That's not a business model choice, though. That's a reality of running SOTA models.
If OpenAI or Anthropic could squeeze the same output out of smaller GPUs and servers they'd be doing it for themselves. It would cut their datacenter spend dramatically.