Framework DependenceWide moat

The most powerful thread in the CUDA moat is that virtually every widely used AI framework — the software researchers actually build with — was written assuming Nvidia hardware underneath, and tuned for it over years of collective effort. When PyTorch and the tools around it are optimized first and best for Nvidia¹, the default path for anyone starting new work runs straight through Nvidia without a decision ever being consciously made.

This dependence is decisive because it operates below the level of choice. A researcher does not sit down and select a hardware vendor; they open a framework, and the framework has already made the choice, silently, by working best on the hardware it was built for. The tutorials assume it, the example code assumes it, the benchmarks everyone quotes assume it — and the path of least resistance leads, every time, to the same place.

The advantage compounds because the frameworks and the hardware improve together. Nvidia works hand in glove with the framework developers to ensure each new chip is supported on day one, fully optimized, while a rival's hardware waits at the back of the queue for support that is often partial, later, and slower. The gap in raw silicon may be small; the gap in how well the software runs on it can be large.

For a long-term owner, framework dependence is the quiet mechanism by which the whole field's software labor accrues to Nvidia's benefit. It did not have to be bought each quarter; it was built, over a decade, by an army of outside developers optimizing for the hardware they assumed everyone would use — and in assuming it, they made it so. That is the sturdiest kind of moat: one your customers dig for you.