https://blog.doubleword.ai/fast-sglang-starts yeah, this is part of the idea. If you get cold starts down to seconds or subseconds, then you can run many models multiplexed on the same GPUs
I have read a Linkedin post in french recently with a university lab (one big machine serving hundreds of users) mentioning that they could swap models on demand. Will try to find back the link.