Google has joined Amazon in attacking Nvidia’s stronghold with AI app developers

Google has announced that TorchTPU, its 'long-awaited' version of PyTorch, is now an open-source project, which is the corporate equivalent of leaving the garden gate unlocked while installing a turnstile. The thing runs AI apps on Google's Tensor Processing Units and, per the announcement, makes it 'easier to move existing PyTorch apps to TPUs and Google Cloud.' Easier. The word is doing a lot of quiet lifting there — the same way 'easier to relocate' does when the relocation is into a company town with a company shop and a company billing department.

The headline has Google joining Amazon in attacking Nvidia's stronghold with AI app developers, and 'stronghold' is the telling noun. A stronghold is only worth storming if the defenders' tenants have somewhere else to sleep, and Google is offering a bed, a compatibility layer, and a meter. Open-source is a lovely label for a tool whose main job is to make the exit from Nvidia look like a gentle stroll rather than a re-platforming project. You can hear the sales deck already: no lock-in whatsoever, just a small, friendly lock-in with us, renewable annually.

'Long-awaited' is corporate for 'late, but we are calling it anticipation.' The underlying work — getting a framework to run on bespoke silicon — is real, unglamorous, and exactly the kind of thing that never gets a keynote unless it can be dressed as generosity. So the announcement leads with the open-source flag and the word 'easier', because 'we have rebuilt the on-ramp so your existing code walks onto our hardware' is less quotable. Easier for developers, certainly; also easier for the cloud billing system, which does not care whether your app speaks PyTorch or TorchTPU so long as it speaks Google. The moat is being widened one convenient migration at a time, and the drawbridge is open-source. Nvidia's stronghold, per the title, is with AI app developers — the people who write the code and then discover the code only runs cheaply in one place.