Google joins Amazon in tooling to move PyTorch workloads off Nvidia GPUs

Google has released a developer tool for migrating PyTorch applications from Nvidia GPUs to its own silicon, following Amazon. The pitch competes on porting cost rather than benchmarks, targeting CUDA's ecosystem lock-in.

• Google and Amazon now offer paths to move PyTorch code off Nvidia hardware
• The real barrier is engineering time rewriting custom kernels, not card price
• A second accelerator source becomes a defensible procurement position