecture. Because VDURA runs on standard servers and drives from multiple certified vendors, customers can deploy where their data must live, on-prem, in a national AI factory, in a Neocloud or across a hybrid estate, without depending on a proprietary appliance supply chain.
V12 was built for exactly this. Multi-tenancy with per-tenant namespaces, VLAN isolation and encryption keys means one shared fleet can serve multiple organizations while each keeps control of its own data. AES-256 encryption at rest and in flight, with per-tenant KMIP key management, keeps keys in the customer’s hands. We already operate in more than 50 countries, and we launched V12 in the Middle East at Ai Everything Abu Dhabi in October, which reflects how central sovereign AI infrastructure has become to our business.
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?
Claffey: Keeping accelerators fed is the core of what we do, and V12 brings several advances. RDMA data paths with NVIDIA GPUDirect Storage move data directly between storage and GPU memory, bypassing the CPU. Persistent context extends KV cache beyond the pod so inference sessions resume without re-prefilling. Instant snapshots make checkpoints fast and space-efficient. And because S3 objects are files within the same volume, there are no staging copies between ingest, training, inference and archive.
On tiering, Context-Aware Tiering automatically places data on the right media by access pattern, so customers get flash performance where it counts and HDD economics everywhere else. And VeLO removes metadata as one of the most common hidden bottlenecks in AI pipelines.
VMblog: HPC environments are now multi-tenant, AI-enabled, and increasingly agent-accessible, which widens the attack surface. What security frameworks, zero-trust approaches, and supply chain protections are you implementing, and how do you avoid sacrificing performance?
Claffey: In a multi-tenant AI factory, isolation is security and performance at the same time. V12 provides per-tenant namespaces, VLAN isolation, encryption keys and QoS on a shared fleet, so one tenant can neither see another’s data nor degrade its performance. Data is encrypted with AES-256 at rest and in flight, with per-tenant KMIP key management.
Resilience is part of the security picture too. Failure domains can be as small as a single VPOD, and VDURA Sentinel monitors the fleet proactively, opening service requests with diagnostics attached before customers notice a fault. Qualifying V12 on a single, validated Supermicro hardware platform also gives customers a well-understood hardware supply chain.
VMblog: Talent shortages and lock-in fears are real. How are you making HPC and AI infrastructure easier to operate and more portable through open software stacks, containers, Kubernetes integration, or cross-vendor programming models, so lean teams can do more with less?
Claffey: Lean teams can’t spend their days tuning storage. V12 is API-first, with REST APIs, Kubernetes CSI and infrastructure-as-code provisioning, so storage fits into the deployment and billing pipelines teams already run. Context-Aware Tiering removes manual tuning, there are no manual rebuilds or downtime windows, and clusters grow online from three nodes to thousands.
VDURA Sentinel and VDURACare Premier take much of the operational load off the team: proactive service requests and parts dispatch, with 24×7 expert response under one contract. On lock-in, we’re software-defined on standard Supermicro hardware and speak open interfaces including S3, so customers keep their options open.
VMblog: Looking toward SC27 and beyond, which bets are you making, whether post-Moore architectures, photonics, neuromorphic, AI-designed systems, or something else? What will attendees be talking about a year from now that they aren’t talking about today?
Claffey: A year from now, I think the conversation will be about storage as part of the inference serving tier, not just the training pipeline. As agentic and long-context workloads grow, persistent context becomes a first-class resource. Where KV cache lives, and how quickly a session can resume, will directly shape cost per token.
The second shift is that neoclouds and AI factories will increasingly run like hyperscalers: multi-tenant, API-driven and obsessed with utilization. Power will remain the hard constraint, so performance per watt and capacity per rack will become standard procurement metrics. We’re building V12 and beyond around those bets.
##
