← The AI Systems Map
L2 · ChipsStandard / protocol⬦ ChokepointUSA

CUDA / software moat

The lock-in layer of the accelerator stack

CUDA plus cuDNN/NCCL/TensorRT is the software moat that makes Nvidia hardware sticky and undermines nominally competitive silicon from AMD, Groq, and hyperscaler ASICs. Any complete AI value-chain map must show that the accelerator chokepoint is as much software as silicon; ROCm and Triton/OpenAI kernels are the counter-plays.

How it fits the stack

CUDA / software moat with what it depends on (above) and what it feeds (below). The figure renders as a crawlable diagram and upgrades to an interactive 3D graph as it scrolls into view.

used bydepends ondepends ondepends ondepends ondepends ondepends oncompetes withcompetes withcompetes withcompetes withCUDA / software moatChipsClaude (Opus / Sonnet)NCCL (NVIDIACollective Comms)chokepointNvidia NeMo / NIM /DynamoNVIDIA TensorRT-LLM /DynamoPyTorchSGLangvLLMAMDHuawei CANNMoore ThreadsTriton (OpenAI)
CUDA / software moatFeeds ↓Related

CUDA / software moat in the AI stack. CUDA / software moat with its immediate upstream dependencies (top) and downstream dependents (bottom) in the AI value chain. Hover a node in 3D, or read the full relationships below.

Graph data (text) — 12 entities, 11 relationships
  • Claude (Opus / Sonnet)used byCUDA / software moat
  • NCCL (NVIDIA Collective Comms)depends onCUDA / software moat
  • Nvidia NeMo / NIM / Dynamodepends onCUDA / software moat
  • NVIDIA TensorRT-LLM / Dynamodepends onCUDA / software moat
  • PyTorchdepends onCUDA / software moat
  • SGLangdepends onCUDA / software moat
  • vLLMdepends onCUDA / software moat
  • CUDA / software moatcompetes withAMD
  • CUDA / software moatcompetes withHuawei CANN
  • CUDA / software moatcompetes withMoore Threads
  • CUDA / software moatcompetes withTriton (OpenAI)