HPE Discover 2026: Agentic AI and the self-driving network

HPE Discover 2026 reframed enterprise AI infrastructure around a single argument: 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. Across three press releases issued June 16–17 in Las Vegas, the company tied its $14 billion Juniper acquisition to that thesis and shipped the operational scaffolding to back it.

What makes this Discover different from a routine product cadence is where the substance sits. Most enterprise-AI coverage this year has been about getting agents deployed. HPE’s announcements are mostly about what happens afterward — a centralized agent registry, token-based cost observability, policy enforcement at runtime, and a data-protection rewind for agents that go off the rails. That is the day-2 operations problem, and it is the part of the agentic story that has been under-served.

This guide walks through what HPE announced, separates the vendor-stated performance claims from the independently confirmed structural moves, compiles every availability date into one table, and ends with what a marketing or engineering team should actually take from it. Everything below is sourced from HPE’s newsroom releases and contemporaneous independent trade coverage.

For the first time since the $14 billion Juniper acquisition closed, HPE used its flagship conference to explain what the deal actually buys it in the AI era. Antonio Neri’s keynote put the network at the center of the story — the layer that every byte, every token, and every agent decision has to cross. Industry analysts at HyperFrame Research described Discover 2026 as the company’s clearest articulation to date of how Juniper fits its long-term AI strategy.

The argument is more than rhetorical. In a distributed agentic system, agents call models, models query data, and data lives across edge, campus, data center, and AI-factory environments. The network is the connective tissue that determines whether that orchestration is fast, reliable, and observable — or whether it fragments into a sprawl of black-box calls. Rami Rahim, who leads HPE’s networking business, made the dependency explicit.

The centerpiece software announcement is GreenLake Intelligence — an agentic AI framework for hybrid cloud and AI operations. Three components do the heavy lifting: a centralized agent registry, intelligent planning and orchestration, and secure governance controls. The registry is the piece worth dwelling on. It is HPE’s structural answer to the problem that every enterprise scaling agents will hit — knowing which agents exist, what they are permitted to do, and what they are costing.

On the operations side, the OpsRamp Operations Copilot — available now — provides observability purpose-built for AI agents and LLMs: monitoring utilization, governing token-based consumption, and making the cost of agents, multi-vendor AI factories, and workloads legible. HPE paired this with a ServiceNow integration that connects GreenLake Intelligence to ServiceNow’s autonomous AI workforce, aiming for a single source of truth from full-stack observability through to end-to-end autonomous service delivery.

The infrastructure-automation half is HPE Morpheus Software (announced as Morpheus 9). It adds Morpheus Central for federated multi-site management, integrated software-defined networking, and a Morpheus Orchestration Copilot for natural-language provisioning. HPE states that Morpheus SDN is now generally available and reduces provisioning time by up to 60% — a vendor-stated figure with no independent corroboration, so read it as a directional claim rather than a measured benchmark. Citrix also announced a deepened collaboration to deliver Desktop-as-a-Service on GreenLake for on-premises and sovereign deployments, integrated with Morpheus.

HPE’s self-driving network story has a clear arc. A May 2026 announcement moved the concept from vision to reality with autonomous networking capabilities; the June 16 Discover announcement is the production-scale follow-on. The company extended self-driving networks across edge, campus, data center, and AI factories, introducing agentic AIOps innovations including HPE Mist platform support for HPE Networking CX switches and HPE Marvis AI-driven self-healing automation in Aruba Central.

Two new Juniper-derived switches anchor the data-center side. The QFX5140 is designed for inference clusters and edge AI, and the QFX5252 switch tray targets AMD Helios — a scale-up module for AMD’s AI rack-scale platform delivering low-latency, high-bandwidth switching. HPE also announced a new AI-native SASE platform that converges SD-WAN and Security Service Edge in a single management console to accelerate zero-trust adoption and protect self-driving networks, and integrated Mist Networking Data Center Assurance into GreenLake for a unified cross-domain experience.

A precision note matters here, because HPE uses self-driving broadly. The term spans at least three distinct mechanisms: Marvis self-healing automation in Aruba Central, the agentic root-cause-analysis reasoning engine in Mist, and intent-based closed-loop provisioning in Morpheus. They are related but not interchangeable, and the speed claims attached to the Mist reasoning engine — remediation in minutes rather than hours or days — are sourced to an HPE executive, not an independent test. Treat self-driving as a maturity direction, not a single switch you flip.

