Dan Welble said agents pushed HPE token use up 5 to 10x in one year on The TWIML AI Podcast with Sam Charrington. He named full lifecycle software work as the driver, from requirements to test and support. Chatbots did not create this load. Not financial advice. Markets move fast.
Tokonomics is token discipline that ties AI spend to business value, Welble said.
Model churn makes that discipline harder, with Databricks reporting a new flagship model roughly every week since May 2026. Chat HPE is HPE’s full stack platform that runs in cloud and on prem, Welble said. An AI gateway is a single front door that sees and routes every model call, Welble said. Intelligent routing is software that picks the cheapest model that can still do the job well, Welble said.
That router learns from evals on prompt, app and user use, Welble said. HPE wants to steer 60 to 80% of tokens to on prem implementation to gain cost certainty, Welble said. Cloud remains for elasticity, experimentation and advanced reasoning. Open weight models got much stronger in the last six to nine months, Welble said, which made on prem viable for reasoning heavy coding work.
HPE does not run a token maxing board, though it tracks use by person, org and app, Welble said. A one year roadmap for Chat HPE gave way to a quarterly plan with reprioritization, Welble said. Governance runs in two lanes, with fast review for experiments and deeper checks for production in security, privacy and integration, Welble said.
An MCP gateway is a shared control that standardizes how agents call company systems, Welble said. HPE’s study from about six months ago put the build decision near 84 billion tokens, a level Welble said many firms pass fast in a multi agent world. Big bets are company processes rebuilt around AI and tracked against historic KPIs and unit economics, Welble said. Watch gateway coverage, eval feedback and on prem share before you argue about model price.




