Vol.019 — How Could Anyone Spend That Many Tokens?
Date: 2026-05-04 | Newsletter
Key Summary
Uber rolled out Claude Code to its engineering team in December 2025 and burned its entire 2026 AI budget by April — with 95% of engineers using AI monthly and 70% of committed code now AI-authored. Zenta reads this not as failure but as a success that the budget didn’t anticipate: “adoption-driven cost overruns are the new failure mode of success.” The structural danger he flags for ASEAN owners is the instinct to respond by locking into bundled multi-year vendor contracts — the same pattern Functional AI Partners calls “marrying the LLM.” Against the Uber headline, Zenta holds up a quieter story: an Indian developer whose AI product flopped while his boring website-with-7-day-delivery service now earns more than his old annual salary, with the insight that “risk removal beat AI sophistication.” The playbook Zenta distils for ASEAN: spend on what’s measurable in 30 days, stay vendor-portable, and bet on boring simple agents that solve one bounded problem well. Do it simple. Bet on boring. Stay vendor-portable.