There is no inference moat Hasn’t been since 2023 with model compilation from torch 2.0 and consolidation to transformers from DiT Nvidia loses inference market long term on batch to lower TCO (AMD) and real-time (TPU, ASICS)