Are serverless AI providers operating at a loss?

I've been hearing a lot about how Cursor, Anthropic, and other AI companies based on proprietary models and charging monthly subscriptions might be operating at huge losses.

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?