HPE’s private AI economics: $45 per node-hour, not just tokens
HPE says Private Cloud AI can cost about $45 per node-hour versus $102 for rented cloud, but Sadeepa Wijesekara warns that agents and hidden operating costs can erase apparent savings.

Why it matters
- HPE estimates Private Cloud AI at about $45 per node-hour versus about $102 for an equivalent rented cloud configuration under specified assumptions.
- A representative HPE comparison used about 400 tokens for a basic chatbot exchange and about 21,500 for an agentic RAG request.
- HPE has scaled AI adoption across more than 70,000 employees.
- Wijesekara cites a scale shift from 10 users to 10,000 as the point where cost, latency and reliability require business-level planning.
- Wijesekara expects standard AI-efficiency measures to reach company P&Ls over the next three to five years.
By Sophie Lindqvist
HPE estimates that its Private Cloud AI costs about $45 per node-hour, compared with about $102 for an equivalent rented cloud configuration under one set of workload assumptions. Wijesekara called this a workload-specific estimate, not a universal price, and said enterprises should recalculate it for each environment.
Sadeepa Wijesekara, Vice President of Private Cloud AI at Hewlett Packard Enterprise, leads the company’s work on enterprise AI software and platforms for private-cloud environments. He focuses on secure deployment at scale and connecting infrastructure choices with business needs, including workload management, model selection and reliability for increasingly autonomous systems.
HPE says it spans AI, cloud, networking, compute and storage. Its Private Cloud AI portfolio supports deployments on infrastructure customers control, alongside hybrid cloud and external services.
For Wijesekara, the comparison matters because it moves AI infrastructure into the cost-and-value conversation that finance leaders must manage. His forecast is simple: as agents take on more steps, companies will need to see the cost of completing a process successfully and carry that result into the P&L.
What does HPE’s $45 estimate actually show?
HPE has been expanding its private AI infrastructure as enterprises move more workloads into production, Wijesekara said. He said private AI can have a strong economic advantage for sustained workloads that make effective use of capacity the company already owns.
Utilization drives the advantage. Owning infrastructure lets companies run extra AI experiments during periods of low usage, turning the purchase into an asset that can produce value beyond its first assigned workload. Private AI earns its place when owned capacity stays busy and supports a measurable process.
When do AI pilots become a business discipline?
HPE has scaled AI adoption across more than 70,000 employees, Wijesekara said. He said a pilot can show that AI performs a task, but a small user base can hide inefficiencies.
Moving from 10 users to 10,000 forces teams to consider cost, latency and reliability at the same time. The conversation must then include finance and the business leaders who own the processes AI supports.
IT can measure the resources an application consumes, while business leaders decide whether the results justify the investment. AI economics becomes a business discipline at that point.
Why does cost per business outcome beat cost per token?
Wijesekara put the accounting principle plainly: “Cost per token tells you how much AI you consumed, but it does not tell you whether that consumption produced a useful result.” He recommends defining the business outcome first, then calculating the full cost required to produce it.
That total extends well beyond the model:
- The model used for inference.
- Software, storage and data movement.
- Governance and the people needed to operate the system.
Companies should compare those costs with the value created inside a specific business process. That comparison shows whether the investment makes economic sense, even when token prices look attractive. The useful unit of AI spending is a completed business result, not a token count.
How much more work can one agent request consume?
As organizations move from individual copilots toward agentic and multi-agent systems, one request can trigger a lengthy chain of work. An agent may retrieve information, call tools and work through several steps before returning a result.
In one representative HPE comparison, a basic chatbot exchange used about 400 tokens. An agentic RAG request used about 21,500 tokens.
Wijesekara said total consumption can still rise sharply as agents take on more complex work, even when inference prices fall. Token discounts do not guarantee a smaller bill when each request triggers more steps. Agent economics depends on the full chain of work, not the sticker price of a token.
What should prove an AI use case creates value?
Wijesekara said enterprises should attach a business case and clear KPIs to a use case before scaling it. Teams need to know what the process costs today and how AI changes that cost.
