June 24, 2026
By Bob O'Donnell
Taking an established company in a new direction is always a challenging task. Doing so in the midst of one of the biggest technological evolutions the tech industry has ever witnessed, even more so.
Yet, that’s what HPE made clear at its recent Discover event in Las Vegas. In the throes of the rapid shift to agentic AI (enabled through hybrid AI architectures in the enterprise), HPE’s message was clear: the future opportunity for differentiation and value in these enterprise environments is through networking and everything that it enables. To put it a different way, instead of concentrating on competing against Dell Technologies, the company seems to be more focused on competing with Cisco.
For long-time observers of the company, this development was a bit surprising at first. After all, the company’s heritage is as the server and enterprise infrastructure part of HP that broke off on its own when HP split into HP Inc and HPE just over 10 years ago. Additionally, like its competitors, much of HPE’s recent success has arguably been due to the quickly growing opportunity for on-prem AI infrastructure. However, this was the first Discover event where the company’s Juniper acquisition was fully complete, and that company’s influence on the overall message and direction of HPE was clearer at this event than it’s ever been.
In part, this is likely due to the fact that Juniper brought with it an extensive set of unique technology IP, custom silicon, and products that the combined company can use to generate higher margin products and services. In fact, one of the biggest themes from the show was the company’s focus on what it called truly autonomous, self-driving networks that can not only predict and warn about potential network issues but also leverage AI to perform any necessary fixes automatically.
By leveraging Juniper’s Mist technology, as well as its custom LEM model with 10 years of application-layer telemetry network data, and Marvis Minis (which are digital twins that simulate user experiences and proactively prevent issues), the company showed how this AIOps architecture can significantly enhance network performance and stability. The company’s newly enhanced SASE (Secure Access Service Edge) capabilities also integrate security capabilities into the autonomous network. As expected, HPE also highlighted how it has combined the capabilities of Aruba Central’s wireless network management capabilities with the wired managed solutions from Juniper Mist as part of this new self-driving network.
In addition to the pure networking capabilities the company discussed, another reason for the shift in focus is how AI computing systems are being built. While individual racks offer a lot of capabilities, it’s the multi-rack systems-level approach to computing that is quickly winning the day. At Nvidia’s GTC, for example, the message was less about individual chips and racks and more on the system-level capabilities of multiple racks linked together into a more powerful whole. HPE seems to be building on these developments as well as the incredibly fast developments in “tokenomics,” or the creation and consumption of tokens from AI factories in enterprise environments.
The context is that companies are realizing that the tokenomics argument is driving organizations to make large investments in their own data centers and create hybrid AI environments. The basic logic is relatively simple. The entirely new cost basis of AI-based token consumption is quickly getting out of hand and changes need to be made to get it under control. Developments like tokenmaxxing, where individuals are using (or encouraged to use) as many AI tokens as possible and leverage today’s impressively powerful LLM for increased productivity, are adding to the challenge. This is particularly true with agents, which are often set into looping workloads that consume significantly more tokens than human beings could on their own.
Plus, much of the work is being sent straight to cutting edge LLMs, which are the most expensive per token, even though most of it doesn’t require that level of power. The net result is that many companies are reporting that they’ve spent through an entire year’s worth of AI token budget in a quarter or less.
Thankfully, important developments that enable even frontier models to run in on-premises environments are opening up the opportunity to approach the token creation and consumption model in an entirely new way. Hybrid AI architectures—a topic that TECHnalysis Research researched and reported on last year in its “The Future of AI is Hybrid” report—are being developed to allow companies to split the token generation capabilities between locally owned computing resources and the cloud-based tools currently being used. In some organizations, some of the token generation is even being split off to client devices, enabling a potential three-way split of the task.
While the details of how to most efficiently break up and orchestrate AI and agentic workloads across these two or three sets of computing resources are still being refined, even early hard-coded efforts are showing positive signs. If just 30% of tokens can be generated locally, that translates directly into a 30% savings on token costs, making the ROI case for on-prem infrastructure extremely easy and compelling.
In the context of all this, HPE announced a number of important new AI-related developments at Discover. New Nvidia-powered computing solutions that HPE debuted at the show included the Proliant Gen12 server built around Nvidia’s new Arm-based Vera CPU, as well as support for Nvidia’s new NemoClaw agentic software tools. At the same time, Nvidia’s overwhelming dominance can make it challenging for companies like Dell, HPE, Lenovo, and SuperMicro to really differentiate their AI factory offerings—yet another reason why HPE could be shifting its focus toward networking.
To that end, the company also highlighted some new Juniper-powered wired networking switches and routers specifically designed for agentic AI-focused computing racks. One of the more interesting announcements was for the QFX5252 switch, specifically optimized for AMD’s forthcoming Helios rack architecture, which many organizations hope will provide a viable alternative to Nvidia-powered systems.
Recognizing that while AI workloads are clearly the sexiest part of the enterprise computing world these days, they are far from the largest, HPE also showed a number of developments focused on more traditional enterprise applications. The latest version of the company’s Morpheus software, for example, provides a virtualization-based alternative for companies looking to break away from the large price hikes that Broadcom has placed on existing VMware deployments. HPE also showed off some impressive new additions to its other software tools, such as GreenLake Intelligence, that can be used to track and manage all types of workloads, including AI-powered agentic ones.
The worlds of enterprise computing and AI infrastructure are quickly coming together in some interesting and novel ways, and companies like HPE that are serving these environments are having to adjust rapidly as well. HPE’s latest Discover announcements highlight that the company is working to make those adjustments and focus its efforts in the areas where they can make the biggest impact. As with all the latest developments in enterprise computing, how these all play out over the next year or two is going to be very interesting to watch.
Here’s a link to the original column: https://www.linkedin.com/pulse/hpe-focuses-network-side-agentic-ai-bob-o-donnell-bpsyc
Bob O’Donnell is the president and chief analyst of TECHnalysis Research, LLC a market research firm that provides strategic consulting and market research services to the technology industry and professional financial community. You can follow him on LinkedIn at Bob O’Donnell or on Twitter @bobodtech.
| 
