By Tom Whitfield

Hewlett Packard Enterprise reduced the operating cost of its customer-support AI by more than 30 times after moving the workload onto its own on-premises infrastructure, saving nearly $100,000 a month.

The system runs on HPE's GreenLake Intelligence framework of AI agents, paired with Private Cloud AI, an on-premises AI stack engineered with Nvidia. Fidelma Russo, HPE's executive vice president, president and general manager of hybrid cloud, and chief technology officer, disclosed the figures from the keynote stage at HPE Discovery 2026 in Las Vegas.

"It allowed us to govern that really important customer data and it gave us better performance," Russo said. "It also helped us significantly minimize the token spend associated with operating AI at scale. We stopped being consumers of AI and we became producers of intelligence."

Why does inference economics favor on-premises?

Public clouds still dominate AI training, but the economics of inference under agentic AI are pushing enterprises to bring the compute home.

"Once agents have continuous access to data, every interaction consumes a token," Russo said. "That includes every decision, includes some validating the decision, and includes them taking the action. Unlike traditional AI, agents don't stop after one response. They continuously reason, they continuously coordinate, and they continuously interact with other systems."

The arithmetic is unforgiving. Russo cited OpenClaw, a widely used virtual personal AI agent, which processed more than 600 billion tokens in a single month to support roughly 100 continuously operating coding agents. The bill comes to about $13,000 per agent per month.

Validation firm Signal65 puts the token multiplier at 4x to 15x over standard chat interactions. Autonomous agents could eventually drive 1,000 times the inference demand of reasoning AI. Takeaway: agentic AI turns inference into a continuous operating expense that behaves like power and cooling, not like a search query.

What are the other vendors doing?

HPE is not alone. Dell Technologies used Dell Technologies World 2026 in May to push its own on-premises AI expansion. Founder and chief executive officer Michael Dell cited a company study showing 67 percent of AI workloads now run outside the cloud, and 88 percent of surveyed organizations run at least one AI workload in their own datacenter.

Cisco Systems took a similar position at Cisco Live 2026 earlier this month, leaning on its networking heritage. HPE, for its part, is folding the Aruba branch and campus networking lineup into its AI fabric after completing its $14 billion acquisition of Juniper Networks last year.

All three vendors are now selling AI factories, together with software to manage the resulting hybrid environments. Takeaway: the on-premises AI datacenter has moved from niche to mainstream procurement category within two years.

What software is HPE shipping?

Software is where HPE is putting its heaviest weight. GreenLake Intelligence includes a central agent registry that tracks which agents exist, where they run, and what they are permitted to do. OpsRamp Copilot monitors token-based consumption across agents, AI factories, and individual workloads.

The Morpheus 9 release extends that control plane. It adds Morpheus Central for federated management of distributed sites, Morpheus Orchestration Copilot for natural-language provisioning, and integrated software-defined networking.

"Morpheus Central provides a single operational layer across every distributed Morpheus instance, regardless of where it runs or what it manages," Russo said.

Storage is wired into the stack. HPE is integrating Alletra MPX 10000 storage with Private Cloud AI to apply governance and metadata policies automatically. The Alletra Storage X10000 serves as active memory for AI by holding KV cache, which lets systems avoid rebuilding context on every query.

"In AI, memory is not a technical detail or a supply chain challenge. It is a strategic resource," Russo said. The MX 10000, she added, delivers 20 times faster time-to-first-token and 17 times higher throughput in internal testing.

Takeaway: HPE is selling the on-premises AI datacenter as a full stack, with storage, networking, and software wrapped around Nvidia silicon.

What silicon anchors the roadmap?

Private Cloud AI servers will ship with Nvidia's Agent Toolkit, including the OpenShell secure runtime, NemoClaw blueprints, and Nemotron models for designing, building, and orchestrating multi-agent systems.

The upcoming ProLiant DL394 Gen12 server, due early next year, will include Nvidia's new Vera CPUs, designed to support rapid tool calls, orchestration, and real-time data processing for agent workloads.

Cheri Williams, senior vice president and general manager of HPE's private cloud and flex solutions, framed the shift in commercial terms. "There's still a place for experimentation and model training in the public cloud, and you still see customers doing that," Williams said during a panel. "But when it comes to production, on-prem is where most of the enterprise customers are going. The economics don't make sense to be in production in the public cloud."

Takeaway: the same arithmetic that drove HPE's own workload back on-prem is now steering the broader enterprise AI roadmap toward Nvidia-equipped, HPE-branded racks.