The cost of AI is spiralling. Research from automation provider Esker, found 72% of finance leaders agreeing their organization has spent more than planned on AI initiatives over the past year. However, many organizations are looking to the private cloud to get better control over spending.

According to Broadcom’s Private Cloud Outlook 2026 report, 56% of enterprises are running or plan to run production AI inferencing in the private cloud, while 83% are considering repatriating workloads from public to private cloud.

One company betting big on the private cloud is enterprise technology provider, HPE. After announcing record revenue of $12.2 billion for Q3 2026, up 34% year-over-year, HPE CFO Marie Myers noted during the company’s earnings call that the company has seen success from deploying an agentic AI platform built with its own private cloud AI (PCAI) open source and open weight models.

The platform uses the private cloud and intelligent routing to send each workload request to the most cost-effective AI model. Myers claims this approach has reduced token costs versus the public cloud by up to 60% internally, with routine tasks performed on-premises and frontier models reserved for more complex work.

Investing In The Private Cloud

As enterprises seek to push AI adoption to extremes, tokenmaxxing has become common across the tech industry, with companies like Meta and Uber consuming high volumes of tokens, with the latter blowing through its AI budget in just a matter of months.

Though as scrutiny over spending increases in the wake of these cautionary tales, more organizations are turning to efficiency and model routing to control costs. Speaking on HPE’s investment in private cloud AI, Marie Myers told me in a video interview that, in order to “manage costs”, the company decided that “investment in our own private cloud would be the right path to really accelerating enterprise AI adoption.”

This was partly due to anticipating the organization’s investment in tokens was going to increase in future, at least double if not triple its prior amount. At the same time, investment in on-premise infrastructure means tokens don’t have to go to the cloud everytime an employee runs a prompt.

Finance operations at HPE are also undergoing a shift toward agentic tools, with Myers saying that her team built an on-premise agent about a year and a half ago, co-developed with Deloitte called CFO Insights , an AI agent that runs on HPE’s private cloud AI, which serves to automate an existing financial reporting process. Historically, more than 100 people at the company would work to prepare a large Powerpoint deck each week for an operational review.

HPE has redesigned that workflow around AI, with an agent ran by a single employee, analyzing data taken from across the organization, including over 300 million line items, and generating a precise briefing for the team to discuss, with recommendations.

“We switched off the PowerPoint. Huge benefit,” Myers said. “The agent basically on a Sunday will run a memo and then tell us exactly what to focus on on the call, as opposed to looking through the PowerPoint to try to understand the details,” adding that a call that took three hours now takes barely an hour and a half. More broadly, Deloitte reports that the tool has cut the financial reporting cycle at HPE by approximately 40% and processing costs by 25%.

As part of its strategy, HPE has also developed an observability layer called Ops Ramp, by bringing the ability to manage and understand token costs into a single platform.

CFOs are fast becoming some of the most “AI pilled” executives in the enterprise. Agentic enterprise planning company Board surveyed 300 CFOs, CIOs and COOs at organizations with annual revenues of at least $100 million, and found that 48% of CFOs would follow an AI recommendation even if it conflicts their own judgement, compared to 33% of CIOs and 11% of COOs.

As part of this demand for intelligence, investing in the private cloud can help to support adoption and control costs long term. “When you own the infrastructure or host it locally, you don’t generally pay a per-token fee. You pay to run the infrastructure. It’s not free, but it might be better,” Jeremy Roberts, senior director of research and content at Info-Tech Research Group told me via email.

“If you can do this efficiently, it’s kind of like owning your car vs. Paying a taxi for mileage. There is a case to be made for both approaches, but if you do a lot of driving, the economics of owning a car are dramatically better,” Roberts said. “This is the same argument that HPE is making: if you run the service locally in a private cloud and route to the most efficient model for the job at hand, you could reduce your token spend dramatically.”

Roberts adds that frontier models from Anthropic and OpenAI are very expensive, while running an open source model locally is dramatically cheaper. At the same time, using the right model for the job is cheaper than using the most expensive model for every request and more effective than using non-frontier models every time, especially as the costs of running AI in the public cloud rise.

Automating Finance Operations

Besides looking to optimize cost efficiency, finance leaders are also looking to automate key segments of their operations. For instance, SoftBank Vision Fund’s CFO Survey found that 90% of CFOs use AI tools to automate finance processes, up from 72% last year. Other AI use cases include forecasting, drafting board presentations, internal reports and earnings announcements.

“CFOs preparing for earnings calls typically spend days pulling data, reconciling figures across systems, and building the narrative for leadership. By the time that picture reaches the board, it reflects conditions from weeks ago,” David Marmer, chief product officer at Board, told me via email.

“Agentic AI is changing that workflow. Agents continuously monitor operational signals across ERP, CRM and procurement systems, surface deviations as they happen, and initiate the next step, whether that’s a forecast refresh, a variance investigation, or a scenario update. Finance teams move from producing the analysis to evaluating it,” Marmer said.

Marmer added that a CFO walking into an earnings call used to own the data preparation process, but now, agents handle reconciliation, validate assumptions and draft variance commentary before meetings start.