One of my frustrations building an AI startup at the application layer was this feeling that I was basically a token reseller. You can have huge revenue, great retention, happy customers, all of it. But if a big chunk of every dollar goes straight to the model provider, you start wondering how much of the business you’re actually building for yourself. And yes, you can use open source models. For some tasks they’re great, for others you still need the frontier. But even if you own the weights, you still need somewhere to run them. Without scale, compute is expensive. Need a lot of it quickly? You pay another premium. Meanwhile the labs are locking in enormous amounts of capacity years ahead, with economics you probably can’t get as a startup. So switching away from an API doesn’t magically solve the margin problem. You go from buying tokens to buying compute, and you’re still buying from someone with more scale and negotiating power than you. That’s fine as long as you’re building something customers value enough to pay a healthy margin on top. It gets uncomfortable when your supplier starts competing with you. They can charge less for the same product and still make more money on each customer than you do. In an ideal world, you want to own the weights without having to spend billions training the model yourself. That’s why open source is so awesome. You want to be able to switch models and providers without getting locked in, and run inference with a cloud provider that has the scale you don’t and actually passes those economies of scale on to you. You don’t need to own everything. But you want the freedom to switch, and the economics to build a business that’s yours.