Comment by minimaxir | Hacker News Reader

Comment by minimaxir | Hacker News Reader

At the end of his previous article (https://www.wheresyoured.at/ai-is-slowing-down/), Ed hyped this news as "a story that will possibly burst the AI bubble" and "imagine what the worst possible thing for me to get would be and you’re probably close." This news doesn't fit either criteria: OpenAI losing billions of dollars isn't shocking news and both AI boosters and AI skeptics have likely assumed that. If anything, the news that OpenAI has $25B on hand in cash as reported here, plus the $122B raised in March, show that OpenAI won't implode for another year or two if it does...and that doesn't say anything about the AI bubble. There's also the confounder that Codex wasn't released until this year which turbocharged revenue with an uncertain increase in operating costs, so it will be difficult to extrapolate 2025 finances to 2026 and beyond.

When I read "the worst possible thing for me to get" I had assumed it would be evidence that inference/Codex is fundamentally unprofitable (as Ed often blogs about) but there isn't enough information here to support that argument either: revenue is still greater than cost of revenue, and the major losses are clearly delineated.

> I had assumed it would be evidence that inference/Codex is fundamentally unprofitable (as Ed often blogs about)

I'm not sure where they'd get that idea from? If inference was fundamentally unprofitable, I don't think we'd have seen the massive CapEx spend & VC cash flooding into AI, it'd be a negative gross margin trap if that were the case.

It looks unprofitable because of the massive CapEx spend right now to build data centers.

People that think inference is not profitable are mistaking the total compute cost as inference cost, when really it needs separated into training compute vs. inference compute.

The bigger question is, is when does training slow down, if at all? If we hit plateaus with LLMs, at that point inference becomes nearly pure profit once you own the compute (and a hardware refresh cycle every 3-5 years).

LLMs eventually hitting a dead end for more advanced capabilities is what would spell trouble for the labs. Any existing hyperscaler cloud can run inference all day long, as long as they have access to a model. They don't need OpenAI or Anthropic for that. The frontier labs entire valuations rely purely on them staying ahead of the commodity curve. The moment they can't do that, they're done.

I think Ed would argue that if 90% of your customers are only using your product because their usage is subsidised and the money to cover that subsidisation is coming from unsustainable customers tokenmaxxing then you are “fundamentally unprofitable”.

The question is, can OpenAI survive if customers start tokenminning? A pure inference business could be profitable but that’s not the business OpenAI are in. OpenAI has a billion users that OpenAI loses money on.

> I had assumed it would be evidence that inference/Codex is fundamentally unprofitable (as Ed often blogs about)

Ed's claim is that they haven't shown inference to be profitable. Which is true? And that he personally believe it is unprofitable (his personal opinion, not what his data report).

I think that's a meaningful distinction with your statement

Yeah, this pretty much seals it for me that Ed has basically nothing. Sure OpenAI isn’t currently profitable, but this doesn’t say to me that they can’t become so soon(ish).

Reply on news.ycombinator.com