The Endpoint Inversion: Why the Work AI Does and the Revenue AI Providers Can Collect Are Diverging

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Critics of very large AI market forecasts point to low willingness to pay for AI services. This paper argues that the same observation supports a different conclusion. Mass AI adoption need not require mass incremental spending on AI services if capable intelligence diffuses through the ordinary replacement cycles of consumer devices and enterprise computing infrastructure, a process the paper calls refresh-mediated adoption. Under this mechanism, the economic value of AI work can become very large while the share of that work generating revenue for frontier-model providers declines. The paper separates three quantities that the debate treats as interchangeable (economic value, adoption and externally monetized revenue) and expresses frontier revenue as R_F = W x mu x f x p: total AI work, the share performed through externally operated AI services, the frontier share of that externally monetized work, and realized frontier price. Evidence that f and p are already moving is direct: Ramp reports a 41% fall in the effective price of AI tokens since March 2026 and company-wide defaults that steer employees away from frontier models, and Vercel reports that open-weight models processed 56% of its gateway tokens in August 2026 while earning 14% of spend. Evidence that mu could move is earlier and rests on product commitments and distribution signals. Enterprises, the paper argues, route work to the cheapest trusted tier capable of performing it, so hosted models that are cheaper per token do not settle where work runs. The paper shows why existing market instruments will read this diffusion as a slowdown, quantifies the revenue growth current plans require under plausible price declines, and specifies prospective tests.