Descriptions:
Matthew Berman details how a two-week vibe-coding sprint resulted in an unexpected $800 Vercel bill, walking through each cost driver and the specific settings changes that brought his spending back under control. The culprits: Vercel’s default Turbo build machine at $0.12 per build minute (versus $0.003/min on the Elastic tier), concurrent builds triggered by rapid AI-assisted iteration, and bloated build pipelines taking three to four minutes when they should take seconds. After changes — switching tiers, disabling on-demand concurrent builds, and offloading build steps to GitHub hooks — Berman’s builds dropped to roughly one minute each and his costs fell to a few dollars per week.
The deeper argument the video makes is about what vibe coding does to developer judgment. Berman describes how AI coding assistants consistently recommend the same services — Vercel, Fly.io, Railway, Resend — without prompting any review of pricing tiers, configuration defaults, or platform dependency risk. He introduces the concept of GEO (Generative Engine Optimization) to explain why services like Resend grew from one million to two million users in four months: AI agents are becoming the primary recommender layer for developer tooling decisions.
The video is a practical cautionary tale for anyone building and deploying with AI assistance at speed, covering both the specific Vercel settings to audit and the broader habit of letting AI make infrastructure decisions without human review.
📺 Source: Matthew Berman · Published April 17, 2026
🏷️ Format: Workflow Case Study
