Sora failed due to data ownership

The core issue was not model quality alone, it was data ownership. ByteDance, Google, and Kuaishou could train on their own giant video platforms and learn from both the videos and the engagement signals around them, while OpenAI had to buy access through licensing and strategic partnerships. That made video a more expensive and slower game for OpenAI, because every step toward better training data depended on external rights holders rather than a captive content surface.

• Owning the surface matters because video training is not just raw files. TikTok and YouTube also generate labels from what people watch, skip, replay, and share. That feedback loop helps a model learn what motion, pacing, and scenes people actually respond to, which outside licensing deals do not naturally provide at the same scale.
  1 sacra 3 sacra 4 sacra
• The comparison with Runway shows the alternative. Rather than competing for mass consumer video, Runway paired proprietary model work with editing tools for filmmakers and added licensed datasets through partners like Lionsgate and Getty. That is a narrower but more workable strategy than trying to match platforms that mint fresh training data every day.
  5 sacra 6 sacra 7 sacra
• This also explains why workflow products kept winning share around video. Interviews across the market show that customers rarely want one shot text to video alone. They want transcript editing, shot fixes, avatar tools, clipping, and publishing in one place. OpenAI had the model and app, but not the deeper production workflow that made AI video sticky for repeat use.
  2 sacra 3 sacra 5 sacra

Going forward, the winners in AI video are likely to split into two camps, platforms that own massive native video data and applications that own a concrete workflow. Labs without either advantage will keep paying up for content, distribution, or both, which makes video harder to justify than coding, search, or enterprise AI where the data loop is cheaper to build.