Pay per pre-configured GPU hour or per token for open-source LLM inference?

Is it worth paying for a pre-configured GPU for open-source LLM inference instead of using a traditional pay-per-token provider?

Source: [Hacker News](https://news.ycombinator.com/item?id=50420874)