CUDA ☆

CUDA is how programs talk to NVIDIA GPUs: the compiler, runtime, and nearly two decades of libraries (cuDNN, cuBLAS, NCCL and hundreds more) that every major AI framework targets first. Launched in 2006, long before deep learning, it means a new accelerator must beat not a chip but an ecosystem — the reason 'CUDA moat' is shorthand for NVIDIA's position. It is free to use and runs only on NVIDIA hardware.