Architecting for Agentic: Key Takeaways from HPE Discover 2026

HPE Discover 2026 ran June 15 through 18 at The Venetian Convention and Expo Center in Las Vegas, drawing more than 20,000 IT professionals across four days of keynotes, sessions, and partner activities. This year's conference was the first Discover to fully reflect what HPE's $14 billion acquisition of Juniper Networks actually means in practice, and the answer turns out to be a thesis rather than a product list: agentic AI does not succeed on compute alone. It succeeds on the network that connects the agents, the data, and the governance layer that keeps them in line.

CEO Antonio Neri opened the week with that argument and did not deviate from it. Every major announcement, whether in compute, storage, or networking, traced back to the same organizing idea. The industry is shifting from generative AI, which responds to prompts, to agentic AI, which acts autonomously and pursues multi-step goals without constant human oversight. That shift changes what infrastructure needs to do, and it falls hardest on the network. HPE's position is that it now owns that layer end to end, from campus and branch through the AI data center, and that this is what separates it from competitors who can sell you a GPU rack but not the coherent fabric around it.

Compute

The compute announcements at Discover centered on one new platform and a broader expansion of the AI factory portfolio.

The headline system is the HPE ProLiant Compute DL394 Gen12, powered by a CPU architecture from HPE's industry-leading AI compute partner and designed specifically for agentic AI, reinforcement learning, and CPU-intensive workloads requiring low-latency memory access. This is a meaningful departure from the traditional compute roadmap. The processor comes from that AI industry-leading partner rather than from the traditional x86 vendors, and pairing it with HPE's ProLiant management and security stack creates a system purpose-built for the inference and orchestration workloads that drive agentic workflows. The DL394 Gen12 is currently announced with availability in 2027, which matters for pipeline timing.

Alongside the DL394, HPE broadened its AI factory solutions with next-generation accelerated rack-scale systems, the HPE Compute XD700 built on the latest rack-scale GPU platform, the HPE Cray Supercomputing GX240 Compute blade designed with that same industry-leading partner CPU, and next-generation 800G InfiniBand support for the HPE Cray Supercomputing GX5000. For organizations running large-scale training today, HPE also announced that its newest rack-scale systems deliver four times fewer GPUs for equivalent AI training versus the prior accelerator generation, and one-tenth the cost per million tokens for inference. These figures are vendor-stated and should be treated as directional rather than independently verified benchmarks until HPE publishes test methodology.

The HPE Superdome Flex line received a successor announcement as well, retaining its position as the SAP HANA certified leader and introducing a new model with the highest certified memory capacity in its class. For SAP environments specifically, this matters: scale-up architecture continues to be the preferred path for HANA at enterprise scale, and HPE is reinforcing that position heading into a refresh cycle where many organizations are re-evaluating their platform choices.

On the software side, HPE CloudPhysics helps customers identify over-provisioning and free up budget ahead of refresh cycles, which is increasingly relevant as organizations right-size their compute estates before committing to AI infrastructure investments.

The broader compute picture from Discover is that HPE is building toward a full-stack AI factory architecture where compute is validated, networked, and governed as a system rather than sold as individual SKUs. The HPE AI Factory now incorporates an integrated agent toolkit, confidential computing, and open foundation models as built-in components. Confidential computing will become standard across the entire HPE AI Factory portfolio in Q4 2026, which is the more significant enterprise-readiness move of the week for regulated industries and sovereign deployments. It protects data in use by executing sensitive workloads inside hardware-based trusted execution environments, extending security posture well beyond encryption at rest or in transit.

Storage

Storage announcements at Discover followed two tracks: raw platform performance for AI workloads, and a unified data fabric for governing the data those workloads consume.

