LLM price pressure from China will not lead to a collapse of US Labs

Hi,

here are my two cents on why I think that the open weights models will not make the AI labs (especially OpenAI & anthropic) go insolvent.

First of all, I’d differentiate in two product groups relevant to LLM-based services. For one APIs, directly reselling tokens, and second applications, like the Claude app.

If you look at the API I think the price pressure is real. For my workflow automations that use a lot of LLM tokens I barely use the more premium US models no more. I do believe that the distance between the Chinese models and the Claude models is massive; but for many tasks you just simply don’t need more intelligence. If I extract values from a PDF, I don’t need Fable, a mistral or qwen model is just as good but cheaper.

But when I look at the application offering, Claude and ChatGPT are a lightyears ahead. Imo you can make an argument for cursor which I have also used for a year now (I really like the IDE-like UI), but the actual output just is worse, probably because of the orchestration. If you look at non-coding tasks, it not even a debate.

Imo for most humans, a lot of value is going to be in the application. Yes, we will do a lot of automation in the back-end using LLMs, but most people will want to work in an easy to use application that is as powerful as possible.

For these apps, businesses and people are to an extend price sensitive. If I could have Qwen app that is 90% cheaper (lets say $20 instead of $200) but just 75% as good, I would save $180 for 25% if the Application-based upside. For most businesses, paying the additional $180 is a no-brainer.

Following this logic, I see two issues: other apps catching up and the fact that the applications are not profitable.

For the risk of others catching up, I think it’s extremely hard to compete with the talent and resources of an anthropic and OpenAI. For me, developing the app seems to be seen as one of the top priorities. Dianne Penn, anthropics first PM said on Lennys podcast, that at anthropic people believe that frontier models need frontier apps. Trying to compete on the application layer would be like competing with Google on search.

With the profitability case, it gets more interesting imo. Since we don’t have an ipo we have to calculate on rumors. This is a bit more dubious, so I will make the worst case argument. If people max out their Claude tokens and don’t buy any additional tokens, can Claude become profitable? For the API token margins at anthropic, I’ve heard numbers ranging from 50% - 80%. I’ll assume 50%. If you calculate the max tokens you can use, you currently get 8x for OpenAI and 6x the tokens for anthropic. I’ll use the 8x. So assuming you have $200 * 8x tokens * 50% margins you get $800 worth of value in your subscription if we take the very worst numbers we find and assume max use at any given time.

This looks pretty bad, but I think what you’ll do (and they arguably already did) is shrinkflation. We are still on the exponential, token costs drop -90% per year and models get +30% better every 6-9 months. So if we map this one year in the future, your $800 worst case would become $80 in value. If you would double the available tokens in the plan in one year, Chatbots would get up to $160 in value, giving you a 20% margin. Doubling the capacity in one year would be borderline insane tho. I’m not sure when our exponential curve would become more of an S curve, but it seems pretty evident that profitability for the Chatbots is in sight. And there are still a few more years of strong innovation ahead.

The api business is more a plus. I think there are many use cases where having the best model is worth it; we have some use cases where you’d even want the additional intelligence or the volume just doesn’t justify testing different models.

This doesn’t mean that the companies will not be overvalued at IPO (nor undervalued, you’d have to check the actual data and the valuation); it just means that I think that there is a pretty stable business underneath which should stabilize the companies to an extend that they will not go tits up.

Happy to hear any thoughts / different opinions :)