He also separates an improvement in an activity from an improvement in the business result. Time saved matters only when it resolves more customer issues, reduces errors or completes work that previously went undone.
Evaluation must include result quality and any extra review or rework. Wijesekara said a successful demonstration proves a capability, while measurable business-process improvements support expansion. He added that some analysis will show AI is not the right answer. Scale only what changes the underlying business result.
How can model routing control enterprise AI costs?
Wijesekara said his team often sees companies send very different workloads to the same expensive frontier model. Their early priority was getting AI running, a choice that works briefly with few users but becomes costly as adoption grows.
His task-to-tool map is straightforward:
- Frontier models for complex reasoning.
- Smaller models for routine tasks.
- Specialized models for particular domains.
- Traditional software for predictable, rules-based work.
Orchestration brings those choices into one workflow and calls a more capable model only when the task requires it. As open-weight models become more capable, Wijesekara said companies gain more chances to run suitable workloads on infrastructure they control.
At enterprise scale, those routing decisions can materially change cost without sacrificing required quality. Route each task to the cheapest tool that still meets the quality bar.
When should enterprises rent compute and when own it?
“I think of it as renting versus buying,” Wijesekara said. Renting fits experimentation and demand that varies significantly, while larger and more predictable usage can make ownership or control economically attractive.
Owned infrastructure offers another advantage: companies can run additional AI experiments during low-usage periods and get more value from an investment they already made. The trade-off includes power, management work and the cost of keeping systems current.
Companies must also weigh each workload’s characteristics. For many large enterprises, Wijesekara expects a mixed approach that combines private infrastructure with external services rather than forcing every job onto one model. The buying decision follows workload stability, utilization and operating burden.
Which costs do agent budgets usually miss?
Wijesekara believes teams most often underestimate the effort required to keep agents reliable, secure and useful over time. Inference represents only one part of the bill, and the rest sits across the workflow.
- Retrieving, preparing and moving data between systems.
- Accessing external tools and making repeated calls.
- Recovering from failed attempts that produce no completed task.
- Evaluating results and monitoring agent behavior.
- Managing access permissions, security and governance controls.
- Human review of exceptions and correction of errors.
Teams need visibility into the full sequence to calculate the cost of a successfully completed task. Failed attempts and rework can consume money without creating the intended result. Count every failed retry and human correction, not just the final inference call.
What should an internal AI measurement framework track?
Wijesekara’s framework links resource use to the business case for each application. Teams should identify which application consumes resources and what result that application should produce.
At HPE, he said, the company tracks KPIs and underlying unit costs for each use case. He wants that view at the department and business-process level, where a CFO can compare today’s process cost with its cost after AI enters.
The review must continue because usage and application economics can change over time. Wijesekara also recommends reassessment whenever a model, price or workflow changes materially, because each can alter cost and value. Measure AI against a named business result, then measure it again.
What changes in the next three to five years?
Looking three to five years ahead, Wijesekara expects AI to sit inside most business processes and AI economics to build on FinOps. He expects those practices to extend beyond infrastructure consumption toward the quality and value of work agents perform.
As agents take more steps, companies will need the cost of completing a process successfully, including retries, human intervention and whether the result meets the required standard. That responsibility will sit across IT, finance and business operations.
He said consistent comparison methods remain immature across organizations and industries. Over time, he wants standard measures that show how efficiently companies operate with AI and carry that efficiency into the P&L.
Wijesekara expects AI efficiency to become a P&L measure. Until common standards exist, each enterprise must calculate the full cost of a completed task.
Source: unite.ai
More from Sophie Lindqvist
Show full bio
Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.
749 articles
Related articles
- HPE: The Enterprise AI Question Is Now Cost Per Outcome, Not Per Token
- Filtering Data Before AI Cuts Security Costs 79%, WitFoo Says
- HP Scales OpenAI Frontier Partnership After Pilot Wins
- Cloudflare Brings OpenAI's GPT-5.4 to Agent Cloud for Enterprises
- Private AI Needs Infrastructure Built to Evolve, Not Just GPUs