Brian hosted Patrick Osborne, Senior Vice President of Hybrid Cloud Technology Acceleration at HPE, for a live podcast on LinkedIn on October 26. The subject was whether enterprise storage is ready for AI, and the answer ran through the Alletra Storage MP architecture decision HPE made five years ago, block storage’s second life under Oracle AI Database 26ai and vector search, the power and density math behind 245TB SSDs and a record year for LTO tape, Morpheus VM Essentials as the runtime under HPE’s private cloud systems, and cyber resilience built into the array.
We have covered most of that ground in the lab: the B10000 cyber resilience architecture and its ransomware lab, the X10000 with the Data Protection Accelerator Node, and a race weekend with Mercedes F1 and HPE at COTA the week before this recording.
Patrick has spent more than 16 years at Hewlett Packard Enterprise on storage, cloud infrastructure, and big data, running the Big Data, Analytics, and Scale-out Platforms business and then the Storage Business Unit before taking the cloud data infrastructure role that covers the Alletra Storage MP line today. He was last on the podcast for episode 119, recorded at HPE GreenLake Day for Storage in Houston in 2023.
HPE storage has been a regular in the lab: we reviewed the B10000 when it launched, measured backup and restore throughput through the Data Protection Accelerator Node in Raleigh, worked through the NIST CSF 2.0 cyber resilience architecture with HPE’s UK ransomware lab, and covered X10000 Release 4 and its move to 16 nodes and native NFS. The SimpliVity and VM Essentials stack came through last year as well.
The episode runs about 58 minutes. The numbers that stuck with us: Alletra Storage MP is going on five or six years old and scales from entry systems to tens of petabytes on one architecture, the “is my data AI-ready” question has led HPE’s customer conversations for 18 months, VM Essentials shipped monthly releases in its first year while GreenLake console updates land in weeks, LTO tape shipped more capacity last year than ever, and Mercedes F1 runs trackside with an IT team of about 18. If you don’t have time to listen end-to-end, we have broken it into segments so you can choose what matters most to you.
[00:00] Storage Is Interesting Again
Two storage veterans open on why the category went from commodity to center stage.
- Patrick describes a period of commoditization in storage and data management that pushed customers toward management experience, deployment patterns across data center, hybrid, and edge, and GreenLake economic models.
- For the last 18 months, he says, every conversation starts with whether the customer’s data and infrastructure are fit for AI workloads, a question with a wide surface area across security, performance, and the application ecosystem.
- Brian’s framing: the data revolution is bigger than AI, and the renewed interest in storage is off the charts.
[02:36] Why HPE Converged on Alletra Storage MP
Brian asks Patrick to revisit the decision to collapse bespoke storage platforms onto one architecture.
- Alletra Storage MP launched going on five or six years ago as what HPE calls the industry’s first disaggregated storage architecture, with separate compute and data nodes so customers scale capacity, performance, cost, and resiliency independently.
- The design uses standard x86 components, a low-latency networking fabric, and a media range from high-performance TLC to high-capacity QLC, with the differentiation at the software layer; earlier architectures leaned on battery-backed memory, crossbars, and custom silicon.
- Two flavors cover structured and unstructured data under one management paradigm, from entry systems to tens of petabytes of mission-critical storage, and Patrick cites unit growth and platform quality as the measure of its success.
[08:02] Block Storage’s AI Renaissance
AI storage usually means file and object, but Brian is seeing demand around Fibre Channel block for transactional databases adding vector search.
- Patrick frames AI as a pipeline of services with different I/O patterns: block for vector and semi-structured databases, file for brownfield unstructured data over NFS and SMB, and S3 object for net-new data from modern applications, with inference the fastest-growing use.
- One platform for block, file, and object plus data services (metadata tagging, resiliency, backup, DR) is the draw, and Brian adds that noisy-neighbor controls and QoS matter again when an Oracle 26ai vector workload lands beside production.
- GPU scarcity pushes customers to multi-tenancy and time-slicing, and HPE works with NVIDIA blueprints on storage-level integration such as KV cache and S3 over RDMA so parts of the data pipeline run in the storage layer.
[12:57] Tiering for Inference and Taking the Mystery Out
Brian asks what customers planning enterprise inference should think about beyond a shared KV cache tier.
- Patrick says the model and runtime landscape has normalized over 12 to 18 months, and HPE’s private cloud offers aim for first time to token in days or hours.
- The push since HPE Discover in June is a vertical stack for the enterprise, with MCP agents, data intelligence, and support for existing data types, so the customer is not the technology integrator and does not rip and replace.
- The two cover what HPE saw with Mercedes F1 at COTA: distributed sites, large data volumes moving back to a central factory, and orchestration that looks like any enterprise with branch locations.
[18:06] From Self-Driving Networks to the Self-Driving AI Data Center
Sports partnerships, Juniper, and why the storage conversation keeps widening.
- Beyond Mercedes, Patrick points to the Ryder Cup and NFL, NBA, EPL, and La Liga partnerships as real-time analytics use cases, each tuning a process for an edge.
- Customers who chose Juniper as their networking backbone get HPE’s CloudOps suite integrated with Mist, which Patrick positions as the self-driving AI data center, with networking as one piece of it.
