Nvidia announced an open source alliance for artificial intelligence. The list of participants drew less attention than the list of absentees: neither OpenAI nor Anthropic, the two companies most associated with closed frontier models, appear on it.
The absence is consistent with each company's business model, which is precisely why it's informative. An open alliance led by the company that makes the accelerators exists to widen the set of models that run well on that hardware. For a company selling access to a proprietary model, joining offers less utility and more strategic cost.
In the same week, research found that the lineage behind 69% of available open models was never documented in a traceable way. Most models presenting themselves as open don't let you reconstruct which earlier model they came from, what data they were tuned on, or under what licence.
For anyone putting one of those models into production, that stops being an academic curiosity. Without traceable lineage, there's no way to answer whether the model inherited a licence restriction from an ancestor, or to assess exposure in a dispute over training data of the kind Suno just lost in Germany.
Open has become a word with different meanings depending on who says it: sometimes published weights, sometimes a permissive licence, sometimes just free to download. The three imply very different responsibilities.