The HPE Alletra Storage MP X10000 is the platform generating the most attention. It delivers unified file and object storage on a single architecture and is among the first object storage platforms validated under a leading enterprise-AI storage certification program. For organizations building out Private Cloud AI environments, this validation removes a qualification step from the deployment process and positions the X10000 as the reference storage tier for HPE's AI factory architecture. HPE also confirmed the X10000 is being brought directly into the Private Cloud AI stack, making it a native storage tier within the platform rather than a separate add-on. In her keynote, CTO Fidelma Russo framed why this matters: storage is becoming "active memory for AI." Using KV cache offload, the X10000 keeps an agent's working context available beyond GPU memory rather than rebuilding it on every inference, which is where a surprising amount of token cost and latency hides. HPE's stated results, measured against traditional architectures without that capability, are up to 20x faster time to first token and roughly 17x higher throughput, achieved not by adding GPUs but by keeping the ones you already own productive. As with the compute performance claims from this event, these figures come from HPE's own materials without published test methodology, and procurement models should request configuration details before relying on them.

The HPE Alletra Storage MP B10000 rounds out the portfolio as the high-performance block storage tier, positioned as the fastest all-flash block storage array in its class for mission-critical workloads. Together, the X10000 and B10000 give HPE a two-tier Alletra story that covers both unstructured AI data pipelines and traditional mission-critical applications from a software-defined disaggregated architecture.

On the software layer, HPE announced Data Fabric 8.2, delivering agentic workload integration, a global data catalog, and turnkey deployment. This is the piece that ties storage to the broader agentic AI architecture. As agents multiply and make repeated requests across distributed data sources, the cost of unmanaged data movement scales quickly. Russo made the stakes concrete with HPE's own example: an AI-first support platform, code-named Milestone and run on-prem on GreenLake Intelligence and Private Cloud AI, that she said cut costs roughly 30x and saved nearly $100,000 a month. Her framing was that HPE "stopped being a consumer of AI and became a producer of intelligence." Data Fabric 8.2 is HPE's structural answer to that tokenomics problem, with general availability targeted for October 2026.

For competitive displacement and commercial opportunity, HPE launched a storage takeout program beginning July 1 offering 15% front-end margin plus up to 9% in rebates, totaling 24% combined. The HPE Smart Choice SKUs with aggressive pricing and accelerated delivery windows were also positioned as a response to customers who need near-term refresh options without long lead times.

Networking

Networking was the defining story of Discover 2026, but the bigger story wasn't simply faster switches or larger fabrics. It was HPE making the case that networking has evolved from moving packets to operating AI infrastructure through intent, automation, and continuous validation. With the Juniper acquisition now complete, Discover was the first opportunity to demonstrate how Juniper Apstra, Mist AI, Aruba, and the broader HPE portfolio fit together as a unified architecture designed for modern data centers.

Throughout the conference, executives reinforced a message that resonated with customers: AI changes applications, but networking determines whether those applications operate reliably at scale. As AI environments become more distributed, organizations can no longer afford manual configuration, inconsistent deployments, or troubleshooting by CLI. That is where Juniper Apstra becomes one of HPE's most strategic assets. Its intent-based networking model, continuous validation, closed-loop automation, and graph-based understanding of the network allow operators to build fabrics faster while dramatically simplifying Day 2 operations. Instead of simply deploying a network, Apstra continuously verifies that the network remains in its intended state and quickly identifies configuration drift before it becomes an outage.

The new QFX switching portfolio complements that vision. New QFX platforms expand support from 25GbE through 800GbE while providing the high-density leaf-and-spine infrastructure required for modern AI and enterprise data centers. Combined with Apstra, these platforms become more than high-performance switches. They become part of an automated fabric where provisioning, validation, lifecycle management, and operational assurance are handled consistently across the environment. Rather than managing individual devices, operations teams manage business intent, allowing infrastructure to scale without a corresponding increase in operational complexity.

Operational simplicity was another recurring theme. HPE expanded Mist AI capabilities across Aruba CX switching and Aruba Central, bringing additional self-healing and proactive operational intelligence to campus and branch environments. Together with Apstra's intent-based automation in the data center, HPE now offers a consistent operational philosophy that spans from the campus edge to the AI data center. While the management platforms remain purpose-built for different environments, both reduce manual effort, improve visibility, and help IT teams spend less time troubleshooting and more time delivering business outcomes.

