My thesis here is that we are going to see a bifurcation based on model type. Text based models will become a commodity since so much data is publicly available. The lack of a moat (proprietary data) and stiff competition will drive down margins. As a result, companies will increasingly open source their text based models. Training data for vision models is much more scarce, with stronger copyright laws to contend with. Thus, these will tend to be more proprietary.

• Update 2025-05-30 (PST) (AI summary of creator comment): The Vision leaderboard on LMArena for this market will be sourced from: https://lmarena.ai/leaderboard/vision