Most vibe coders trust their AI to pick the right services. One creator trusted the defaults and got hit with an $800 bill after just two weeks.
This AI professional, Matthew Berman, just shared how a surprise Vercel invoice forced him to rethink the entire state of AI coding. He shipped multiple products during one of his best months. Then the bill landed.
Here’s what happened.
The old way vs the new way
Old way: devs picked services, tuned configs, read the code.
New way: AI picks the stack, AI writes the code, you hit deploy and pray.
The original poster admits he never looked at his Vercel setup. His AI coding agent recommended it, so he used it. By default, Vercel picked the ‘Turbo’ build machine at 12 cents per build minute. The Elastic tier he runs now? 0.3 cents per minute. Same deploys, a tiny fraction of the cost.
Three defaults that drained the budget
The creator broke down exactly where the money leaked:
- 🔧 Turbo build machine instead of Elastic (40x price difference)
- ⚙️ Concurrent builds on by default, so duplicate deploys stacked up
- 🐢 Builds taking 3 to 4 minutes each instead of seconds
He was deploying dozens of times a day. Every rushed fix triggered another build while the last one was still running. All of them billed.
What the fix looked like
After Theo and other devs jumped into his X thread, the post’s author made these changes:
- Switched to the Elastic build machine
- Disabled on-demand concurrent builds (queues instead of stacking)
- Optimized the build process itself, dropping from 4 minutes to about 1 minute
- Moved builds to GitHub Actions and used Vercel only for the deploy step
Now his bill runs a couple of dollars a week instead of hundreds.
The bigger pattern
Here’s the part I found most interesting. The expert zoomed out and pointed to something the whole industry is sliding into.
AI coding agents keep recommending the same services: Vercel, Resend, Fly.io, Railway. Resend just crossed 2 million users, up from 1 million four months ago. These companies are winning the GEO race (generative engine optimization) because AI assistants default to them.
The downside? Builders stop thinking about platform risk, pricing tiers, or uptime. You just accept whatever the model suggests.
Why this matters for your stack
The creator also flagged a deeper issue. Anthropic shipped 13 features in the first two weeks of April. That pace is only possible because humans aren’t reading the code anymore. IDEs like Cursor, Codex, and Claude Code have de-emphasized the code view entirely. The chat is the product now.
His takeaway, which I agree with: knowing the fundamentals still matters. Not to write every line yourself, but to catch the $800 defaults before they catch you.
Takeaways worth saving
- Check default settings on any service your AI picks
- Read the pricing page before shipping, not after
- Queue your builds instead of running them concurrently
- Separate your build step from your deploy step
- Treat platform dependency risk as your problem, not the model’s
Watch the full video for the exact settings he changed and the thread where other devs piled on with optimizations.