The Burn Rate Advantage
AI didn't remove the gate on who codes well. It moved the gate onto who can afford to be wrong the most times before getting it right.
That's the part the "AI democratizes coding" pitch skips over. The pitch holds at one level: someone with no formal training can now describe a feature in plain English and get a working first draft. The floor came up. What the pitch leaves out is that the tilt above the floor simply moved.
The old advantage was skill: years of practice, the pattern recognition that catches a wrong abstraction before you've typed it. That kind of advantage was personal and slow to build, but it was available to anyone with time and a free compiler. Nobody had to buy their way into being good at debugging.
Per-token and per-seat billing put a second gate in front of that one. Coding well with an agent rarely happens in one pass. It comes as a series of attempts: spin up a branch, let the agent try an approach, throw it away when it's wrong, tighten the prompt, run three variations in parallel to see which one survives contact with the actual codebase. That's where the edge lives. Everything before it is just typing.
That loop costs money in proportion to how much you're willing to waste on it. A metered account watches the count. A funded one doesn't. The skill gap between two engineers running the same model might be small. The gap in how many failed attempts each one can afford before they stop exploring and ship whatever half-worked is not small, and it has nothing to do with either engineer's skill.
Call it an exploration tax: a cost that only the metered notice, and one that tracks treasury more closely than it tracks competence. A well-funded team can let an agent thrash on a hard problem for an hour, throw the run away, and try three more framings, because an hour of tokens is invisible against payroll. A developer on a personal plan feels every retry as a real number climbing toward a cap, and that feeling changes behavior before the model ever gets a chance to be wrong: fewer branches attempted, and more willingness to settle for the first plausible answer instead of the fifth, better one.
This stays hidden because the cheap tiers are good enough to produce something that looks exactly like the future being sold. A demo, or a working prototype for a real problem, is reachable on a low-cost plan with patient prompting. That's the case everyone points to. What it leaves out is what happens after the demo, when the work turns sustained and iterative, when getting the tenth attempt right depends on having been willing to pay for the first nine. The tier that makes demos possible and the tier that makes that kind of sustained production possible are not the same tier, and the gap between them is exactly where the democratizing story stops getting checked.
Capital has always bought more attempts at a hard problem than labor could afford alone. What's new is the sales pitch insisting the tool erased that gap, when what it actually did was move the gap onto the retry, a place it wasn't load-bearing before.
The people who can absorb a long burn rate were already a narrower group than "everyone who wants to learn to code." That was true before any of this shipped. What changed is that the narrower group can now buy, with a subscription tier, an edge that used to require actually being better than everyone else at the craft.
Everyone gets the demo. Not everyone gets to keep going after it.