The all-flash consensus is underwater.
The cause is structural, not cyclical. AI inference demand has outstripped storage supply. Memory manufacturers are shifting NAND production lines to DRAM, where margins are fatter. North American hyperscalers are stockpiling. Short-term capacity tightness is "difficult to resolve."
The economics made sense when NAND was cheap, and capacity was abundant. Neither holds. The organisations locked into all-flash architectures are now exposed.
One storage company saw this coming.
Who are they?
What is VDURA?
VDURA is a modern data storage infrastructure software built for AI factories and Neocloud operators.
Not a general-purpose filer. Not an all-flash array betting on stable SSD pricing. A true parallel file system with 25 years of production hardening, now re-architected around HYDRA - a hyperscale-inspired, software-defined platform delivering distributed metadata, extreme durability, and hybrid economics under a single namespace.
The company targets three buyer profiles:
Neocloud operators: GPU cloud providers who must maximise utilisation, adapt to changing workloads, and control costs as data volumes grow exponentially.
AI factories: Large-scale training and inference environments running continuous production workloads.
Traditional HPC: Life sciences, energy, manufacturing, federal, research - the verticals where Panasas built its name across 1,000+ deployments in 50+ countries.
VDURA is Panasas evolved. Founded in 1999 by Garth Gibson, who co-invented RAID at UC Berkeley and built Carnegie Mellon's Parallel Data Lab. The company shipped the first enterprise-grade parallel file system, pioneered file-level erasure coding, and deployed at NASA, RTX, Boeing, and Airbus. That production-hardened foundation now runs as VDURA's software platform.
On May 7, 2024, Panasas rebranded to VDURA and pivoted from proprietary appliances to software-defined storage with subscription pricing. The company has raised over $150 million to date.
What Problem Does VDURA Solve?
AI workloads have outgrown general-purpose storage. VDURA breaks the problem down into three structural failures:
Storage stalls kill GPU economics. Any bottleneck in data delivery leaves expensive accelerators sitting idle - a direct hit to Neocloud margins.
Centralised metadata cannot scale. Training and inference pipelines involve thousands of concurrent reads, writes, and checkpoints. Systems that serialise metadata access or rely on single-controller architectures collapse under AI concurrency.
All-flash economics are volatile. VDURA believes SSD pricing follows boom-bust cycles that organisations cannot forecast over a three-year planning horizon. HDD pricing moved 35% over the same period, and the flash market surged - a structural gap that punishes rigid all-flash commitments.
VDURA solves this with HYDRA - a hybrid-native, distributed architecture built for AI-scale parallelism.
HYDRA delivers true parallelism across both data and metadata, supports flash-first performance where GPUs demand it, and enables scalable capacity tiers for checkpoints and retention under a single namespace.
Performance and capacity scale independently. The architecture tolerates failure as a normal operating condition.
For Neocloud operators, this eliminates the need to stitch together multiple storage systems across different pipeline stages, enabling the entire AI lifecycle to run on a single platform.
Who Runs VDURA?
Ken Claffey (CEO): Scaled ClusterStor at Xyratex (later Cray/HPE) to power 40% of top supercomputers globally. Led Seagate's enterprise storage business. Now driving VDURA's AI-first strategy for GPU clouds and AI factories.
Garth Gibson (Chief Technology & AI Officer): Co-invented RAID at UC Berkeley. Co-founded Panasas. Former CEO of the Vector Institute. Has advised on $1B+ AI infrastructure initiatives. Returned September 2025 to architect VDURA's AI-scale storage foundation.
Paul Hiemstra (CFO): Former CFO of HPE's HPC & AI business. 10 years as Treasurer at Cray.
Erik Salo (SVP Marketing & Business Operations): Former VP at Seagate, Chief Strategist at AMD. Leads VDURA's flash economics research and the Flash Volatility Index.
Hanish Vaghela (EVP Sales): Built a $1B+ ClusterStor business at Xyratex. Now leading VDURA's global go-to-market across Neoclouds and AI factories.
Samantha Clarke (VP Neocloud & AI Business Development): Enterprise infrastructure leader from Seagate and Xyratex. Owns VDURA's strategy with Neocloud operators and GPU-centric infrastructure partners.
Christopher Girard (VP Product Management): AI and HPC storage product leader across Dell, Seagate, HPE Cray, and Cohesity.
Peter Maddocks (VP Engineering): 25+ years building enterprise storage systems at Seagate, Dot Hill, and HP.
Rex Tanakit (VP Technical Services): HPC practitioner with hands-on experience across CFD, signal processing, and GPU-accelerated workloads. Responsible for deployment success and operational reliability in large AI and Neocloud environments.
