I think @Openrouter has one of the most interesting funding/business model here and that should be seriously studied

> raise
> buy gpu
> make more revenue
> collaterize gpu with multiples
> borrow more $
> buy more gpu
> reduce inference costs
> train own models

fundamentally it benefits inference operators like openrouter on 2 fronts:

1) owning compute lowers inference costs and improves margin

2) they act as insurance against open-source supply shocks so if Chinese labs like Qwen, GLM, or DeepSeek close restrict access, inference providers with their own GPUs can move from pure routing into training smaller domain-specific models

https://x.com/arndxt_xo/status/2059671311518208336