- On economics, he says tokenomics and model matching only work when every layer of the stack is observable, which has been a challenge even inside HPE over the past year.
[21:53] Power, Density, and a Record Year for Tape
With GPU and CPU power budgets rising, Brian asks what the stress looks like for storage.
- The move from disk to hybrid to all-flash helped power and cooling, and SSD density and new form factors such as rulers push terabytes per rack unit when a rack can only take so much power; fully liquid-cooled systems are now part of the mix.
- Patrick notes the industry shipped more LTO tape capacity last year than in any other year, and Brian makes the case for LTO-10 libraries as a local archive that costs almost nothing to keep while older data becomes valuable again to newer analysis tools.
- Brian traces SSD capacity from a long plateau at 2 to 4TB through 15, 30, 61, and 122TB to 245TB, with E1.S, E3.S, and E3.L splitting platform support so no one can keep a few cases of drives on the shelf.
[28:23] Blueprints, Appliances, and Not Buying the Biggest Box
How HPE packages the stack for enterprises without deep AI infrastructure skills.
- Alletra Storage MP is the data foundation for HPE Private Cloud AI, which customers can build from an NVIDIA blueprint or buy as an appliance with NIMs, blueprints, and foundation models ready to run, then scale GPUs and storage as they go.
- Brian’s point is that the right answer is often not the largest system: a server with a couple of RTX PRO 6000 or 4500 cards can be the fit, and the hard part is matching a new platform announcement to the actual business outcome.
[30:42] Audience Question: Morpheus VM Essentials and Storage
The first live question asks how much VM Essentials will shape HPE’s storage hardware and architecture.
- VM Essentials launched about two years ago into a disrupted virtualization market, and Patrick says CSI driver downloads for Alletra Storage MP have gone through the roof as customers run bare metal, VMs, and containers side by side.
- Integration work runs from confidential computing and storage plug-ins for snapshots and replication to software-defined networking, and VM Essentials is now the runtime inside HPE Private Cloud 1000, 3000, and Private Cloud AI, managed through the CloudOps suite.
- HPE keeps a wide hardware compatibility list and plugs into Veeam, Commvault, Cohesity, Rubrik, and others, while offering a “better together” experience on HPE storage.
- Brian recalls the SimpliVity and VM Essentials lab work: deceptively simple to deploy, with data protection built into VM creation, and a fit for SLED, SMB, and branch offices.
[37:40] Multi-Tenant Isolation and Release Cadence
A second audience question on metadata leakage across tenants, and how fast HPE ships.
- Patrick takes the multi-tenant isolation question as a follow-up and a blog topic; HPE’s multi-tenant architecture isolates environments regardless of protocol, and his team owes the asker a post with specifics.
- On cadence, GreenLake console updates ship in weeks while array firmware stays put, breaking the old golden-copy habit of upgrading every 12 to 18 months; VM Essentials shipped monthly releases early on and has eased the pace as it matured, with a customer advisory board noting how far it has come in 18 months.
[42:09] Data Protection: DPAN, Zerto, and the Ransomware Lab
Brian turns to the Raleigh lab work on the Data Protection Accelerator Node and asks where security sits for HPE.
- Patrick says security is a board-level topic that is accelerating as attackers use AI to bundle and scale exploits, and HPE is a member of Project Glasswing alongside its technology partners.
- The Zerto acquisition brought a recovery point objective mindset into the storage layer, and the array now raises alerts, takes snapshots, and quarantines copies, including an off-prem golden copy, without a team of people to build and run it.
- HPE leads customer conversations with three topics: the AI opportunity, the VM reset, and security and compliance; Brian points to the UK lab where HPE exposes its systems to live ransomware.
- From the DPAN testing, Brian notes the restore bottleneck was the SSDs inside the target servers, not the fabric, which argues for sizing component choices down to the drive when a recovery time objective is on the line.
[48:22] Resiliency From Day One, and the IT Generalist
Platinum service levels defined at provisioning time, and what a small team at Mercedes F1 showed.
- With Morpheus Enterprise, performance, resiliency, and cyber posture attach to a service at deployment, bare metal, virtualized, or containerized, and follow it through its life.
- Brian says the Mercedes trackside team of about 18 uses the same tooling so engineers spend their hours on the car and not on administration, a tradeoff every business makes.
- Patrick rarely meets customers with dedicated backup, storage, and compute teams anymore; everyone owns a workload and a service level, and the generalist needs visibility and ease of management across the stack.
[52:30] What Comes Next for Enterprise Storage
Brian asks for the themes customers should plan around over the next six to 12 months.
- Patrick sees customers rethinking storage architectures for a mix of virtualization, cyber resilience mandates, and AI pipelines, with tight governance on data sharing, and a shift toward vertically oriented systems where one vendor supplies the full stack.
- He expects more AI data services embedded in the storage layer, including key-value stores, metadata queries, and new security standards, and eventually GPUs and models running in the storage tier itself.
- Brian adds that DPU-class devices will show up on both the compute and storage sides for offload, and the two close on five years of a platform that has not met a workload it could not handle.
You can read our B10000 cyber resilience paper, the X10000 Data Protection Accelerator Node review, the SimpliVity and VM Essentials review, and the Mercedes F1 race weekend piece for the work discussed in the episode.
Amazon