Renting Intelligence: Vendor Concentration Risk and the Pricing of AI Dependency
Abstract A firm that puts artificial intelligence into a product must either license models from a provider or train and serve its own. Providers are widely reported to price inference below the cost of serving it, so a firm that rents holds an input priced by another company’s strategy, while a firm that owns has already converted that exposure into capital. Whether equity markets price the difference is a matter of commentary rather than evidence. Classifying the model architecture that United States registrants disclose in their annual reports, I find firms that rent and firms that build indistinguishable on realized volatility, on market beta and on the implied cost of equity. That result is uninformative, and the disclosure is the reason: most registrants who write about artificial intelligence never say where their models come from, and two independent classifications of the same text agree on a registrant’s architecture only about half the time. Dependence on large customers became a priceable attribute because a reporting rule obliged firms to disclose it. Dependence on external model providers carries no such rule, and until it does the exposure cannot be assessed from public filings, by investors or by supervisors.