“Sherman Wu believes the cannibalization risk from open-weight models is very low because inference at scale is an extremely difficult technical problem that OpenAI has a specialized team to solve.”