Nvidia doesn't own a chip factory, yet it kept 75 cents of every dollar as gross profit last quarter. The hardware isn't the reason. CUDA is.
For a developer, the lock-in is mundane. If you've ever typed torch.cuda.is_available(), you're in it. PyTorch, the inference servers, the math libraries were built for Nvidia first and ported later. Moving off means rewriting and re-testing the kernels that make a model fast, and for most teams that costs more than the chip savings.
The catch is who can afford to pay that bill. Nvidia's latest 10-Q says five customers account for about 70% of what it's owed. Those same buyers are the ones building TPUs and custom chips, with gigawatts of capacity coming online from 2027.
Switching costs protect you from the long tail. They don't protect you from customers big enough to write their own stack.
I wrote up the full business breakdown (the fabless model, where each dollar goes, and what could crack it) here: https://medium.com/@glennlenormand/nvidia-doesnt-make-chips-so-how-is-it-worth-4-4-trillion-4d0743ab12fc?sk=31dc985237f94b9afcf44da840f5de0c
Top comments (0)