The integration extends beyond technology into HPE's partner ecosystem. Beginning November 1, the Juniper Partner Advantage and HPE Partner Advantage networking programs will merge under HPE Partner Ready Vantage, creating a unified competency framework that reflects the company's single networking strategy. For customers, the message is equally clear: networking is no longer just about speeds and feeds. It is about delivering an intent-based operational model where Juniper Apstra provides the intelligence, automation, and continuous assurance needed to build the next generation of AI-ready data centers.

The broader picture

Three themes cut across all of the Discover announcements and are worth holding onto as you work with customers in the months ahead.

The first is governance as infrastructure. Industry surveys consistently show that fewer than a quarter of enterprises have mature governance models for autonomous AI agents, even as the majority expect agentic AI to be a meaningful part of their operations within two years. HPE's bet is that the customers who get this right will have invested in the governance layer before it becomes a crisis. Russo described GreenLake Intelligence as the intelligence layer for that problem, built around an agentic mesh and a centralized agent registry that applies identity, governance, and policy controls to every agent. Her analogy was direct: you would not onboard an employee without those controls, and agents deserve the same scrutiny. The integrated agent toolkit, OpsRamp's token cost observability, and Zerto's extended data protection for agent rollback round out that answer. Taken together, they represent HPE's push to own not just the infrastructure where agents run, but the operational layer that keeps them accountable.

The second is the VMware displacement opportunity. HPE entered Discover with a credible commercial offer: a unified private cloud portfolio under the PC 1000, 3000, 7000, and AI brand family, all orchestrated through Morpheus 9, with a 25% total partner incentive built from a 10% base and a 15% competitive takeout bonus, plus a migration program offering dollar-one licensing to start workloads moving. European renewal clusters concentrated in Q1 next year create a near-term entry window. The product still needs to mature, and partners are asking for better stability and more referenceable customers, but the commercial packaging is more competitive than it has been at any prior Discover.

That orchestration layer also gained real substance at the show. HPE announced the general availability of Morpheus 9, calling it the biggest release in the platform's history and positioning it as an enterprise control plane for hybrid infrastructure and AI workloads alike. Morpheus Central now delivers a single operational view across sites and regions through GreenLake, offering federated multi-site management as a cloud service today, with an air-gapped on-premises option coming for disconnected environments. Stretched clustering extends resilience across sites for critical workloads. For a VMware-displacement conversation, that is the part that resonates: the argument is no longer only about licensing economics, it is about a common operating model that spans traditional and AI workloads from a single control plane.

The third is that HPE's financial momentum gives this strategic narrative real weight. HPE reported $10.7 billion in revenue for fiscal Q2 2026, up 40% year-over-year. That growth gives HPE the investment capacity to back its architectural commitments and signals to the market that the Juniper integration is not a distraction. It is delivering.

What this means for WWT

The announcements from Discover 2026 do not live in a brochure for WWT. They live in the Advanced Technology Center.

WWT's ATC gives customers a hands-on environment to validate exactly the kind of architecture HPE was describing on stage in Las Vegas. The same converged infrastructure principles, the AI factory reference architectures, the Private Cloud portfolio built on Morpheus, and the Alletra storage tiers are testable, configurable, and demonstrable before a single purchase order is written. For customers navigating a VMware renewal, evaluating HPE's private cloud maturity, or trying to understand whether an AI factory makes sense for their workload profile, the ATC removes the guesswork from the conversation.

Within the ATC, WWT's AI Proving Ground takes that capability a step further. It is where agentic AI architectures get stress-tested against real enterprise workloads in a production-representative environment. The governance and observability questions that HPE addressed at Discover, specifically how you monitor agents, control token costs, enforce policies, and recover from agent failures, are not hypothetical in the AI Proving Ground. They are active design constraints. Customers working with WWT on AI infrastructure strategy benefit from that accumulated operational experience before they commit to an architecture.

HPE left Las Vegas with a coherent story, strong financial momentum, and a partner program that recognized WWT as Solution Provider of the Year. The real work now is translating those announcements into customer outcomes. That is what the ATC was built for, and it is where WWT's differentiation in this market actually shows up.