Asked how he deals with stacked dependencies on GPUs and model APIs, Pieter Levels tells how the service he used to fine-tune images for Avatar AI raised its price after he shared his revenue publicly. This is a 2023 episode, but the platform-risk lesson still holds.
…somehow it's like very, let's not go too deep, but it's very useful, very fast for AI Startups. You need a GPU for AI pretty much. Otherwise, it's very slow. So, and there's not enough GPUs being built. Like NVIDIA makes almost all of them. So that makes the cost of these servers very high. It's
…you need those for machine learning systems that also are run by somebody else. And then there's an API that is run by a company, and you build on top of that. So you depend on all these layers. How do you deal with this? Because I don't think NomadList is that it's not dependent on anything. No,
…were following me. So they quickly, within a month, they did the same. And they got way bigger because they're VC funded and they made, I think, $40 million. I think I made like maybe half a million dollars of it or $400,000. A lot within a month or something, within two months. It's insane. But the funny part what I wanted to say is that the service I used to do this, to do this fine-tuning, the cost of fine-tuning was $3. And then when they saw my tweets where I was sharing my revenue that I was making so much money with this, they said, sorry, we need to increase the price to $20 per training. And I was selling them for like $25 or $30. So for like a month, they got all the money. And I couldn't switch. And they said, yeah, it's difficult. Maybe they were telling the truth. They had problems with like getting GPUs also. But they increased the price. And I felt kind of like scammed, you know? And so this is a good example. You're dependent on a supplier who can, when they see you're successful, they increase the price. This happens to people who have co-worker spaces too. There's a Bali…
Yeah. Well, if there's one thing that you can learn from this experience, it's just to think about alternatives from the start, right? If you have a service that you're building on, you have to find a way to abstract it enough so you can have another service to plug into your system.
Exactly. So you have to go one level higher. So you have to get your own virtual private server with GPUs, you know? But then you need to learn the code Python and do all this stuff. And man, this Python is too much for me. This GPUs, it's too difficult. So I hired an AI developer for this to help…
it's kind of one of these things that is so specific that as a solopreneur kind of device to build stuff. You don't want to dive into something that takes you three years of university to understand. Man, I think