Source: Dwarkesh Patel
Anthropic and OpenAI's ability to monetize inference workloads gives them a structural advantage in bidding for scarce training compute. If these labs extract more revenue per FLOP deployed, they can outbid everyone else—including nation-states and Big Tech—for the next generation of chips. This creates a self-reinforcing cycle: early dominance in inference revenue feeds back into dominance in training compute acquisition, which sustains dominance in frontier model capability. The mechanism concentrates AI development among two organizations and forecloses meaningful competition in frontier model development.