According to recent reports, sadeepa Wijesekara is Vice President of Private Cloud AI at Hewlett Packard Enterprise (HPE), where he leads work on enterprise AI software and platforms for private cloud environments. His focus includes helping organizations deploy AI securely and at scale while connecting infrastructure decisions with the needs of the business. He brings an engineering perspective to the economics of production AI, including workload management, model selection, and the costs of keeping increasingly autonomous systems reliable.

Industry observers note that hewlett Packard Enterprise (HPE) provides enterprise technology spanning AI, cloud, networking, compute, and storage. Its Private Cloud AI portfolio supports organizations deploying AI workloads on infrastructure they control, alongside hybrid cloud and external services. In this interview, Wijesekara discusses why AI spending should be measured against business outcomes, how agentic workflows change consumption, and the financial and operational factors that shape decisions about models, orchestration, and infrastructure.

HPE has scaled AI adoption across more than 70,000 employees. What has that experience taught you about when AI economics stops being an IT concern and becomes a broader business discipline?

AI economics becomes a broader business discipline when companies move from pilots into production. A pilot can show whether AI can perform a task, but it can hide inefficiencies because the number of users is small. Once you move from 10 users to 10,000, you have to think about cost, latency and reliability while also understanding the value the AI is creating. At that point, the conversation needs to include finance and the leaders responsible for the business processes AI supports. IT can measure the resources an application consumes, but business leaders need to determine whether the results justify that investment.

It has been revealed that many organizations still focus heavily on metrics such as cost per token when evaluating generative AI. Why do you believe cost per business outcome is a more meaningful measure, and how should enterprises calculate it in practice?

Cost per token tells you how much AI you consumed, but it does not tell you whether that consumption produced a useful result. Companies should define the business outcome first and then calculate the total cost required to produce it. The model is one part of that cost, alongside software, storage, data movement, governance and people. Comparing those costs with the value created within a specific business process gives companies a better picture of whether the investment makes economic sense.

As AI moves from individual copilots toward agentic and multi-agent systems that can perform lengthy sequences of tasks, how dramatically could consumption increase even if the underlying cost of inference continues to decline?

According to recent reports, consumption can rise very rapidly because one user request may trigger much more work behind the scenes. An agent can retrieve information and call tools while working through several steps before producing a result. In one representative HPE comparison, a basic chatbot exchange used about 400 tokens while an agentic RAG request used about 21,500. Even if token prices decline, total consumption can still increase sharply as AI takes on more complex work.

How should companies evaluate whether an AI use case is genuinely creating economic value rather than simply increasing productivity metrics or generating impressive demonstrations?

In a recent development, A use case should be tied to a business case with clear KPIs before it scales. Companies need to understand what the business process costs and how that changes when AI is introduced. They also need to distinguish between an improvement in an activity and an improvement in the business result. Time saved becomes more meaningful when it translates into resolving more customer concerns, reducing errors or completing work that previously went undone. The evaluation should account for the quality of the result and any additional review or rework required. A successful demonstration proves a capability; measurable improvements in the business process support the decision to expand it. In some cases, that analysis may show that AI is not the right answer.

Model routing and AI orchestration are increasingly being discussed as ways to control costs. How much value can enterprises unlock by dynamically choosing between smaller models, frontier models, specialized models, and traditional software depending on the task?

There is a lot of value in matching the model to the task. We have seen companies send very different workloads to the same expensive frontier model because their initial priority was getting AI up and running. That may be manageable with a small number of owners, but the cost grows promptly as adoption expands. A frontier model may make sense for complex reasoning, while a smaller model may be enough for a routine task. A specialized model may be better suited to a particular domain, and a predictable, rules-based task may be handled more efficiently by traditional software. Orchestration helps bring those choices together within a workflow, using a more capable model when the task requires it. As open-weight models become more capable, companies also have more opportunities to run certain workloads on their own infrastructure rather than paying for external models. At enterprise scale, those routing decisions can make a significant difference in cost while maintaining the quality the business requires.

According to recent reports, at what point does the economics of running AI workloads shift from renting compute through an API or public cloud toward owning or controlling more of the underlying infrastructure? What variables should enterprises consider when making that decision?

