I'm not sure I understand why Poolside are training their own models at all - what's the perceived or real advantage of splitting up model training efforts into smaller companies and dividing up resources like this? Is it just to have a US-domiciled LLM lab?

Limiting resources steers research towards different routes.