1. The network's essential role in sustained AI success

In the opening keynote, HPE CEO Antonio Neri said, "Architecting for AI starts with the network." As enterprises ramp up their AI initiatives -- expanding to more data and more use cases -- the network can very quickly become a bottleneck.

For initial deployments, nearly every server vendor partners with Nvidia to offer AI factory platforms, including Dell Technologies, Cisco, HPE, Lenovo and SuperMicro. As a result, enterprise organizations have a variety of options to get started. But at scale, the network, data and data storage ecosystem become increasingly important to ensure positive ROI.

AI isn't a single workload that can be isolated to an infrastructure silo. AI is a category of technologies that spans a variety of use cases and requires diverse data and infrastructure demands. Enterprise production AI environments often necessitate a hybrid infrastructure environment, on- and off-premises, to achieve success. Although you can start without modernizing your network, at some point it becomes essential to ensure your compute and GPU investments stay fully utilized. The importance of the network for AI likely played a key role in HPE's acquisition of Juniper Networks, which closed last year, strengthening HPE's AI networking portfolio and helping position it alongside Arista, Cisco and Nvidia as leaders in enterprise AI networking infrastructure.

2. Understanding and optimizing tokenomics is vital

AI projects can quickly become expensive. Specifically, as model adoption increases, token costs can rise very quickly. With some organizations strongly encouraging AI usage by internal teams, AI bills can rapidly cripple budgets. In the day-two keynote, HPE CTO Fidelma Russo demonstrated how the shift from prompt-based AI to an agentic, orchestrated AI environment can increase token usage by 250,000 times.

In response, businesses are actively prioritizing controlling token costs, which often requires shifting from being a token consumer to a token producer by owning your own infrastructure. Although the expense for infrastructure can still be high, on-premises deployments can provide better cost control as usage increases.

For on-premises environments, improving your ROI typically requires maximizing the value of your GPU investments. Doing that requires modernizing the surrounding ecosystem, networking, data and data storage to best ensure that your GPUs are continuously optimized and fully utilized across the multiple phases of the AI data lifecycle, including data prep, training and inference. For data preparation, for example, the HPE Alletra MP X10000 offers an integrated data intelligence engine to help reduce the infrastructure investment needed to prep data for AI.

For AI inference, high-performance storage has become a necessity to accelerate the rate at which storage can serve data to the GPU and to enable storage to serve as a key-value (KV) cache offload.

When a language model processes a prompt, it creates new calculations for each output or token. Those calculations are then used to create subsequent tokens. The KV cache within the GPU stores previous calculations to reduce the need for redundant calculations and accelerate new results. The limited capacity of GPU memory restricts how many calculations can be retained and increases the likelihood that the GPU must redo calculations in multiturn conversations.

When a data storage platform, such as HPE's X10000, can expand KV cache capacities using KV cache offload, GPUs can provide a better, faster experience for repeated prompts. In addition, fewer calculations reduce the burden on existing GPUs and lessen the need for additional GPUs as demand scales, thereby reducing costs.

HPE X10000 isn't alone; Dell, NetApp and Everpure have announced similar KV cache offload capabilities. For IT, the ability of the storage to serve as a KV cache offload is just one example of the new requirements IT teams should prioritize when selecting data storage infrastructure for AI. A storage platform's ability to improve AI economics by simplifying data preparation and optimizing GPU utilization, while enhancing the AI experience, should be part of the evaluation process when making new investments. 

3. Embrace AI-based automation to optimize IT personnel

IT operations have long faced increasing complexity. According to survey results from a June 2026 IT modernization report by Omdia, a division of Informa TechTarget, 75% of organizations agreed that over the past two years, IT administrators have faced additional responsibilities to support digital initiatives. The pressure to support new AI projects is only adding to that complexity. The hope for admins is that the integration of AI into infrastructure management, observability and automation tools for servers, networking and storage can significantly offload the burden and free up cycles to work on strategic projects. 

For networking, HPE's Marvis AI Assistant provides an AI-enabled conversational interface for Juniper Networks products and services. HPE announced at Discover that Marvis AI-driven insights and self-healing automation will be extended to HPE Aruba Central, with new agentic reasoning functionality to speed up root cause analysis and remediation.

Beyond networking, HPE GreenLake Intelligence provides AI copilots -- based on OpsRamp technology -- for compute, orchestration and observability. Combined, these tools provide multiple avenues for admins to use AI to simplify their understanding and management of distributed infrastructure environments.

Other infrastructure players, such as Cisco, Dell, Everpure, IBM, Hitachi Vantara and NetApp, also provide integrated generative AI capabilities for infrastructure management. The breadth of HPE's AI operations capabilities, spanning compute, networking and storage, provides real strength, and the potential to further integrate its multiple tools could have a cumulative effect on simplifying distributed IT operations.

4. Optimize legacy operations to free additional budget

Concerns over hypervisor costs and lock-in continue to affect IT strategies. According to Omdia's survey results, 69% of organizations said they are evaluating alternatives to replace their primary hypervisor.

At Discover, the largest crowd I saw on the expo floor was at the booth demonstrating HPE Morpheus VM Essentials, which integrates HVM -- HPE's version of KVM -- to provide a hypervisor alternative for VM-only environments.

More importantly, this week, HPE announced that new VM Essentials customers can receive up to one free year of VM Essentials licenses, a year of HPE Zerto for $1 to support nondisruptive migration to HPE VMs and 0% interest on software through HPE Financial Services. For organizations still paying for their existing hypervisor licenses, migrations can often result in a period of double payment. HPE plans to provide relief. When comparing enterprise hypervisor options, there are often multiple considerations and no exact "like-for-like" replacement options. That said, the cost-saving potential of HPE's offering should make it a consideration.

By providing an in-house virtualization alternative, HPE's strategy differs from that of other infrastructure vendors, such as Dell Technologies, which focus on providing tools to improve flexibility. For example, the Dell Automation Platform helps accelerate the deployment of multiple private cloud blueprints, using technology from Microsoft, Red Hat, Broadcom or Nutanix.

AI presents tremendous potential for business transformation, but unchecked adoption can break your budgets. Success in AI requires optimizing the economics of AI. Overall, HPE Discover highlighted the breadth of innovation across both infrastructure and software -- all designed to optimize production AI environments at scale. The question for HPE moving forward is whether it can build on its portfolio breadth in servers, networking, storage and software to create real, differentiated value by simplifying the entire IT experience or whether this range of competing priorities will limit HPE's overall pace of innovation.  

Scott Sinclair is practice director with Omdia, covering the storage industry.

Omdia is a division of Informa TechTarget. Its analysts have business relationships with technology vendors.