I'm yet to see any convincing argument that inference is subsidised in any substantial way. Training and speculative expansion are where the spend is from what I can see.
I'm yet to see any convincing argument that inference is subsidised in any substantial way. Training and speculative expansion are where the spend is from what I can see.
A few days ago Gemini redid their rate limits, making images/audio/video generation much more expensive, shrunk limits across the board (including a new weekly limit) and added more expensive tiers.
At the moment you can pay $20/month to do thousands of expensive queries a month (involving file uploads, the Pro model, extended thinking), and evidence suggests that heavy users are not profitable.
I agree that heavy users are probably not profitable but that's the way the economics of subscription services tends to work across the board.
I'm arguing that even if inference isn't profitable right now it's not orders of magnitude off. Whatever pricing emerges for models equivalent to current frontier models won't be significantly higher than the current API pricing.
There are already enough small companies without tons of VC money to burn that are serving up nearly-frontier llms at prices lower than the big players are charging. They can't all be subsidising? These are companies without any moat or any IP.
[dead]
If inference was profitable - they'd tell us. Msft, goog, public companies. They'd break out the numbers and show us, if they were good.
But instead, all we get is known liars going on podcasts and repeating "stylized facts" that aren't literally true about their supposed profitability on inference, from companies losing billions per year in a situation where they don't have to tell the truth.
That is VERY far from a convincing argument that they are profitable. So I can & will safely conclude that the opposite is true.