CUDA is NVIDIA's parallel computing platform, introduced in 2007, that exposes GPU compute capability through a C-like programming model. The practical significance of CUDA is that almost all production deep learning code runs through it — either directly or via higher-level frameworks (PyTorch, TensorFlow, JAX) that compile to CUDA underneath.
CUDA's dominance is the single largest factor in NVIDIA's competitive position in the AI hardware market. Competing GPU vendors (AMD, Intel) have technically comparable hardware but lack an equivalent software ecosystem. Porting a large deep-learning codebase from CUDA to an alternative platform is a multi-quarter engineering project for a well-resourced team, which produces enormous switching costs and keeps NVIDIA's market share in training-workload GPUs at approximately 90 percent through 2026.
CUDA is the canonical example of a software moat in AI infrastructure. Hardware eventually commoditizes; the ecosystem surrounding it does not.