Hacker News Reader

Comment by camillomiller | Hacker News Reader

This is a very biased example. Also, it is possible only because right now the tools you've used are heavily subsidised by investors' money. A LOT of it. Nobody questions the utility of what you just mentioned, but nodoby stops to ask if this would be viable if you were to pay the actual cost of these models, nor what it means for 99.9% of all the other jobs that AI companies claim can be automated, but in reality are not even close to be displaced by their technology.

View on HN

Why is it biased?

So what if it's subsidized and companies are in market share grab? Is it going to cost $40 instead of $20 that I paid? Big deal. It still beats the hell out of $2k - $3k that it would have taken before and weeks in waiting time.

100x cheaper, 1000x faster delivery. Further more, v0 and ChatGPT together for sure did much better than the average web designer and copy writer.

Lastly, OpenAI has already stated a few times that they are "very profitable" in inference. There was an analysis posted on HN showing that inference for open source models like Deepseek are also profitable on a per token basis.

LLMs are particularly good at web development (granted that's a big market), probably due to a lot of the training material being that.

We don't know what AI should cost but if you look at the numbers then 2x more expensive is much too low.

Think about the pricing. OpenAI fixed everyone's prices to free and/or roughly the cost of a Netflix subscription, which in turn was pinned to the cost of a cable TV subscription (originally). These prices were made up to sound good to his friends, they weren't chosen based on sane business modelling.

Then everyone had to follow. So Anthropic launched Claude Code at the same price point, before realizing that was deadly and overnight the price went up by an order of magnitude. From $20 to $200/month, and even that doesn't seem to be enough.

If the numbers leaked to Ed Zitron are true then they aren't profitable on inference. But even if that were true, so what? It's a meaningless statement, just another way of saying they're still under-pricing their models. Inferencing and model licensing are their only revenue streams! That has to cover everything including training, staff costs, data licensing, lawsuits, support, office costs etc.

Maybe OpenAI can launch an ad network soon. That's their only hope of salvation but it's risky because if they botch it users might just migrate to Grok or Gemini or Claude.

  Then everyone had to follow. So Anthropic launched Claude Code at the same price point, before realizing that was deadly and overnight the price went up by an order of magnitude. From $20 to $200/month, and even that doesn't seem to be enough.

Maybe it was because demand was so high that they didn't have enough GPUs to serve? Hence, the insane GPU demand?

I’ve wondered if it makes sense to buy Intel along with Cerebrus in order to use Intels newest nodes while under development to fab the Cerebrus wafer level inference chips which are more tolerant of defects. Overall that seems like the cheapest way to perform inference - if you have $100B.

If it subsidized it's a problem because we're not talking about Uber trying to disrupt a clearly flawed system of transportation. We're talking about companies whose entire promise is an industrial revolution of a scale we've never seen before. That is the level of the bet. The fact they did much better than the average professional is also your own take and assessment that is purely self evident. Also, your example has fundamentally no value. You mentioned a marginal use case that doesn't scale. Personal websites will be quicker to make because you can get whatever the AI spews your way, you have basically infinite flexibility and the only contraints are "getting it done" and "looking ok/good". That is not how larger business work, at all. So there is a massive issue of scalability of this. Finally, OpenAI "states" a lot of things, and a lot of them have been proven to be flat out lies, because they're led by a man who has been proved to be a pathological narcissistic liar many times over. Yet you keep drinking the kool aid, inlcuding about inference. There are by the way reports that, data in hand, prove quite convincingly that "being profitable on inference" seems to be math gymnastics, and not at all the financial reality of OpenAI.

The vast majority of highly valuable tech companies in the last 35 years have subsidized their products or services in the beginning as they grew. Why should OpenAI be any different? In particular the tokenomics is already profitable.

Reply on news.ycombinator.com