Strategy

HPE's Private Cloud AI Runs at $45 Per Node-Hour vs $102 Rented

HPE pegs Private Cloud AI at $45 per node-hour vs $102 for rented cloud. VP Sadeepa Wijesekara explains why token costs mislead and agents multiply bills 50-fold.

By Nathan Brooks

4 min read

Sadeepa Wijesekara, Vice President of Private Cloud AI at HPE – Interview Series
Sadeepa Wijesekara, Vice President of Private Cloud AI at HPE – Interview Seriesblickpixel / pixabay

What's News

  • HPE analysis estimates Private Cloud AI at ~$45 per node-hour vs ~$102 for equivalent rented cloud configuration
  • An agentic RAG request consumed ~21,500 tokens vs ~400 for a basic chatbot exchange in an HPE comparison
  • HPE has scaled AI adoption across more than 70,000 employees
  • Sadeepa Wijesekara is Vice President of Private Cloud AI at Hewlett Packard Enterprise
  • Wijesekara expects AI economics to build on FinOps disciplines across IT, finance and business operations

HPE puts the cost of its Private Cloud AI at roughly $45 per node-hour, against about $102 for an equivalent rented cloud configuration, according to one internal HPE analysis under specified workload assumptions. That 55% gap is the sharpest number in an argument Sadeepa Wijesekara, Vice President of Private Cloud AI at Hewlett Packard Enterprise, makes about how enterprises should think about production AI spending. His warning to practitioners: the metrics most companies track won't tell them whether their AI investment actually pays.

HPE has rolled out AI internally across more than 70,000 employees, and Wijesekara says that experience taught him exactly where AI economics stops being an IT line item. "AI economics becomes a broader business discipline when companies move from pilots into production," he said. A pilot with 10 users hides inefficiencies. At 10,000 users, cost, latency and reliability all become boardroom questions, and finance has to join the conversation.

Why cost per token misleads buyers

Cost per token tells you how much AI you consumed, not whether that consumption produced anything useful. Wijesekara's alternative: define the business outcome first, then add up everything it takes to produce it. The model is only one component — software, storage, data movement, governance and people all belong in the denominator. Compare that total against the value created in a specific process, and you get an answer about whether the investment makes economic sense.

What do agentic workflows do to the bill?

They can blow it up. One user request to an agent triggers tool calls, retrievals and multi-step reasoning before anything reaches the user. HPE ran a representative comparison: a basic chatbot exchange consumed about 400 tokens, while an agentic RAG request consumed about 21,500 — a roughly 50-fold increase. Falling inference prices don't rescue the budget when total consumption scales that fast.

That's before the hidden costs. Wijesekara says companies most often underestimate "the work required to keep agents reliable, secure and useful over time." Budget lines that get skipped include data preparation, external tool access, repeated or failed tool calls, monitoring, access management, security controls, and the human time spent reviewing exceptions. Understanding the cost of a completed task means visibility into the entire workflow.

Routing models and the rent-versus-buy call

Wijesekara has seen companies send wildly different workloads to the same expensive frontier model because their first priority was simply getting AI running. That works at small scale and fails at scale. His framework:

  • Frontier models for complex reasoning
  • Smaller models for routine tasks
  • Specialized models for particular domains
  • Traditional software for predictable, rules-based work

Orchestration ties those choices together, and as open-weight models improve, companies get more chances to run workloads on infrastructure they already own.

On rent versus buy, he's pragmatic. Renting makes sense while experimenting or when demand swings hard. Once usage grows large and predictable, ownership starts to pay — and owned infrastructure can absorb extra experiments during low-usage periods, squeezing more out of sunk capital. "For many large enterprises, the right approach will likely include both private infrastructure and external services depending on the workload," he said. Private AI also changes the accounting: infrastructure becomes an asset producing value, not just an expense.

Measuring like a CFO

At HPE, the internal framework tracks KPIs and underlying unit costs for each use case, broken down to the department and business-process level. The goal is a comparison any CFO can read: what the process costs today versus what it costs with AI. Reassessment isn't annual — it happens whenever a model, price or workflow changes materially, because those shifts move both cost and value.

Wijesekara is also blunt that the analysis can kill a project: "In some cases, that analysis may show that AI is not the right answer." Time saved only counts when it converts into more resolved customer issues, fewer errors or work that previously went undone.

Where this heads in three to five years

He expects AI economics to build on FinOps disciplines — cost visibility, allocation, accountability — while extending them to measure the quality and value of the work AI performs. As agents take on more process steps, companies will need to know the full cost of completing a process, including retries and human intervention, as a shared responsibility across IT, finance and business operations. Consistent cross-industry measures don't exist yet. Wijesekara wants to see standard efficiency metrics that companies can reflect directly in their P&L — the practical signal that AI economics has matured from experiment to discipline.

Source: unite.ai

Share this article:

More from Nathan Brooks

Nathan Brooks

Show full bio

News editor covering regulation, policy and their impact on entrepreneurship, startups and venture funding at Business Bearings.

2,274 articles

Related articles

« Previous article