The HPE–NVIDIA half of Discover focused on Private Cloud AI, the co-engineered stack for running agents in production. The new capabilities center on an NVIDIA Agent Toolkit — including NVIDIA Nemotron open models, NemoClaw, and the OpenShell secure runtime — that functions as an agent operating system, letting enterprises monitor agent behavior, enforce policies, and reduce deployment risk. Private Cloud AI also added multi-node inferencing for up to 256 GPUs, a unified model gateway for governed frontier-model access, and active workload prioritization. These features are slated for July 2026 availability.

The security posture is the more interesting part. HPE announced integration with NVIDIA Confidential Computing to protect models and private data during execution through cryptographic attestation and encryption at every stage, targeted at the HPE AI Factory at-scale and Sovereign AI Factory configurations (availability stated as Q4 2026). And HPE Zerto Software is being extended to identify rogue agent actions and provide continuous data-protection rewind to a clean slate for AI environments — also stated for Q4 2026. That last capability is the clearest signal of where HPE thinks the risk actually lives: not in deploying agents, but in containing the ones that misbehave.

One product needs careful framing. The HPE ProLiant Compute DL394 Gen12, built around the NVIDIA Vera CPU (88 cores, 176 threads, 1.2 TB/s LPDDR5X memory bandwidth, roughly twice that of traditional x86), was introduced for agentic and reinforcement-learning workloads at AI-factory scale, with the New York Stock Exchange named as an early customer exploring the configuration. Its availability inside HPE Private Cloud AI is stated as 2027 — so it should not be conflated with the July 2026 Private Cloud AI features above. It is a roadmap item, not a shipping component of this release.

HPE’s most quotable performance numbers come from integrating Alletra Storage MP X10000 into Private Cloud AI, which lets enterprises automatically apply metadata and governance policies to unstructured data. The headline figures: token response time improved by up to 20x, and token throughput improved by up to 20%. Both are real claims worth understanding — and both are vendor-stated, footnoted in the HPE press release with no independent corroboration at the time of writing. The chart below shows them for what they are: directional vendor figures, not audited benchmarks.

Most coverage handles one HPE press release at a time, so no single article tells you what is actually available now versus what is a 2027 roadmap promise. The table below compiles the announcement and availability dates across all three Discover releases, with the domain each feature sits in. It is the fastest way to separate the shipping capabilities from the futures.

The pattern is clear once it is all in one place. The governance and observability layer — OpsRamp Copilot, Morpheus SDN, Marvis self-healing — is available today. The deeper agent-operating-system and security capabilities cluster in July and Q4 2026. And the most-quoted hardware, the Vera-CPU DL394 Gen12 inside Private Cloud AI, is a 2027 item. Anyone planning a deployment should map their timeline to this distribution, not to the keynote energy.

Here is the through-line worth taking away. Every vendor is talking about deploying agents. The substance of HPE’s Discover announcements is about what happens after deployment — the registry that catalogs them, the observability that prices their token consumption, the policy enforcement that constrains them, and the Zerto rewind that contains the ones that misbehave. That is the day-2 operations problem, and it is the part of the agentic story that most enterprise coverage skips.

Looking forward, this reframes the enterprise-AI buying conversation for the next 18 months. The early-adopter phase rewarded teams that could get an agent into production at all. The next phase will reward teams that can run hundreds of agents without losing track of what they cost, what data they touch, and what they are authorized to do. HPE is betting that observability and governance become the differentiator once deployment is commoditized — and the structure of these announcements suggests that bet is shared across the infrastructure vendors, not unique to HPE. Whatever stack a team lands on, the registry-plus-rewind pattern is the one to evaluate against.

Most readers of this blog are not buying a GreenLake contract. But the patterns HPE is productizing apply directly to anyone running agents on any stack. Here is how to translate the announcements into decisions, by situation.

If you are weighing where agents should actually run — public cloud, on-prem, or a hybrid split — our breakdown of self-hosted vs. hybrid AI agent deployment maps the decision the same way HPE’s Private Cloud AI and air-gapped configurations are designed to address. And if the governance side is where you feel exposed, the enterprise agent platform reference architecture lays out the registry, observability, and policy layers that GreenLake Intelligence is productizing — useful whether or not HPE is your vendor. For a structured self-assessment, our 100-point agent stack readiness checklist covers the same governance, data, and operations gaps across 100 checkpoints.

For teams that would rather build the governance spine into a custom stack than buy a platform, our AI and digital transformation engagements start with exactly this kind of readiness assessment — agent inventory, cost observability, and policy design before you scale. Where the network is the bottleneck, our custom development work connects the infrastructure decisions to the application layer that actually runs the agents.