Nvidia gives away the software so it can keep selling the chips
CUDA is free, open to any developer, and the most expensive thing about leaving Nvidia. That is not an accident.
The decision
Nvidia has never charged for CUDA, its programming platform for running general-purpose code on its GPUs. It has poured engineering into it for years, kept it free, and made it available to anyone with a card. From the outside this looks like generosity, or at worst a marketing expense.
The strategic why
The chip business is a treadmill. Every generation, someone builds a faster part, and every generation the advantage decays. A CEO looking at that cycle has two options: run faster, or change what customers are buying.
CUDA changed what customers were buying. Once a team’s models, tooling, and hiring pipeline assume CUDA, the question is no longer “which chip is fastest” but “which chip runs our code.” Only one does without a rewrite.
What it defends against
Every serious competitor — cloud providers building their own accelerators, chip startups, the rest of the semiconductor industry — can eventually match Nvidia on raw performance. None of them can retroactively make the world’s AI software target their hardware. The moat is not the silicon. It is the accumulated weight of everyone else’s work.
The risk
Abstraction layers are the threat. If a framework makes hardware truly interchangeable, CUDA stops being the thing people write to and becomes an implementation detail. Nvidia’s answer has been to keep CUDA ahead of the abstractions — the fastest path is still the native one. The day that stops being true, the moat starts draining.