How about companies like Groq, Fireworks, and other providers that offer serverless inference based on open-weight models? I use them a lot presuming that, since they don't have training costs and also only charge per token rather than monthly subscriptions, they would be more economically viable. Is that assumption reasonable though?