Amazon Takes a ‘Blowtorch’ to Nvidia’s Lock-In With TorchNeuron, Its One-Line Escape Hatch to Trainu
Nvidia, meet Amazon Basics…
Sooo apparently, the online bookstore that took down Borders (RIP, legends) has its sights set on Nvidia next.
On Monday, Amazon Web Services dropped TorchNeuron, a free tool that lets AI apps built with PyTorch move off Nvidia GPUs and onto Trainium, Amazon’s homemade AI chip. Which, you know… is a BFD considering Amazon claims it takes one line of configuration and zero code rewrites. Vibe coders just went from six to midnight.
“PyTorch?” - you, probably…
Yes, the free, open-source toolkit developers use to build, train, and run AI models. The Facebook built it before handing it to the Linux Foundation in 2022, and Amazon says about 85% of AI apps are written with it.
Which brings us to Jensen Huang, Denny’s most valuable ex-employee. Nvidia spent years making sure CUDA (read: the software that lets apps actually run on its chips) plays nice with PyTorch. Meaning, Nvidia hardware comes baked into the toolkit by default, and CUDA has been living rent free in there ever since.
However, Amazon isn’t the first to come knocking, of course. PyTorch users can already port their apps to AMD GPUs, or even Apple’s chips, with a config change. It’s just that AWS says its version comes with the entire cloud attached. Which, convenient, since Amazon would love for startups to quit renting from Nvidia-tied neoclouds (read: GPU rental shops) like CoreWeave, Nebius, and Firmus.
Colin Brace, Amazon’s Chief Not-Nvidia Officer, put it like this: “You’re not betting on a single-purpose compute provider and scrambling to build everything else around it.” No one tell Firmus, which pulled its IPO earlier this month. That said, AWS admits Trainium isn’t the cheapest per token. Its pitch is that the same chip handles both training and running the model, and it’s plugged into the cloud startups already use, so the whole bill shrinks. Exhibit A, courtesy of Amazon: Deepgram, the ElevenLabs-we-have-at-home, moved over and now processes 70% more audio per dollar.
Oh, and Amazon is also contributing code to PyTorch itself to evict Nvidia “remove code that ties the framework to specific chips.” Small problem: the entire point of PyTorch is hiding all the hardware sh*t from developers.
Meaning, once Nvidia stops being the default, somebody has to pick the chip and babysit the settings. Andy Jassy be like “sounds like a you problem…”
Friendly reminder, the same Amazon preaching freedom from Nvidia is apparently asking investors to help finance a cool ~$8B of Nvidia chips. So Nvidia is still Amazon’s sneaky link… it just wants everyone else to get over him.
Until next time, friends!

At the time of publishing, Stocks.News holds positions in Amazon and Apple as mentioned in the article.