Why the panic over Chinese AI models is aimed at the wrong target

In a widely discussed essay published on 20.07.2026, Ben Thompson argues that the anxiety gripping Washington and Silicon Valley over China's open-weight AI models is largely misdirected. Chinese labs are not trying to out-build American frontier models — Beijing benefits simply by commoditizing the model layer itself, since a world of interchangeable, freely distillable models erodes the pricing power of OpenAI, Anthropic, and Google DeepMind while leaving China's lead in robotics and physical-world deployment untouched. Xi Jinping has explicitly tied model openness to AI's migration "from the digital world into the physical world" — a domain China already dominates. Thompson's charge against the frontier labs is that they remain anchored in a training-cost mindset from an era when pretraining, not inference, ate their budgets, an assumption he says no longer holds as agentic use of AI explodes. He proposes two US legal fixes: enshrining data collection for model training as fair use, and voiding terms of service that ban distilling a paid model's outputs, so open-weight American alternatives can compete with Chinese labs on equal footing. The argument reframes the AI race not as a scoreboard of raw model capability, but as a fight over who controls the layer where users and value actually sit.