AI training clusters have unique networking patterns that TCP handles poorly - small messages, many parallel connections, and tolerance for some loss.
Develop a UDP-based protocol with ML-specific congestion control and message batching. Package as a drop-in library for PyTorch/TensorFlow.
Sell to cloud providers and enterprises running large training jobs. Open source core with paid management tools.
MVP is a basic Python library that shows throughput improvements in benchmarks.
Risk is large players developing in-house solutions first.