Access to capability × Deployable task coverage

The takeoff story is a distraction: the real weight of this corner isn't compounding capability at all. One moat is liability — when output is dependable but scarce, whoever underwrites "this is correct" owns the market, so the chokepoint migrates from the compute holders to the indemnifiers and reinsurers. The other is permission — a gatekeeper's deployment choices become de facto industrial policy nobody voted for, and state and firm end up each other's hostages. Either way, the corner's stability is the actual hazard: reliable output starves the verification economy a chokepoint would normally spawn, so no aggrieved constituency ever forms to contest it.

Universal capability doesn't end scarcity, it relocates it to everything capability needs in order to act — energy, compute siting, capital, land, the legal right to sign. Politics reorganizes around permitting fights and occupational licensing as the last moat; guilds return not because AI can't do the work but because the law still requires a human signature, and the income-tax base quietly erodes as employment becomes optional for firms that dissolve into one person orchestrating an agent swarm. The institutions keyed to that disappearing relationship — pensions, employment law, health insurance — lose their foundation without anyone voting on it.

Elite access to a tool that only half-works looks like ordinary hype, which is exactly why nobody watches it closely enough. The gatekeepers' real asset isn't the capability, it's the narrative of the capability — they're the only ones who can substantiate or refute claims about how well it works, and they have every incentive to inflate rather than admit the limits. A small priesthood of prompt-whisperers and integration specialists forms around the brittleness, and because documenting the failure modes would cost them their job security, they have a standing incentive to keep the tool's quirks mysterious rather than fixed.

Cheap generation without matching verification doesn't create abundance so much as re-create the chokepoints it was supposed to dissolve — one layer up, at whoever owns trust: identity providers, provenance registries, notaries, in-person reputation. The inversion nobody predicted: brittle AI is good enough to flood every application channel — jobs, grants, courts, admissions — but not good enough to triage the flood, so institutions defend themselves by raising the cost of applying, which regressively excludes exactly the outsiders cheap access was supposed to empower. And because nothing here can adjudicate what's real, people don't get more skeptical, they retreat to tribal source-trust faster than they would under a genuinely capable AI that could at least settle facts.