HPE walked into Discover 2026 with two things to prove: that its AI Factory is more than a reference architecture with a logo on it, and that the Juniper acquisition advances the AI story rather than sitting beside it. On both counts, the company delivered.
That’s not a sentence I write often after a vendor event. These shows usually produce a long list of incremental updates and a keynote that asks you to connect the dots yourself. This one was more coherent. The pieces fit together around a single idea: the enterprise that wants to run AI in production needs a different foundation than the one most companies run today.
And while networking – Juniper – was the star of the show, there was plenty more that HPE rolled out that’s worth exploring.
This was the first Discover since the Juniper acquisition closed, and networking carried the show. The industry has spent the last three years fixated on GPUs, and for good reason. But for large-scale AI, the next bottleneck isn’t compute. It’s connectivity. Moving data between nodes and serving inference at low latency are now the hard problems to solve. And a GPU waiting on the network is an expensive idle asset.
HPE made its case directly. New Juniper QFX switches target inference clusters and scale-up fabrics, with a pretty simple message – keep GPUs processing instead of waiting. Maybe more importantly, folding Juniper into the HPE AI Datacenter Solution turns networking from a component you buy into a layer engineered with the rest of the stack. For a CIO, that’s the difference between integrating a fabric yourself and buying one that’s already validated.
And there’s a deeper point under the hardware where Juniper delivers a lot of value. Building an AI network is the easy part. Operating one is the hard part. AI clusters aren’t static. Models change, traffic shifts, and inference demands move around the fabric faster than manual tuning can keep up with. That’s why Mist, Marvis, and Data Center Assurance matter more than port counts. The real value is running a dynamic AI environment that operates and optimizes autonomously – no additional staff to imperfectly manage.
One thing I don’t want to get lost in this analysis. NVIDIA wasn’t the subject of HPE’s networking pitch, but it was everywhere around it, from Spectrum-X and BlueField to Vera CPUs. While HPE continues to integrate Juniper and deliver a differentiated networking and management stack, it remains tightly aligned to NVIDIA.
Both are true at once. The Juniper differentiation is real, and so is the NVIDIA dependence. It also points to where vendor differentiation is heading. As more of the industry standardizes on NVIDIA silicon, the room to stand out on chips shrinks, and the contest moves up the stack to networking, deployment models, lifecycle, and operations. And this seems to be the layer HPE is targeting.
One of the biggest mistakes I see organizations make is treating AI infrastructure as an extension of the environment they already have. Add some GPUs, stand up a model, and you’re done. But the reality is, almost every part of the stack changes.
AI workloads place different demands on compute, storage, and networking than the applications most enterprise datacenters were designed to support. Data has to be accessible, governed, and usable before it can deliver business value. And infrastructure teams have to think about operating environments that will eventually support hundreds or thousands of AI agents interacting with systems and that data around the clock.
That’s why the conversation is increasingly shifting away from models and toward operations. How do you deploy infrastructure? How do you manage it consistently? How do you govern data? How do you support virtual machines, containers, and AI workloads without creating separate operational silos?
When you think about it this way, HPE’s announcements fit together more clearly. VM Essentials, Morpheus, Alletra, Data Fabric, and the broader GreenLake portfolio address different aspects of the same challenge: building an operational foundation that supports AI alongside traditional enterprise workloads.
The first phase of enterprise AI was largely about gaining access to compute. The next phase is about running these environments efficiently, governing them appropriately, and controlling costs as deployments scale. In my view, that’s the harder problem, and ultimately the one that will determine which organizations successfully move AI from experimentation into production.
Most enterprises now run separate consoles for servers, virtualization, networking, cloud operations, observability, and increasingly AI governance. Stitching those domains together has quietly become as expensive as the infrastructure itself, and AI only adds to the pile, with agents, models, and token costs now on the list of things someone has to watch. That’s why, for me, management is where HPE separated itself most. It doesn’t make headlines, and it matters most to the people actually running the environment.
What I really like about HPE’s approach to server management (and I’ve written about this extensively) is that it starts lower than most people even think to look, in iLO, the company’s BMC. The server’s management controller is the root of trust for the hardware. Compute Ops Management (COM) builds on that and lifts it into the cloud, so provisioning, firmware, compliance, and health move from a rack-by-rack chore to a fleet-level operation. New this cycle, COM integrates with HPE Mist Data Center Assurance, tying server-level and network-level context together. And the efficiency gained is hard to quantify, as provisioning and optimization both get smarter when the system sees both sides at once.
The part I like most is the multi-vendor reach. COM can monitor third-party servers, not just HPE’s own. Insights across HPE, Dell, and Lenovo in a single console are genuinely useful because I don’t know a single enterprise that runs end-to-end with a single vendor. Most management tools quietly assume a single-vendor world and punish you for the reality. Acknowledging the mixed fleet and managing it reframes COM from a ProLiant feature into a hybrid IT control point.
