HPE cuts token costs 30-50% with on-prem open-weight AI routing

The pitch arrives as per-token pricing turns agentic coding into a budget problem. Uber reportedly exhausted its entire 2026 AI budget by April after expanding Claude Code use, and GitHub Copilot has moved from flat-rate pricing to metered credits. HPE's answer is not to cap usage but to change who owns the capacity: reserve frontier models for difficult planning, and route code generation, test creation and repeated refactoring to open-weight models run on premises.

The economics cut both ways. Rented inference adds to the bill with every call, while owned capacity commits most of the cost upfront and gets cheaper per call as usage grows — though power, operations, maintenance and depreciation still count. HPE's own deployment, spanning more than 150 internal applications, is presented as evidence the model works, with token costs down 30 to 50%.

Usage concentration complicates budgeting. The top 5% of users generated most of the consumption, and if power users grow from a small minority to half the workforce, monthly spend can rise about 4.4 times with flat headcount. That is why HPE ties its metrics to shipped software rather than consumption alone, and it concedes owned capacity is not for everyone: low or unpredictable usage still favors APIs.