CUDA

CUDA is the Nvidia software ecosystem contrasted with Google’s [[XLACompiler|XLA]] route in E228｜谷歌TPU能撼动英伟达吗？前TPU工程师首次揭秘. The episode treats CUDA as a major reason [[GPU|GPUs]] remain broadly useful: developers, libraries, debugging habits, and model tooling have had years to accumulate around Nvidia hardware.

In this source, CUDA is not only a programming interface. It is part of AI Infrastructure Full-Stack Moat because software familiarity can protect an incumbent even when a specialized chip such as TPU offers better [[AIInferenceCostStructure|cost]] or throughput for narrower, more stable workloads.

国产 AI 算力能凭「超节点」弯道超车吗？｜WAIC 深度观察 S10E23 adds the domestic [[AIAcceleratorSupernode|supernode]] version of the same moat. Even if Huawei CM384 or other domestic systems can exceed [[NvidiaGB200NVL72|NVL72]] on aggregate hardware parameters, the source says CUDA, engineering habits, tooling, and model adaptation still make Nvidia hard to displace.

Connections

• Nvidia and GPU — platform owner and hardware category.
• TPU, XLA Compiler, JAX, and PyTorch — competing or bridge software paths.
• AI Chip Specialization, Domestic AI Chip Catch-Up, and AI Infrastructure Full-Stack Moat — ecosystem and substitution context.
• Proprietary AI Interconnect Fragmentation and Domestic AI Chip Order Validation — WAIC source’s software migration and market-proof extension.