2 August 2026Architecture · Cost

AI-generated code is cheap to write and expensive to run

Everyone talks about AI autonomy. Almost nobody talks about what running these systems actually looks like. It looks like continuous supervision.

The models know syntax, patterns and language better than most engineers. What they don’t know is your business model — or the consequence of a decision measured eighteen months out. So someone has to read every contract, catch every drift, and re-orient constantly.

That’s not a complaint. That’s the job. And it is where the value sits.

Vibe coding produces fast, visible results. Look closer and the same pattern keeps appearing: it works today, and it doesn’t hold up — not for maintenance, not for expansion.

Two Lambdas, 7MB each

After review and re-orientation: 250KB. At a few hundred invocations, that’s nothing. At millions, it’s a number someone has to explain.

Trivial to fix, right? Ten minutes each. That is exactly the point.

An ERD with fields nobody ever reads, paid for on every ingress and egress. A schema proposed with 32 tables for email handling alone — rebuilt to 18 that absorb WhatsApp, SMS, social, and whatever comes next, and do it better.

And the one nobody counts: no shared vocabulary. No single naming standard. So the next person — or the next agent — inherits a project they cannot continue.

Why nobody fixes it

None of this fails a test. None of it breaks a build. That is why nobody fixes it.

You don’t notice a dripping tap. You notice the bill — and by then the code is in production and rewriting costs more than the savings ever would have.

I have watched three cycles do this. Client–server. Web. Cloud. Each promised speed. Each left behind a generation of systems nobody could maintain.

AI is the fourth. It is the fastest. The drips arrive at machine speed.

The tool is indispensable — I use it every day. The problem is that generation has outpaced deliberation about consequence. That is where the cost detonates.

So I am curious about your side of it: what is the smallest thing you have caught in AI-generated code that would have cost real money at scale?