Google and Amazon challenge Nvidia’s dominance in AI development

PyTorch serves as the foundational toolkit for approximately 85% of AI applications, having originated at Meta in 2016. Because Nvidia’s GPUs were the industry standard during the framework’s infancy, the software was architected to rely heavily on that specific hardware. Google is now releasing TorchTPU as an open-source project, allowing developers to shift workloads to its proprietary Tensor Processing Units with simple configuration changes. To ensure adoption, the search giant collaborated with Meta, Hugging Face, and the creators of vLLM and SGLang to maintain compatibility across popular development ecosystems.

Amazon’s strategy mirrors this shift through its TorchNeuron project, which ports applications to its Trainium chips. Beyond providing migration tools, Amazon is actively submitting code to the core PyTorch project to decouple its features from Nvidia hardware entirely, aiming for a more agnostic infrastructure. While these developments do not guarantee an immediate exodus from Nvidia’s ecosystem, they provide a viable exit strategy for startups and developers looking to diversify their compute resources.