Source:ITTech Pulse, Aug. 18, 2026 – Ken, as CEO and President of VDURA, you’ve led the company’s transformation from Panasas into a modern AI and HPC data platform – what core expertise shapes how you approach that kind of reinvention?

I’ve spent my career in enterprise storage and HPC – at Seagate, Xyratex, Adaptec and Eurologic, building and scaling infrastructure businesses through inflection points in the market. At Xyratex, I helped build the ClusterStor HPC storage business that ended up powering 40% of the world’s top supercomputers. That experience taught me that reinvention isn’t about chasing trends, it’s about recognizing when the underlying workload has fundamentally changed and having the discipline to rebuild the architecture around it rather than bolt features onto what came before. Panasas had 25 years of parallel file system pedigree, but AI training and inference demanded a different operating model — software-defined, subscription-based, and built to move data intelligently across flash, disk and cloud. VDURA is that rebuild.

VDURA recently announced a technology alliance with Wasabi Technologies pairing GPU-adjacent performance storage with predictable S3-compatible cloud economics – what problem were you specifically trying to solve for AI infrastructure teams?

AI infrastructure teams are paying GPU-adjacent prices to store data that isn’t actively being used. Datasets, checkpoints and model artifacts don’t lose their value once a training run ends, they become the raw material for the next model, the audit trail for governance, the baseline for comparison, but keeping all of it on performance infrastructure is operationally wasteful and increasingly unaffordable given where flash pricing has gone. Moving that inactive data off high-performance storage has historically meant complexity, access delays or unpredictable cloud costs, so teams just didn’t do it. The Wasabi alliance solves that specific problem: it gives teams a clean, native path to keep the working set close to the GPUs while everything else moves automatically and affordably into S3-compatible cloud storage, without vendor lock-in to a single hyperscaler.

The alliance keeps active AI data close to GPUs while moving inactive data freely to the cloud, mirroring a tiering practice used by hyperscalers – how does this cut cost and complexity for enterprise teams?

Hyperscalers have run tiered infrastructure for years, lean, expensive performance capacity paired with deep, low-cost capacity for everything else because it’s the only economic model that scales. We’re bringing that same design pattern to enterprises and AI factories that don’t have hyperscaler engineering teams to build it themselves. Practically, that means organizations stop overprovisioning expensive flash to hold data they touch infrequently, and data movement becomes a native, automated part of the AI lifecycle rather than a manual, error-prone project. On the cost side, Wasabi’s predictable pricing, no per-GB egress or API charges, means teams can plan retention and reuse costs in advance instead of finding out on next month’s invoice. That combination is what actually lets infrastructure teams control both performance and spend at the same time.

VDURA’s V12 Data Platform recently won “AI Data Management Solution of the Year,” and you unveiled a next-generation control plane and expanded S3 capabilities at ISC 2026 – what sets this architecture apart technically?

V12 is built on our HYDRA architecture, which combines the speed of a true parallel file system with object storage in a single platform, so you’re not stitching together separate systems for performance and capacity. It’s designed to manage data across every stage of the AI pipeline: ingest, model load, training, checkpointing, fine-tuning, inference and archive. NVMe flash handles the performance-critical work, high-capacity HDDs handle cost-efficient retention, and Dynamic Data Acceleration together with our new Context-Aware Tiering engine moves data between them automatically. The next-generation control plane we introduced at ISC adds true multi-tenant management with a modern interface and REST API, and the expanded S3 capabilities, including native object tagging, give data engineers policy-based lifecycle management and automated tiering across training datasets and model artifacts. It’s throughput at 0.5-4 GB/s per GPU, scaling linearly across thousands of nodes, with simplified data protection through new snapshot capabilities.

VDURA is built around an industry-leading 12-nines durability commitment and a single global namespace spanning NVMe flash, HDD, and native S3 – how does that architecture change what customers can trust with their AI data?

Twelve nines of durability mean customers can stop treating data protection as a trade-off against performance or cost, it’s simply built in. When you put NVMe flash, HDD and native S3 under a single global namespace, customers aren’t choosing which tier to trust with which data; the platform handles placement and protection automatically while giving them one consistent view and one set of access controls across the entire estate. For AI specifically, that matters more than people initially assume: training data, checkpoints and model artifacts increasingly carry compliance, governance and reproducibility requirements, and a single namespace with that level of durability means teams can retain and audit that data with confidence instead of stitching together separate protection strategies for each storage tier.

Looking to 2026 and 2027, with NAND allocation tightening and DRAM prices climbing, how do you see AI infrastructure teams rethinking flash-first storage strategies, and where does VDURA fit into that shift?

The flash supply situation has gotten serious, 30TB SSD pricing surged over 250% in about a year, and lead times have stretched out significantly. Flash-first, all-flash approaches that made sense eighteen months ago are becoming very hard to justify economically at scale. What we’re seeing now is teams recognizing that not all AI data needs to sit on the most expensive tier, only the active working set does. VDURA’s whole architecture is built around that principle: put flash where it earns its cost, on the data GPUs are actually touching right now, and let everything else live on HDD or in the cloud without sacrificing performance when that data is needed again. As NAND allocation tightens and DRAM prices climb further into 2026 and 2027, I think intelligent tiering stops being an optimization and becomes the only viable strategy and that’s exactly the position VDURA is built for.

What advice would you give to IT and infrastructure leaders who are building AI data pipelines today but haven’t yet planned for the full lifecycle of their data, from active training through long-term retention?

Don’t treat storage as a decision you make once at the start of a project. The teams that get into trouble are the ones who architect for training day one and never plan for what happens to that data on day two, when the run ends, the checkpoint needs to be retained, the dataset needs to be compared against the next version, or a regulator asks for an audit trail. Start by mapping your data’s actual lifecycle, ingest, training, checkpointing, inference, archive, and build tiering and retention into the architecture from the outset rather than retrofitting it later under cost pressure. And build that plan assuming flash economics will keep getting harder, not easier. The organizations that plan for the full lifecycle now will be the ones with headroom to keep innovating in 2027; the ones that don’t will find themselves making expensive, reactive decisions when their performance tier fills up.