Situation where a technology infrastructure provider acquires asymmetric bargaining power over a company that has become captive after heavily investing in non-transferable specific assets.
Technological hold-up is an application to the digital sector of the fundamental transformation theory developed by Oliver Williamson in his work on the economics of organizations. Klein, Crawford and Alchian (1978) formalized the mechanism: when a company invests in highly relationship-specific assets (in this context, ultra-calibrated prompts, complex workflows, fine-tuned models), these assets lose much of their value if the relationship is broken. This specificity creates an asymmetric power relationship in favor of the supplier: an initially competitive relationship mutates into a bilateral monopoly where the supplier can extract economic rent without the company being able to change partners without prohibitive cost. In the context of agentic AI, this risk is structural for companies whose operational architecture rests on large language model APIs: a unilateral modification of the neural architecture by the provider can render months of prompt engineering obsolete; the abrupt deprecation of a model (documented practice at OpenAI and Anthropic) forces recalibration of all workflows; the censorship of sensitive business queries through opaque internal policies can interrupt critical processes. This risk is aggravated by the oligopolistic concentration of the leading LLM market, which limits substitution alternatives.
When OpenAI announced the deprecation of GPT-4 Turbo in favor of GPT-4o in 2024, companies that had invested months of engineering in ultra-specific system prompts had to rework their pipelines end-to-end, as response behaviors differed enough to degrade their accuracy rates. The migration cost was not monetary (pricing was comparable) but engineering, perfectly illustrating the hold-up logic through idiosyncratic assets.
hold-up technologique, technological lock-in, dépendance API, Williamson hold-up