Nvidia's moat is getting attacked from all sides.
Most AI apps run on PyTorch → PyTorch was built for Nvidia GPUs → that's been Nvidia's lock-in.
But in 48 hours:
→ Monday: Amazon drops TorchNeuron for Trainium chips
→ Tuesday: Google open-sources TorchTPU for TPUs
Google claims migration is just "a simple config change" — no rewrites needed. They got Meta, Hugging Face, vLLM, and SGLang on board. Even built an agent suite to handle the migration.
This isn't just competition. This is a coordinated assault on $NVDA's software fortress.
If PyTorch portability becomes frictionless, Nvidia's pricing power evaporates. Watch how this plays out — could reshape the entire AI infra stack.
Most AI apps run on PyTorch → PyTorch was built for Nvidia GPUs → that's been Nvidia's lock-in.
But in 48 hours:
→ Monday: Amazon drops TorchNeuron for Trainium chips
→ Tuesday: Google open-sources TorchTPU for TPUs
Google claims migration is just "a simple config change" — no rewrites needed. They got Meta, Hugging Face, vLLM, and SGLang on board. Even built an agent suite to handle the migration.
This isn't just competition. This is a coordinated assault on $NVDA's software fortress.
If PyTorch portability becomes frictionless, Nvidia's pricing power evaporates. Watch how this plays out — could reshape the entire AI infra stack.