What 20,000 People Asked After I Said I Don't Pay for AI

September 2026

My post about not paying for AI did 20,000 impressions. I expected arguments about model quality. I got those — plus four other questions, over and over, in the comments and in my inbox.

Here are the five, with honest answers.

1. "Isn't this more expensive than $20 a month?"

In cash, eventually, a little. In structure, no. The API math is $20 a month until it isn't — plan changes, usage tiers, the project that outgrows the allowance two-thirds of the way through. My hardware was already owned: a lab that runs on a 15-amp circuit, electricity that's a rounding error next to the sunk cost. The honest accounting is that the dollars were never the point. The budget I was actually protecting was predictability. Full disclosure: I never put my card in to begin with. I'd spent time on all of them and never found a subscription worth buying — I'd have quit anyway, I just skipped the quitting part.

2. "Is this about privacy?"

Partly, and the boring kind, not the tinfoil kind. Early drafts, client-adjacent thinking, half-formed analysis — nothing that belongs on someone else's training run. Local means my early learnings stay private while they're still wrong. That's not paranoia. That's just good version control for ideas.

3. "But the local models aren't as good, right?"

Right — at the frontier, on the hardest reasoning. For the workloads I actually run every day, the gap doesn't matter. Job-description ranking, cover-letter drafting, RAG over my own career history, site edits, scrape triage: qwen, gemma, and hermes-class models handle all of it locally, fast, with the GPU I already have. I don't need a frontier model to rank thirty job postings. I need one that shows up, offline, every morning at 10.

4. "How much of your life does this eat?"

This is the question people are really asking, and it's the fair one. Dozens of evenings. A GPU watchdog that reset the card mid-inference until I raised the timeout. A context window that silently ran at 4K until I pinned it to 64K. An agent that once reported a file written without ever calling the write tool. None of that appears on the $20 invoice. All of it is the actual price — paid in time, and only worth it if operating the stack is itself the skill you're building.

5. "Would you do it again?"

Yes, for one reason that has nothing to do with money or privacy. I advise organizations on AI adoption. I refuse to do that from vendor slides. Every failure above is now a paragraph I can write from experience instead of theory — what breaks at 2am, what the guardrails miss, what "done" actually costs. The stack is the credential.

So no, this was never really about cancelling anything. It was about where the learning happens. The API teaches you prompting. The lab teaches you operations. If you're leading AI work anywhere in 2026, you need the second one — and you can't rent it for $20 a month.

Which of the five would change your answer — cost, privacy, quality, effort, or the credential?