Hacker News Reader

Comment by Aurornis | Hacker News Reader

We go through this with every startup cycle. Startups are not expected to be profitable because they’re spending so much money on growth and R&D. The concept of running a business in an intentionally unprofitable state is confusing to those who don’t understand startup funding.

The weird thing is that so many people believe that inference is unprofitable. There are large open weights models that companies run at a profit while charging far less than what OpenAI and Anthropic charge. Deepseek V4 just made their 75% off deal permanent and it was already very cheap.

Yes, you have to consider costs of training the models, but as usage grows it’s going to become a smaller and smaller part of the business.

I think we will see some data center businesses and AI companies blow up, but I think the people expecting the entire AI scene to blow up because prices quadruple are going to be disappointed.

View on HN

I wonder how much of this reasoning will make sense in the future. How much of the way you are thinking is based on the past curves reality worked before? Are you taking into account exponential acceleration? I guess abundance will flow in such a way that the idea of debt will be a thing of the past.

> There are large open weights models that companies run at a profit while charging far less than what OpenAI and Anthropic charge.

You have no idea whether those companies are making a profit.

1. All it takes is one of them operating a loss to gain market share to force the other ones to lower prices to compete.

2. There’s not reason to expect that these relatively small companies are correctly pricing GPU depreciation.

> You have no idea whether those companies are making a profit.

I doubt the various providers on OpenRouter are benevolently operating at a loss because they’re so generous.

You can also calculate the cost to run these models yourself. They are open weight and the hardware required to run them is not a secret. They can be modeled and many have done the business modeling.

I’m always surprised at how many Hacker News commenters are unaware that a lot of financial modeling and analysis has been done on these companies and models. It’s naive to think the the hottest topic in tech has not already been dissected and analyzed by the finance industry at every level.

Selling a brands new project at a loss to gain market share or to compete with other companies doing it because you hope you can outlast them isn’t being benevolent or generous.

If you want to link to a specific cost analysis that was performed by someone without a vested interest in generating hype then do it and we’ll discuss that.

Because what you wrote sounds an awful lot like “let me tell you a lot of very smart people are saying it.”

When I go to Amazon and pay them for DeepSeek inference, do you think that Amazon are subsidising that?

It’s a brand new market that they want to claim a share of. I doubt they would be making much money of selling deepseek inference right now even if it were profitable, so why not throw sum subsidies at it for a little while in the hope that you are one of the big names left standing once everyone runs out of money.

You didn’t answer my question: do you think they are doing this?

AWS already have a strategy in place for what you describe. They are very liberal in giving out credits. They don’t do it by subsidising prices.

I don’t know enough to be certain either way. But I will say that I know that Amazon has operated certain product segments at a loss before. Whether that’s with direct price subsidies or credits is irrelevant in the face of a new product with hype unlike anything I’ve ever seen in over 20 years in the industry. It’s highly plausible in the face of this absolute mania and FOMO that Amazon is operating open source inference at a loss to gain market share. They might think that inference prices will drop in the future.

They might be panicking because they don’t have good models of their own. Or they might just be price matching other open source inference providers. They have cut prices to keep up with competition many times over the years.

Whether they are doing it or not, you don’t know they aren’t, and it’s plausible that they are. So the claim that starts with “we know that people are making a profit selling open source inference at X price therefore Y” is unfounded.

GPU depreciation cycles are slowing down a lot. A big chunk of frontier model inference is still being run on Hopper-era GPUs because anything more recent is heavily bottlenecked and it makes more sense to use the newer stuff for training,

You have to be naive to believe that any pricing is permanent.

Reply on news.ycombinator.com