I think another third option is that perhaps there aren't many people with the skills to run large teams for both research and product. I suspect that DeepMind / OpenAI / Anthropic require infrastructure and tooling that's significantly more sophisticated than what smaller independent research/product teams have experience with.

So in addition to large-team leadership skills, there may be some niche technology/systems that only a limited number people have experience with multiple 1-3 year cycles of designing, building, maintaining, and replacing these uniquely large infra systems.