The evidence is that quotas exist, as seen here, and are low enough that people are hitting them regularly. When was the last time you hit your quota of Google searches? When was the last time you hit your quota of StackOverflow questions? When was the last time you hit your quota of YouTube videos? Any service will rate limit abuse, but if abuse is indistinguishable from regular use from the provider's perspective, that's not a good sign.
It's also kind of interesting that they don't think they can do what an economy would normally do in this situation, which is raise prices until supply matches. Shortages generally imply mispricing.
There's a lot of angles you take from that as a starting point and I'm not confident that I fully understand it, so I'll leave it to the reader.
The parent's argument is that the marginal cost of inference is minimal. However, the fundamental flaw is that he's separating inference from the high cost frontier models. It's a cross-subsidy that can't be ignored.
Without any insider knowledge on the economics of these companies, I suspect it's that the amount of infrastructure you have to build is determined by peak usage rather than average usage. If peak usage is much higher for a small part of one day a week (say on Monday morning as software developers across the US get back to work) the cost of fulfilling demand at all times can be insane. That's why companies are implementing batch/standard/priority pricing for the API.
It sounds like it's more of a profit maximization function (and not just demand) with GPU rental prices increasing 48% since Feb.
> Renting one of Nvidia’s most-advanced Blackwell generation of chips for one hour costs $4.08, up 48% from the $2.75 it cost two months ago, according to the Ornn Compute Price Index.
You're assuming they can just stop training. For the entirety of these companies' existence, they have done training. It is part of their price. They must keep pushing out better and better models. That's like saying Nvidia can just stop making new GPUs, they're obviously making so much money with their current models now.
I've seen sources like this before. It's all hearsay and promo. I was asking for any publicly available verifiable information regarding the cost of inference at scale. I haven't seen any such info personally which is why I asked.
I'm dying to see S-1 filing for Anthropic or OpenAI. I don't actually think inference is as cheap as people say if you consider the total cost (hardware, energy, capex, etc)
Well they're not public yet so you'll have to put up with rumors. But the numbers are available for companies like DeepSeek say they have an 80% profit margin, so it stands to reason OAI etc would do similar numbers considering they charge much more.
Ads do not pay enough to cover AI usage. People see the big numbers Google and Facebook make in ads and forget to divide the number by the number of people they serve ads to, let alone the number of ads they served to get to that per-user number. You can't pay for 3 cents of inference with .07 cents of revenue.
You also can't put ads in code completion AIs because the instant you do the utility to me of them at work drops to negative. Guess how much money companies are going to pay for negative-value AIs? Let's just say it won't exactly pay for the AI bubble. A code agent AI puts an ad for, well, anything and the AI accidentally puts it into code that gets served out to a customer and someone's going to sue. The merits of the case won't matter, nor the fact the customer "should have caught it in review", the lawsuit and public reputation hit (how many people here are reading this and salivating at the thought of being able to post an angrygram about AIs being nothing but ad machines?) still cost way too much for the AI companies creating the agents to risk.
Agreed, and the answer is pretty obvious as to how they start making profit. The answer is in this thread, CRANKING the cost up immensely once they establish agreements between the duopoly leaders in the field to do so in tandem and buy up any competition that seeks to challenge them.
I’m thinking 20x what the cost is now is where they’ll land. It’ll be a massive line item for software dev shops.
Or switch to using the way cheaper open weight models from various providers who don’t have to subsidize training costs so can just race to the bottom on inference pricing…
The quality isn’t really SOTA yet but at some point I assume they’ll be good enough (maybe already are?).