Apparent AI-market paradox (2023–present): buyers can usually purchase large volumes of API inference tokens; many advertised token prices have fallen over successive model generations; meanwhile advanced AI compute (training clusters / scarce accelerators) is widely described as capacity-constrained. Is this mainly a real paradox, an exaggeration, or a cat…

Apparent AI-market paradox (2023–present): buyers can usually purchase large volumes of API inference tokens; many advertised token prices have fallen over successive model generations; meanwhile advanced AI compute (training clusters / scarce accelerators) is widely described as capacity-constrained. Is this mainly a real paradox, an exaggeration, or a category error? What best explains any real coexistence of elastic retail token supply / declining unit prices with tightness higher in the stack? Answer requirements - Take an explicit position on real / exaggerated / category error. - Rank explanatory mechanisms by power; do not treat them as equal. - Separate claims about retail API markets vs wholesale/training compute where the distinction matters. - Ground claims in observable evidence categories (pricing schedules, utilization, capex, waitlists, efficiency, competition) without inventing precise figures you cannot support. - Flag the strongest counterargument to your view and what evidence would most change it.