There’s a stat from Day 1 of Pure Accelerate 2026 that I haven’t been able to shake. VP of R&D and Customer Engineering, Shawn Rosemarin, put up the results of a survey Everpure ran with IDC across more than 1,300 IT leaders, and one number stopped him cold mid-keynote: 86% say storage is holding AI back. A companion IDC figure in the press materials goes further—94% of IT leaders call data quality the determining factor in AI success.
Sit with that for a second. Boards are demanding AI, billions are flowing into GPUs and models, and the people closest to the work keep saying the bottleneck isn’t the model. It’s the data plumbing. That gap is the entire reason Everpure spent its Day 1 keynote talking less about flash and more about getting your data ready for AI.
The last-mile problem, in plain terms
When I sat down with Everpure’s Chief Technology and Growth Officer, Rob Lee, I asked him about what I think of as the last-mile problem for AI—data accessibility and movement. His answer reframed the whole enterprise.
Most IT shops, he explained, are a pile of fragmented applications—ERP here, NetSuite for order management, Workday for HR, and on and on. The hidden cost is that the data, the systems of record, got trapped behind each app. Asking a simple question like “tell me everything about this customer” becomes genuinely hard. “You’re lucky,” Lee said, “if you have matching definitions of what customer means in these systems.”
We coped for years by exporting all of it into data warehouses and lakes and post-processing our way to answers. AI breaks that pattern, because now you need answers in real time, across silos. So Lee’s framing is that the architecture itself is outdated—and the fix is to make data the thing that takes the front-row seat instead of the applications.
Three things you actually need
Rosemarin boiled the requirements down to three, and they’re a useful checklist for any IT leader sizing up an AI project:
- AI-ready data — data that’s been refined and is ready to be served to models, not raw and scattered.
- AI-ready infrastructure — a unified platform that stores, understands, governs, and protects the data.
- Ecosystem integration — real partnership with the likes of NVIDIA, not a bolt-on.
What makes Everpure’s pitch different from the crowd, Rosemarin argued, is the refusal to make you copy everything into yet another system. “Other vendors want you to copy all your data into their system,” he said. “A copy is always behind.” The alternative they’re selling is bringing AI to the data where it already lives.
Data Intelligence: discover, classify, contextualize
The foundation here is Everpure Data Intelligence, the rebranded technology from the company’s 1touch.io acquisition. It does three jobs across your whole estate—on Everpure storage or not, structured or unstructured, cloud or on-prem or even mainframe:
- Universal Discovery — finds your data regardless of format, including SQL Server and Oracle, and shows exactly where it sits.
- Automated Governance — scans for sensitive info like PII and PHI, tracks lineage, and maps the landscape for compliance.
- AI-Ready Context — maps raw data to its real-world business meaning, building the semantic knowledge graph that lets agents understand and safely act on data.
The live demo made the stakes tangible. The team walked through an anonymized case—an athlete they called “Shane Falco”—where Data Intelligence surfaced a hereditary heart condition buried in a PDF on a file server, linked it to a dependent’s records siloed in a separate database, and flagged that the son had never been screened. That’s not a storage feature; that’s context doing something a human team would likely never catch. In another real example, the engine found 70% duplicative PII and PHI for a client and saved them $2 million in the first year on cyber-insurance premiums.
Data Stream: months of work down to minutes
The genuinely new product is Everpure Data Stream, built on NVIDIA’s AI Data Platform reference design. Lee described it as the answer for the enterprise that has a mountain of unstructured data and wants to make it usable by agents and RAG systems “but doesn’t have teams of PhDs to go build these systems from scratch.”
Data Stream automates the pipeline end to end—ingestion, curation, classification, indexing, vectorizing—and hands the results to AI agents or APIs. The claim Everpure is making, and it’s a bold one, is that work which takes skilled data engineers months can be automated down to minutes. It also keeps data inside the corporate network with stream-level access controls and uses a scale-out design so storage and compute grow independently as model needs change.
The on-stage demo featured a traffic-court legal assistant that ingested case data, made it searchable by meaning rather than keywords, and even predicted how a case might progress based on prior rulings in the dataset. It also exposed APIs and Model Context Protocol (MCP) support, so you can bring your own agent or build on top.
The NVIDIA co-design, and why hardware is cool again
This is where the partnership gets interesting. Kevin Deierling of NVIDIA joined Everpure CEO Charlie Giancarlo to explain that the two companies spent the last year as “extreme co-design partners” on Data Stream. NVIDIA’s AI Data Platform is the reference architecture; Data Stream is Everpure’s productized take on it. There’s more coming, too—Everpure is developing AI-native storage using NVIDIA Vera and the BlueField-4 STX storage processor to push acceleration and security closer to the data.
The closing panel, hosted by Rob Lee with Deierling and Crusoe’s Omar Lari, drove home why this matters for performance. Omar’s point landed hard: an 800-GPU Blackwell cluster holds about a quarter of a petabyte of HBM memory. If it takes 20 to 30 minutes to load model weights and training data, you’ve got hundreds of millions of dollars in GPUs sitting idle waiting for bytes. “Idle GPUs are economically destructive,” as STN’s CEO put it in the press materials. Storage and networking, in this world, are every bit as important as the GPUs—and Omar’s line that “hardware is cool again” got a knowing laugh from a room full of people who spent years abstracting it away.
The infrastructure underneath
The flashy AI software sits on real iron, and Everpure spent time on the parts that feed the models:
- FlashBlade//S delivers 220 GB/s and 450 million IOPS for low-latency inference.
- FlashBlade//EXA handles 4.6 billion metadata operations per second and pushed past 7,200 on SPEC AI—described as nearly 50% more than any prior benchmark—with over 10 TB/s of read performance feeding more than 10,000 GPUs without an idle cycle.
- KV Cache Accelerator speeds inference response by up to 20x.
- Portworx runs the containerized AI pipelines, with sub-minute failover, from edge to core.
Where this nets out for IT leaders
A fair question came up in the press Q&A about agentic AI and storage demand. Giancarlo’s take was practical: AI uptake has been strong in the enterprise but mostly in the cloud, because companies are still experimenting and don’t want to buy a $10 million AI farm they’re unsure about. As inference grows, though, you don’t want to answer the same question from scratch every time—you store results. And storage is cheaper than memory and cheaper than regenerating tokens. His point on “token minimization” was a nice corrective to the industry’s token-maximization noise: tokens cost money, and good data hygiene plus storage cuts the bill.
So here’s my honest read for the VMblog audience. The AI story underneath the Everpure rebrand isn’t really about faster flash, even though the flash is genuinely fast. It’s about a claim that the hard part of enterprise AI—getting fragmented, messy, ungoverned data into a state where an agent can trust it—can be automated and kept where it lives instead of copied into yet another silo. Data Intelligence ships now, Data Stream is here on NVIDIA’s design, and the IDC numbers give the whole pitch a tailwind. The open question is execution at enterprise scale and whether “bring AI to your data” holds up against the copy-everything-to-my-cloud crowd. Rob Lee’s advice to anyone starting was simple, and I’ll leave you with it: start today, pick a partner that’s nimble, because this AI world changes faster than anyone can plan for.
##