There’s a missed opportunity for HPE in all of this, though. The management story is powerful (leading, even) in aggregate, but the portfolio is a lot to hold in your head. iLO at the chip. COM for the server fleet, with the older OneView now folding into it. OpsRamp for observability. Morpheus for the control plane. GreenLake Intelligence over the top, with its own copilots, plus the FinOps tooling around GreenLake.
Each piece is defensible on its own, but together they’re hard for anyone outside HPE to assemble into a clear picture. If infrastructure management is going to be the moat, and I think it is, HPE has to make it materially easier to understand. Folding OneView into COM is the right instinct, and the portfolio needs more of that. The vendor that wins the management layer won’t be the one with the most tools. It’ll be the one a customer can actually reason about.
Even if HPE hasn’t fully integrated this portfolio (which it hasn’t), it would be well served to present it in a way that’s easier for everybody to understand – not just those who are steeped in the HPE management portfolio. The company is truly missing an opportunity.
Above the hardware sits the layer HPE is betting its narrative on. Morpheus, the cross-vendor control plane, now ships with an Orchestration Copilot that turns multi-step provisioning into natural language while enforcing governance guardrails. OpsRamp’s Operations Copilot now watches AI workloads and the models behind them, not just infrastructure, including token consumption and cost. That point lands harder than it sounds: the cost of running models is becoming a real line item, and most shops have no clear idea what their AI operations actually cost.
Tying it together is GreenLake Intelligence, and it’s worth being precise because the marketing blurs it. It isn’t a single product. It’s a framework for autonomous operations that runs across the stack, with a central agent registry, planning and orchestration, and governance controls. The company’s partnership with ServiceNow extends this from full-stack observability into autonomous service delivery. So what does this mean? A “signal” can move from insight to action with a human in the loop where it counts – or not!
This is where I’d place HPE on the leading edge. Not because the demos are flawless, but because the architecture is right. An agent registry, governance, and observability built for AI workloads and their token costs, along with a control plane spanning VMs, containers, and multi-vendor hardware, is the right shape for where operations are heading.
The tension is that every vendor is racing. And building copilots is no longer the hard part. The hard part is getting enterprises to delegate real operational authority to software, and that kind of adoption tends to lag the technology by years.
I think HPE has built the right framework. The behavioral and cultural change on the customer side is the challenge.
Amid all the AI hype, the VMware discussion seems to have been pushed out of focus. But it’s real, and October 2027 is approaching in terms of a virtualization migration. While Broadcom’s licensing changes prompted many organizations to re-evaluate a platform they had treated as permanent, HPE quietly introduced VM Essentials (VME). This is a per-socket virtualization layer that competes head-on with vSphere, with HPE citing up to 90% cost reduction against legacy per-core licensing. And to help customers, it’s paired with a migration program built to avoid paying for two platforms at once. If I’m a CIO looking to modernize for my AI future, this is an exit ramp.
But really, the question isn’t about replacing VMware. It’s deciding what comes next. A CIO has a handful of options: stay put and absorb the cost, swap the hypervisor for something like VM Essentials or Nutanix, go cloud-native on Kubernetes, or run some hybrid of all three. Most large enterprises will land on the hybrid, because they already live there. That’s the opening, and it’s a bigger one for HPE than any single AI announcement at the show.
This is where Morpheus matters, and it isn’t the hypervisor. It’s the cross-vendor control plane that orchestrates VMware, HPE’s own KVM-based virtualization, Red Hat, and Kubernetes from a single place. The value isn’t swapping one hypervisor for another. It’s giving a CIO a single operating model over the mixed estate they already run. VM Essentials enters a crowded field with Nutanix, Red Hat, and cloud-native options. Morpheus is where HPE can more cleanly differentiate, because few competitors offer a vendor-neutral control plane rather than another platform to standardize on.
Strip away the product names, and the CIO question is pretty simple. Can I run AI in production, under control, without re-platforming my estate? And without all of that unpredictable cost? HPE’s answer was a credible yes. And the proof is operational, not just architectural. Anyone can resell GPUs. Fewer can credibly manage and govern a hybrid, multi-vendor, AI-heavy estate from edge to core.
I think this is a strong position to argue from.
OK, so the HPE strategy is coherent and, in places, ahead of the field. But much of this rolls out over the next twelve to eighteen months, and the autonomous operations pieces have to prove themselves in the real world. And cleanly unifying the Aruba and Juniper networking stacks is real, ongoing work, not finished.
But the direction is right. The industry spent the last several years treating AI infrastructure as a silicon problem. HPE sees the next phase as an operations problem, and I think that bet is correct. If this vision is right, the long-term winners won’t be decided by who ships the most GPUs. They’ll be decided by who can make increasingly complex AI environments consumable and manageable at scale. And consumable and manageable for the commercial enterprise as well.
HPE has assembled most of the pieces required to compete in that future. The challenge is execution. AI infrastructure is becoming more distributed, more heterogeneous, and more difficult to manage. Enterprises don’t need more complexity. They need a way to consume and operate that complexity without having to become experts in every layer of the stack. The vendors that solve that problem will define the next phase of enterprise infrastructure.
HPE is positioning itself to be one of them.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。