[scuff3d]

It always makes me laugh when people say this, because its so utterly pointless. That percentage assumes literally no other costs exist besides the direct inference cost.

Even if they quit trying to make better models today, there are a mountain of recurring costs that will never go away. Retraining the models with new data, replacing/upgrading old hardware, enormous infrastructure costs related to maintaining the actual platforms, data collection costs, payroll...

I'm not aware of a single player in the LLM space actually turning a profit, even if they're only providing inference.

[vessenes]

Anthropic.

Listen carefully to Dario’s public statements; you could just pull his most recent Dwarkesh interview for example - worth a listen in any event.

He is guilty of an engineer’s use of the word profit when he says “we never made a profit.” But he always follows up with the real story — “every model we trained has returned 2-4x in free cashflow, counting R&D and inference”

You could say “the industry is engaged in possibly ruinous competition training ever-larger models and sucking cash to do so, and in fact if anyone stops, they’ll lose forever” and those statements might be true, but to be clear the fact that these companies are posting a loss right now is a FEATURE of how R&D works, one that lets them spend more on a race. It’s not tied to the sort of financial reality accrual accounting is designed to talk about.

[B56b]

How do you reconcile this with Ed Zitron's reporting that just the AWS bill in 2024/2025 was more than their entire revenue?

https://www.wheresyoured.at/costs/#how-much-did-anthropic-an...

[ben_w]

While true, the obvious counterpoint is that open-weight models exist, that high-end desktops can run them, that said hardware doesn't yet appear to have reached the end of the road for improvements to both purchase and operational costs, and that even if it had the moment people stop having VC money to constantly churn expensive training runs for new models it suddenly makes sense to etch the weights of whatever is SOTA at that point onto a silicon wafer and run it as a much more efficient hardware circuit without wasting the overhead that comes with software doing the same thing on general-purpose hardware.

Even if the bubble burst while I was writing this comment, even if every single current LLM provider goes the way of pets.com, AltaVista, and GeoCities, that can all happen without ending vibe coding.