Right now sota models requires a lot of iron.
It's possible that this will always be the case. But its is not a certainty!
We've seen software improvements shave orders of magnitude of compute requirements before. This could totally happen here. Iron could easily become stranded assets.
But that said, models have already become commodities, well somewhat. Is the value in running inference or applying it?
Today, we dare not use vibe coded libraries for mission critical things, HTML sanitization as an example.
But one day, who is to say the industry won't be disrupted by a vibe coded database with ~100% Oracle compatibility? Made by a nerd in a garage.
Established code bases is a moat today. It might not be in 5 years. Big tech won't be well positioned to take advantage, because trusting vibe coded crap is risky.
My point is mostly: the future is uncertain. Big established software companies might see their moat challenged by nerd in a garage running LLMs in the cloud.
What about the Adobe suite? AutoCAD? Office, etc. (To be fair, it's possible that software never was the moat).