To leave Nvidia, first follow the signs to Google. If you prefer, Amazon has opened another exit. Both companies would be delighted to free your AI application from the terrible burden of sending its business to somebody else.
The exits are real. That is what makes the sales pitch interesting.
PyTorch is the widely used software toolkit developers use to build and run AI models. As TechCrunch explains, Nvidia’s chips were already standard equipment in AI research when the toolkit was developed, so many features grew around that hardware. An open-source toolkit can still develop expensive habits. Nobody has to padlock the door if leaving requires rewriting the application.
Google’s newly released TorchTPU lets existing PyTorch applications move to Google’s Tensor Processing Units with a configuration change instead of a rewrite. Amazon’s TorchNeuron offers a similar route to its Trainium chips and AWS. The practical improvement is substantial: developers get another place to run their work without first reconstructing the work.
And the helpful destination happens to sell the computing.
The Daily Heckler: five or six selected stories.
Political satire for your morning inbox.
This is competition doing something useful while keeping a firm grip on its own business card. Google and Amazon want Nvidia’s software advantage loosened because an easier move can bring applications into their clouds. The customer gains an option; the liberator gains a prospective customer. There is no need to pretend either side arrived by accident.
Amazon is also contributing code to the main PyTorch project to separate features from Nvidia hardware. That could make the toolkit more open to different chips by default. Google has worked with other developers to keep familiar tools compatible and supplied an agent suite to help with migration and errors. One company is widening the exit; the other has thoughtfully stationed assistance beside its own door.
The changes do not mean everyone will leave Nvidia. TechCrunch makes that qualification explicit. Nor does easier migration between chips turn a cloud provider into a public utility. As the banking fallout from a cloud outage demonstrated, the address of the computing still matters after the application arrives.
Developers should welcome a useful escape route without mistaking its owner for an abolitionist. These companies have not lost interest in being indispensable. They have discovered a principled objection to Nvidia being indispensable first.
The application may no longer need rewriting. The destination still needs your account.


Comments (0)
Have something to say? Give this story a heckle.
Comment Rules
Argue the point. Add a joke. No threats, harassment, hate speech, spam, private personal information or unsupported accusations. First-time comments and comments with links are moderated. Full Comment Rules
No approved comments yet. Give this story its first heckle.