AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Amazon is taking its biggest shot at Nvidia yet with new TorchNeuron dev tool

Collected Oct 19, 2026

Amazon Web Services released a developer tool called TorchNeuron on Monday that lets AI applications built with PyTorch run on Trainium, Amazon's custom AI chip. According to AWS, moving an app onto Trainium requires changing one line of configuration rather than rewriting code. The release puts Amazon in more direct competition with Nvidia, whose CUDA software stack has long been the path of least resistance for running PyTorch workloads on GPUs.

PyTorch is a free, open-source toolkit developers use to build, train and run AI models. Facebook originally developed it, and since 2022 it has been maintained under the Linux Foundation. AWS puts adoption at roughly 85% of AI applications written with PyTorch. Nvidia has spent years making sure CUDA supports PyTorch smoothly, which is a large part of why its hardware sits close to the default path for new projects. TorchNeuron is adapted from PyTorch itself, so developers who already know the framework should find the interface familiar.

Amazon is not the first to make this kind of move. PyTorch users can already retarget applications to AMD GPUs or to Apple's chips through a configuration change, so the pitch is less about novelty and more about how tightly Trainium is wired into the rest of AWS. That integration is where Amazon makes its argument: a startup that picks Trainium gets the surrounding cloud services for storage, deployment and scaling without assembling those pieces itself.

Colin Brace, VP of Technology Software at AWS Annapurna Labs, the unit responsible for Amazon's custom silicon, framed the choice as one of scope rather than raw chip price. He told TechCrunch that teams on AWS are not betting on a single-purpose compute provider and then building everything else around it, and that the infrastructure to move from prototype to production already exists on the platform. AWS also argues that because the same chip handles both training and inference, and because it is closely integrated with the cloud, customers save money across a full workload even when the per-token price is not the lowest available.

Brace pointed to Deepgram, an AI voice provider competing with ElevenLabs, as a customer that moved to Trainium using native PyTorch support and now processes 70% more audio per dollar at production-grade accuracy. That figure is a vendor-supplied benchmark tied to a single customer, and it is the sort of claim that is easier to make about voice workloads than about the largest model training runs, which depend heavily on scale-out networking and memory bandwidth. This suggests the comparison developers should demand is not one configuration change against another, but the cost of a full pipeline, including the engineering time spent chasing numerical differences between backends.

Amazon is also going after the software layer itself. The company says it is contributing changes upstream to the main PyTorch project, working with the community to strip out code that binds the framework to particular chips. AWS describes the goal as making PyTorch hardware-agnostic for everyone, not only for Trainium users. In practice, a more neutral core could make it easier for any accelerator vendor, including smaller challengers, to offer a working backend and reduce the amount of bespoke work a team needs to tolerate differences between chips. The counterweight is that abstraction layers often hide the machine-specific tuning that makes accelerators fast, so the interesting question will be how much performance survives when the chip-specific paths are generalized.

There is a financial logic behind the upstream push. Vendors that sell hardware and cloud capacity competing with Nvidia benefit when the default choice stops being Nvidia, and contributing to a shared open-source project is a cheaper route to that outcome than building a proprietary framework nobody wants to learn. Amazon's own framing, as expressed by Brace, is that the real story is removing lock-in that kept developers from having a real choice. A likely trade-off is that the PyTorch abstraction exists precisely to spare developers the details of getting code to run on specific hardware; a framework that leans harder on per-vendor backends could hand some of that detail back to users in the form of configuration and debugging.

Why it matters: Teams already running PyTorch on AWS, and startups weighing Nvidia-based neoclouds such as Coreweave, Nebius and Firmus, now have a lower-friction path to evaluating Trainium, which in principle makes price-performance comparisons more practical. If Amazon's upstream work succeeds, backend portability becomes a property of PyTorch itself rather than a per-vendor project, which would weaken Nvidia's software moat over time. The main things to watch are whether the one-line claim holds for real production workloads and whether the customer numbers AWS cites replicate beyond a single voice provider.

Read at TechCrunch · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Amazon wants to break Nvidia’s hold on PyTorch, the free toolkit used to build most AI apps, so that more developers run their apps on alternative chips like its own.