Open weights are quietly closing up - and that's a problem

Open weights LLMs have served as a competitive check on frontier lab pricing, but that dynamic is quietly eroding. Meta has stopped releasing open weights for its newest models, Alibaba increasingly gates models behind its API, and others like Kimi and Mistral have added restrictive license conditions. Drawing on contestable markets theory, the author argues that open weights models act like generic pharmaceuticals — keeping frontier prices honest. If the trend toward closed or restricted models continues, a small oligopoly of Western and Chinese labs could capture the enormous consumer surplus AI currently generates, with little competitive pressure to keep prices down. Hardware improvements and distillation are raised as potential mitigations but are deemed insufficient on their own.