On Monday, Amazon Web Services (AWS) launched a tool called TorchNeuron, which enables AI startups to easily migrate their AI applications to Trainium—Amazon’s proprietary AI chip. This move directly challenges the software advantage held by Nvidia—an advantage that has made Nvidia the default hardware choice for AI applications and fueled its rise as the premier supplier of AI hardware.
TorchNeuron is designed for use with PyTorch, a free, open-source toolkit that developers use to build, train, and run AI models. According to Amazon, AI applications written in PyTorch can be made to run on Trainium by changing just a single line of configuration—no code rewriting required.
Amazon notes that approximately 85% of AI applications are written using PyTorch. Originally developed by Facebook, PyTorch has been managed by the Linux Foundation—which oversees open-source projects—since 2022. For years, Nvidia has worked to ensure that its CUDA software—the infrastructure that allows applications to run on Nvidia GPUs and other hardware—provides seamless support for AI applications built with PyTorch. Consequently, Nvidia’s hardware has become deeply, and almost inherently, integrated into the toolkit.
Because TorchNeuron is built on PyTorch, developers already familiar with the framework will naturally know how to use it. Amazon is not the first to simplify the process of porting applications to its own chips; PyTorch users can already migrate their applications to AMD GPUs or even Apple silicon with simple configuration adjustments.
However, Amazon argues that migrating to Trainium offers more than just ease of use—it also grants AI startups access to Amazon’s entire cloud platform. Unsurprisingly, AWS claims this is a superior option for startups compared to using providers like CoreWeave, Nebius, or Firmus—platforms that are largely tethered to Nvidia and its hardware. “You’re building on AWS, rather than betting on a single-purpose compute provider and then scrambling to build everything else around it,” Colin Brace, VP of technical software at AWS Annapurna Labs—the division responsible for Amazon’s custom chips—told TechCrunch via email. “On AWS,” he continued, “the infrastructure from prototype to production is already in place, which allows founders to move fast and scale when needed.”
AWS notes that because the same chip can handle both training (developing AI models) and inference (deploying finished models, such as for answering user queries)—and is tightly integrated with the cloud infrastructure that startups require—its cloud services can save startups money, even if the price per token offered by AWS isn’t the absolute lowest.
“The value of Trainium lies in its exceptional price-performance ratio and its ability to integrate both training and inference on a single platform, which improves the overall economics of the workload,” says Brace. “Take Deepgram, for example—by migrating to the Trainium platform and leveraging its native PyTorch support, they are able to process 70% more audio data for every dollar spent, all while maintaining production-grade accuracy.” Deepgram is a provider of AI-powered speech applications and a competitor to ElevenLabs.
However, Amazon is not content with merely offering developers of PyTorch-based applications an easy path to migrate to its cloud platform.
Amazon also stated that it is working to reduce PyTorch’s reliance on Nvidia, and intends to contribute code to the core PyTorch project to achieve this. In a blog post announcing the new tools, the company wrote:
Amazon’s strategy extends beyond its own hardware. The company is contributing code upstream, collaborating with the PyTorch community to remove elements that tie the framework to specific chips. The goal is to make PyTorch more hardware-agnostic, thereby benefiting the entire developer community—not just those using Trainium.
While this strategy certainly favors companies that sell hardware or cloud services competing with Nvidia, a key value proposition of PyTorch is that it abstracts away the complex, thorny challenges of getting AI code to run on hardware. It remains to be seen whether pushing Nvidia out of the picture will—as Amazon promises—provide developers with more choices and greater cost-efficiency, or if it will simply lead to more configuration headaches.
In any case, if Amazon succeeds in effectively decoupling PyTorch from Nvidia, the implications for the GPU giant could extend far beyond merely contending with rival chipmakers. As Brace puts it: “The real key lies in eliminating the kind of lock-in that leaves developers with absolutely no choice.”