I think the under-analysed point is: > Most work inside most cos are not frontier-model-worthy Particularly in larger cos, this is only true because they haven't been reconfigured to the best "shape" for AI, eg by doing work in a way which automatically provides maximal context, merging teams, removing vetos on processes, etc. This will happen gradually over the next 5 years in F500 cos—larger enterprises are more susceptible to these failure modes.
See also:
On your 2 options: I think it will be challenging for the labs to compete at the application layer with the likes of Palantir, because "companies only do one thing well" & many players can wrap APIs and provide a good product so there's much more competition (because that's more about good product design and customer relationships than model capability if model capability has diminishing returns).
The exception: vertical integration in particular sectors where there is little diminishing returns to intelligence. Hence: bio, and acqusitions of bio companies.
there's a deeper underestimated reason:
consumers just do not have very many long chains of tasks which can be automated end to end to unlock huge value. their chains are shallow almost always
in enterprise, org processes form very long chains, so the reward for automation is much higher
why are most consumer AI agent demos still effectively at the "book a flight for you" level of complexity? that's not a coincidence