An internal analysis found that splitting agentic coding work between frontier models and private cloud hardware trims the bill
Hewlett Packard Enterprise says it has found a way to make AI coding agents cheaper to run. An internal hybrid platform cut its token costs by 30-50%, according to a company analysis conducted in October 2026.
HPE’s approach does not involve abandoning the expensive frontier models. It involves using them only when the job actually calls for them.
How the routing works #
HPE’s platform routes work between AI models. Complex work, such as planning and heavy inference, gets sent to frontier models best suited for those tasks.
More common operations run on open-weight models hosted on HPE Private Cloud AI, or PCAI. That is infrastructure the company owns, rather than capacity rented by the token.
The system includes budget controls. Routine tasks get routed to cheaper capacity by default, and complex jobs escalate to the bigger models when needed.
The platform supports more than 150 internal applications.
Beyond the headline savings, HPE is making broader claims about its private cloud. The company says PCAI delivers 40-55% savings on cost per token and a roughly 50% total cost of ownership advantage over major public clouds, based on multi-year comparisons.
HPE CFO Marie Myers said in September 2026 that private cloud deployments achieved up to 60% cost reduction compared with public cloud. There is also a security argument: keeping workloads on owned infrastructure keeps proprietary code closer to home.
A caveat worth noting: these are HPE’s own figures, from HPE’s own analysis, about HPE’s own product. The company sells private cloud infrastructure, so it has every reason to publicize a win for private cloud.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Why AI coding bills are ballooning #
The core problem is how coding agents behave. A chatbot answers once. An agent plans, writes code, runs tests, finds errors, and loops back to fix them.
Each of those steps burns tokens, the units AI providers use to meter and bill usage. A single agentic task can consume 10 to 50 times more tokens than a basic chat prompt.
Usage is also wildly uneven. HPE found that its top 5% of users generate approximately 450 times more token consumption than the lightest users on the same license.
That gap breaks the logic of flat per-seat pricing. GitHub Copilot moved from flat-rate pricing to metered credits.
Uber reportedly burned through its entire 2026 AI budget by April after expanding its use of Claude Code. That is a budget designed to last twelve months, gone in roughly four.
Gartner predicts that AI coding token costs will surpass average developer salaries by 2028.
What this means for enterprises and cloud providers #
For companies deploying AI coding agents at scale, HPE’s results point toward a practical playbook. The question is shifting from which model is best to which model is good enough for each specific step. Most coding agent activity is repetitive: running tests, reading files, making small edits. If those steps can run on cheaper open-weight models, the frontier model bill shrinks to the work that genuinely needs it.
For HPE, the internal results double as a sales pitch. A company that sells private cloud infrastructure demonstrating savings on its own private cloud is a compelling story, if one that buyers will want to verify against their own workloads. There are also tradeoffs that the savings figures do not capture. Running owned infrastructure means capital spending, operational overhead, and the need for in-house expertise to manage open-weight models.
Open-weight models also vary in capability. Routing works only as well as the system deciding which tasks are routine and which need escalation, and a bad call in either direction costs money or quality.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.