4 ms·

The metric the article focuses on is “revenue,” but it seems like the foundation many of these startups build on (other people’s LLM APIs) are much more expensi

by yojo 1y ago

The metric the article focuses on is “revenue,” but it seems like the foundation many of these startups build on (other people’s LLM APIs) are much more expensive than the last generation of startups. Given the cost of training a SOTA model, it’s not clear these companies have sustainable businesses. If your primary expense is AWS you can always shift to your own hardware once you hit sufficient scale. If you’re Cursor, how big do you need to get to eliminate your 3rd party API dependency?

simonw 1y ago
The cost of training a good model can be as low as $5.5m (DeepSeek v3) or "a few tens of millions of dollars" (Claude 3.7 Sonnet https://simonwillison.net/2025/Mar/2/ethan-mollick/ https://simonwillison.net/2025/Mar/2/ethan-mollick/) so once you have proven revenue against other models it might actually become a reasonable thing to do.
hn_throwaway_99 1y ago
This article is about AI applications, not core foundation models. None of the initial 3 companies mentioned (Cursor, Lovable, Gamma) train there own models from scratch, nor do they need to. E.g. I get tons of value from Cursor, but I also still pay for ChatGPT Plus. There is also enough competition in the core model space that these apps don't need to have an Achilles heel by being reliant on a single vendor. E.g. I think Cursor was smart to let you "bring your own API key".
- yojo 1y ago
  I get that the article is about applications, but the applications have a dependency on 3rd party models, and that dependency is costing them most (all? More than all?) of their revenue. Put another way, if I make $100M in annual revenue, but am paying out $110M to the API I wrap, it’s not nearly as compelling a business as that top-line $100M number makes it out to be. In the previous generation of startups, expenses were mostly dominated by headcount, and the cost of actually delivering the service tended to be small. The story was “keep growing revenue, and if you need to show a profit, stop hiring.” An AI startup built on other people’s models has to hope that the foundational models end up being fungible commodities, otherwise any margins you might gain will get squeezed out by your LLM provider. Alternatively, you can train your own model. I don’t know what Cursor’s userbase looks like. If everyone is paying for Pro but using their own API key, that’s obviously a high margin business.

by yojo 1y ago

simonw 1y ago

The cost of training a good model can be as low as $5.5m (DeepSeek v3) or "a few tens of millions of dollars" (Claude 3.7 Sonnet https://simonwillison.net/2025/Mar/2/ethan-mollick/ https://simonwillison.net/2025/Mar/2/ethan-mollick/) so once you have proven revenue against other models it might actually become a reasonable thing to do.

hn_throwaway_99 1y ago

This article is about AI applications, not core foundation models. None of the initial 3 companies mentioned (Cursor, Lovable, Gamma) train there own models from scratch, nor do they need to. E.g. I get tons of value from Cursor, but I also still pay for ChatGPT Plus. There is also enough competition in the core model space that these apps don't need to have an Achilles heel by being reliant on a single vendor. E.g. I think Cursor was smart to let you "bring your own API key".

yojo 1y ago
I get that the article is about applications, but the applications have a dependency on 3rd party models, and that dependency is costing them most (all? More than all?) of their revenue. Put another way, if I make $100M in annual revenue, but am paying out $110M to the API I wrap, it’s not nearly as compelling a business as that top-line $100M number makes it out to be. In the previous generation of startups, expenses were mostly dominated by headcount, and the cost of actually delivering the service tended to be small. The story was “keep growing revenue, and if you need to show a profit, stop hiring.” An AI startup built on other people’s models has to hope that the foundational models end up being fungible commodities, otherwise any margins you might gain will get squeezed out by your LLM provider. Alternatively, you can train your own model. I don’t know what Cursor’s userbase looks like. If everyone is paying for Pro but using their own API key, that’s obviously a high margin business.