What is VDURA's competitive edge?
VDURA is a true parallel file system with distributed metadata: VeLO metadata engine distributes metadata services across the cluster, delivering billions of inode operations per second. This eliminates the metadata bottleneck that kills checkpointing performance and small-file inference throughput. Data and metadata paths operate in parallel - no centralised controller, no serialisation, no head-node chokepoint.
Extreme durability without hardware lock-in: 12 nines of data durability in all-flash configurations. 11 nines in hybrid environments. Software-defined multi-level erasure coding and automated file-level self-healing, not tied to proprietary hardware designs.
Flash-first performance without all-flash economics: The V5000 all-flash NVMe platform delivers GPU-saturating throughput for training and inference. Mixed Fleet, hybrid configurations extend into scalable capacity tiers for checkpoints, datasets, and retention - all under one namespace. Deployments have demonstrated a total cost of ownership that is more than 60% lower than that of all-flash architectures at similar performance levels.
The Flash Relief Program: VDURA backs its thesis with a commercial commitment: submit a competitor configuration from any all-flash vendor, and receive a matched or lower bid - up to 50% lower - within 24 hours.
Shared-nothing, hyperscale-inspired design: HYDRA handles availability, fault tolerance, and resilience entirely in software. No HA pairs, no RAID controllers, no dual-ported devices. Linear scale, faster recovery, hyperscaler-grade efficiency on commodity servers.
Data-driven sizing tools: The GPU Storage Calculator helps operators size storage based on GPU count, workload type, throughput requirements, and data retention - aligning storage investment with real GPU utilisation. The Flash Volatility Index quantifies real-world SSD price swings and their impact on AI infrastructure costs.
Recent Moves
February 2026: Craig Bernero and Eddie White appointed to board.
January 2026: Flash Relief Program launched. Public commitment to beat Vast/Weka quotes by up to 50%. Flash Volatility Index and Storage Economics Optimiser Tool launched. First tooling to quantify real-world SSD price instability and its impact on AI infrastructure economics.
November 2025: World-record 282 PB data lift at Supercomputing 24 with Hafþór Björnsson ("The Mountain"). Partnership continues through ISC 25 and SC 25, including the AI Data Transfer Challenge, sustaining 5.1 PB/s. Data Platform V12 also announced. Elastic distributed metadata engine, system-wide snapshots for AI checkpointing, SMR optimisation.
September 2025: Garth Gibson returns as CTAIO from the Vector Institute.
July 2025: AMD scalable AI reference architecture delivered. Partnerships expanded with NVIDIA, Dell Technologies, and AIC.
March 2025: V5000 all-flash NVMe platform released. GPU-saturating throughput for AI training and inference.
May 2024: Panasas rebrands to VDURA. Announces software-defined pivot and subscription pricing.
September 2023: Ken Claffey joins as CEO from Seagate Enterprise.
What's Next for VDURA?
Data Platform V12 will reach general availability in the coming months.
This will bring the elastic metadata engine, system-wide snapshots, and SMR optimization into production. Higher concurrency and faster checkpointing are the goal here, unlocking simpler operations for multi-tenant Neocloud environments. Multi-tenant isolation, sustained GPU utilisation under mixed workloads, and operational tooling built for always-on AI services rather than batch-only environments are also on the roadmap.
Partner-led and delivered reference architectures will also expand beyond AMD with VDURA planning validated AI platforms across diverse GPU, server, and networking ecosystems.
The bet underneath all of this is that AI infrastructure is a lifecycle problem, not a storage SKU decision.
The faster organisations move models from ingest to training to inference, the faster they generate revenue. VDURA is positioning to own that compression.
The risk is market position.
Competitors are landing billion-dollar-plus contracts on the one hand, and raising billion-dollar-plus funding rounds on the other. They have ecosystem momentum and enterprise sales motion that VDURA's $150M in total funding cannot match directly.
The Flash Relief Program's 50% undercut is aggressive positioning, but guarantees only work if buyers comparison-shop - and many default to the all-flash incumbents without looking.
The strongest signal in VDURA's favour, however, is production scale.
20PB+ deployments running 800+ GB/s sustained throughput. Exabytes managed globally across thousands of deployments. Parallel file system maturity that competitors cannot easily replicate with funding alone.
Garth Gibson's return is an indicator of the depth of conviction VDURA’s leadership has in their chosen direction of travel - the HYDRA architecture is a deliberate break from legacy storage assumptions.
Given today’s conventions, the question is whether Neocloud operators and AI factory builders will bet on purpose-built infrastructure over the all-flash consensus - and whether VDURA can capture that demand before flash pricing normalises and the economic argument narrows.