One setting instead of rewriting code. A small change like this could have a bigger impact on the AI chip market than it seems.
Amazon $AMZN today introduced TorchNeuron, software that lets users run existing PyTorch code on its custom Trainium AI chips. According to Amazon, all it takes is changing a single setting—there’s no need to rewrite the entire codebase.
Why has it been so difficult to compete with Nvidia until now?
One of Nvidia’s biggest competitive advantages $NVDA is not just its chips, but its CUDA software ecosystem. Developers have spent years building and optimizing their AI models on it. Moving to another manufacturer’s hardware has therefore often meant extra work, costs, and time. Even if competitors offered cheaper chips, migration might not have been worth it. And that’s exactly the barrier Amazon is trying to remove.
According to the company, roughly 85% of AI researchers and developers use PyTorch. TorchNeuron lets them use Trainium without making major changes to their existing code. The software is also set to be free and available as open source.
Reuters also reports that more than 30 early customers in areas such as robotics, voice AI, and real-world simulation model development are already using TorchNeuron. Some are reporting significant cost savings.
An opportunity for Amazon, a risk for Nvidia
For Amazon, this is a chance to bring Trainium to a much broader range of AWS customers. Its custom chips also give it greater control over costs and the ability to offer more competitive prices for AI computing.
For Nvidia, it’s another sign that competitors are no longer just trying to catch up with its GPU performance, but are also working to narrow the advantage it has built over the years through software.
Of course, TorchNeuron doesn’t mean the end of CUDA’s dominance overnight... Nvidia still has a massive lead in optimization, tools, and developer support. So the main question is whether Trainium can deliver comparable performance at a lower total cost on real-world workloads.
If switching to competing chips becomes significantly easier, customers will have more freedom to choose based on price and performance, rather than the hardware their models are already built on.
By the way, Morgan Stanley today raised its price target for Amazon from $335 to $350 and named it one of its top stock picks. Shares gained about 5% today, largely on optimism surrounding AWS.
Do you think Amazon can become a more significant competitor to Nvidia over time, or is its software lead still massive? 👇