The least flashy AI news of October 20 may be the one that changes a budget. Google open-sourced TorchTPU, a way to point ordinary PyTorch at TPUs. Amazon had put TorchNeuron out the day before, for the same job on Trainium.
Why the framework was the moat
Most training code in the wild is PyTorch. Nvidia’s advantage was never only the chip. It was that the code ran there first, and ran there without a science project. Google and Amazon have chips of their own. They do not get used if a team has to rewrite the model to try them.
TorchTPU being open source matters for the reason open source usually matters. A compiler you cannot read is a compiler you cannot blame. Amazon’s TorchNeuron is the matching door for companies that already live in AWS and have been quoted Trainium as the cheaper option they could not actually reach.
Try it on a real step, not a demo
Pick one training job, run it on the GPU you use now, and run it through the new bridge. Compare wall time, the loss curve, and the bill. If those three are boring in a good way, you have a second supplier. If the loss curve wanders, you have a science project and you should say so.
None of this pulls Nvidia out of the rack. It gives a procurement team a sentence they did not have last week: the code we already wrote can at least start on the other chip.