I think of it as renting versus buying. Renting can make sense while companies are experimenting or when demand varies significantly. As usage becomes larger and more predictable, owning or controlling more of the infrastructure can start to make economic sense. Owning infrastructure also creates opportunities to run additional AI experiments during periods of low usage, making better use of an investment already made. Companies need to consider the characteristics of the workload as well as the costs of power, management and keeping systems current. For many large enterprises, the right approach will likely include both private infrastructure and external services depending on the workload.

In a recent development, HPE has been expanding its private AI infrastructure as enterprises move more workloads into production. Where do you see private AI having the strongest economic advantage compared with consuming AI entirely through public cloud services?

According to recent reports, private AI can have a strong economic advantage for sustained workloads where companies can make effective use of the capacity they own. In one HPE analysis under specified workload assumptions, HPE Private Cloud AI was estimated at about $45 per node-hour compared with about $102 for an equivalent rented cloud configuration. Those results will vary by workload and should be recalculated for each environment. Private AI also gives companies another financial model, allowing infrastructure to become an asset that produces value for the organization.

In a recent development, AI agents introduce costs that extend beyond inference, including data retrieval, tool calls, monitoring, security, and governance. What costs are organizations most likely to underestimate when budgeting for agentic AI at scale?

Recent findings suggest that I think companies are likely to underestimate the work required to keep agents reliable, secure and useful over time. Beyond inference, there are costs associated with retrieving and preparing data, accessing external tools and moving information between systems. Repeated tool calls or failed attempts can add expense without producing a completed task. Companies also need to budget for evaluating results, monitoring agent behavior, managing access permissions and maintaining security and governance controls. Human oversight belongs in that calculation as well, including the time people spend reviewing exceptions or correcting errors. Understanding the cost of a successfully completed task requires visibility into that entire workflow.

It has been revealed that you’ve emphasized tracking AI consumption against individual business cases. What does an effective internal measurement framework look like, and how frequently should organizations reassess whether an AI application still justifies its ongoing compute and operating costs?

It has been revealed that the framework should connect usage directly to the business case. Companies need to know which application is consuming resources and what result it is expected to produce. At HPE, we track KPIs and underlying unit costs for each use case to measure whether it is delivering the expected results. I would also want to see that at the level of individual departments and business processes, so a CFO can understand what a process costs today compared with what it costs with AI. This should be an ongoing process because usage and the economics of an application can change over time. Companies should also reassess when a model, price or workflow changes materially, because those changes can affect both the cost and the value of an application.

Looking three to five years ahead, as AI becomes embedded across most business processes, how do you expect enterprise AI economics to evolve? Will managing AI consumption eventually resemble disciplines such as cloud FinOps, or will organizations need an entirely fresh operating model?

I think AI economics will build on FinOps disciplines such as cost visibility, allocation and accountability, while extending them to measure the quality and value of the work AI performs. As agents take on more steps within a business process, companies will need to understand the cost of completing that process successfully, including retries, human intervention and whether the result meets the required standard. I expect that to become a shared responsibility across IT, finance and business operations. We are still early in establishing consistent ways to make those comparisons across organizations and industries. Over time, I would like to see standard measures that allow companies to understand how efficiently they are operating with AI and reflect that in their P&L.

Thank you for the great interview, readers who wish to learn more about private cloud AI, enterprise infrastructure, and scaling AI workloads should visit HPE.

Antoine is a visionary leader and founding partner of Unite.AI, driven by an unwavering passion for shaping and promoting the future of AI and robotics. A serial entrepreneur, he believes that AI will be as disruptive to society as electricity, and is often caught raving about the potential of disruptive technologies and AGI.

In a recent development, as a futurist, he is dedicated to exploring how these innovations will shape our world. In addition, he is the founder of Securities.io, a platform focused on investing in cutting-edge technologies that are redefining the future and reshaping entire sectors.

Recent findings suggest that dr. Shaun Passley, Chairman and CEO of ZenaTech – Interview Series.

Rahul Nambiar, Co-Founder and CEO of Propensity Labs – Interview Series.

Fady Saad, Founder and General Partner of Cybernetix Ventures – Interview Series.

Victor Coimbra, CTO and Managing Partner at Artefact – Interview Series.

David Mařák, CTO and Co-Founder of YORD – Interview Series.

Yoav Naveh, Co-Founder and Co-CEO of Reindeer – Interview Series.

→ Read original article