If you’re investing in model companies, or holding their equity — this piece answers the foundational valuation question: will this position still exist in three years?

Put the layer’s strangeness on the table first. In 2026, the model layer simultaneously has: the fastest revenue growth in history, the highest talent density in history, valuations approaching a trillion dollars — and near-zero all-in profit with near-zero customer switching costs.

Traditional industry analysis can’t process that combination. High growth plus high valuation — the textbook says early monopoly. Zero profit plus zero switching costs — the textbook says late-stage perfect competition. The model layer shows both at once because it isn’t an industry at all — it’s a qualifying race. The contestants aren’t burning money for market share; they’re burning it for a position after the race.

The Loser’s Ending: Three Mechanisms Eat the Bare API

Bare-API commoditization runs on three interlocking mechanisms: capability is a reproducible engineering artifact — architectures are public, methods diffuse within months, and distillation makes “teach a cheap model with a strong one” standard practice; switching costs approach zero — a one-line config change; and buyers are systematically organizing the deflation — every large token buyer now routes: easy tasks to cheap models, hard tasks to the frontier. The frontier can only charge a premium on the sliver open source can’t do — and that sliver narrows every quarter.

The bare API’s long-run equilibrium price is plainly visible: hugging the cost of self-hosted open source — which is to say, hugging the cost of compute.

The Winner’s Escape: Two Routes, One Essence

Look at the two leading labs’ revenue structures. By most estimates, OpenAI earns ~65% from subscriptions, only a quarter from API — it’s running Google’s playbook: use capability to win the entry point, use the entry point to collect distribution rent (open source can copy the capability; it can’t copy hundreds of millions of people’s default habit — Bing was good enough for a decade and nobody switched). Anthropic runs the enterprise-workflow route: Claude Code reportedly passed $2.5B annualized by February 2026 with majority enterprise customers — Microsoft’s playbook: embed in the production process and become something whose uninstall cost is ten thousand times the subscription fee.

Opposite directions, same essence: swap the revenue anchor from “model capability” to “lock-in structure.” Capability deflates; lock-in doesn’t. Which is why exploding lab ARR doesn’t refute the sandwich thesis — it’s the strongest evidence for it: if the model layer could keep profit, why would they be in such a hurry to become something else?

The Most-Confused Distinction: API Revenue ≠ Bare API

“Anthropic’s revenue is mostly API — doesn’t that make it a bare-API company?” No. API is a billing channel, not an economic category. Bare or not depends on who holds the lock-in, not on the interface. Revenue through the same endpoint comes in three grades:

API behind your own workflow (Claude Code bills through the API, but uninstalling it means rebuilding your dev process — API in accounting, downstream in economics); API behind someone else’s workflow (selling tokens to the Cursors of the world — lock-in exists, but it belongs to them; the moment the capability gap narrows, you get routed away. That is what “bare” really means: not the absence of lock-in, but lock-in in someone else’s hands); and commodity tokens behind no workflow (margins already hugging compute cost).

The diligence question for any model company isn’t “what’s your API-to-subscription split” — it’s “within your API revenue, what share of the end workflows do you yourself hold?”

Why There Is No Third Ending

The race’s entry fee runs in the billions per generation and rises every generation — and the ticket can’t be paid in installments. Fall one generation behind and your model drops into open-source range; bare-API pricing collapses to the cost line, and you have no lock-in structure to retreat to. There is no “small and beautiful” in the model layer: either you can afford permanent participation, or you are commoditized immediately.

The endgame: two or three successful escapees (for whom the model is a cost center and a quality bar) + an open-source commons (the capability floor — available to all, profitable to none) + a field of commoditized ruins. “The model industry” will disappear as an investment category — the way “the web portal industry” disappeared after 2005.

Three sentences to close. The correct way to value a model company: split it into “a downstream company + a cash-burning model department,” and pay only for the former. Price its revenue in three lock-in columns: own lock-in, others’ lock-in (perpetually discounted — it’s supplying the future routing layer), and no lock-in. And never pay a premium for “capability leadership” itself — that’s an asset whose shelf life is measured in months.

The model company you’ve backed — what share of its revenue sits behind its own lock-in? Comments open.

Data sources (verified, June 2026): OpenAI ~65% subscriptions (multiple estimates); Claude Code >$2.5B annualized as of Feb 2026 (reported; a fast-staling floor, not a current figure); inference margin 38%→mid-60s (SemiAnalysis, 1Q26).

— From Chapter 6 of a book in progress, working title The Deflation Sandwich

#AI #ModelLayer #DeflationSandwich