Google revealed on Tuesday that TorchTPU—its much-anticipated PyTorch variant—is now available as an open-source project. This move positions Google alongside Amazon in attempting to loosen Nvidia's grip on the AI industry, where the majority of AI applications are designed by default to operate on Nvidia's chips, or GPUs.

PyTorch is the widely-used open-source toolkit that AI developers and startups rely on to construct, train, and execute AI models. By some estimates, it powers 85% of AI applications. PyTorch was initially created by Meta (then Facebook) back in 2016, well before anyone suspected that LLMs would revolutionize the world. At that time, Nvidia's GPUs had already established themselves as the standard hardware for AI research, so numerous PyTorch features were designed to function specifically with them. In 2022, Meta transferred PyTorch to the Linux Foundation, a nonprofit organization that oversees many significant open source initiatives.

Google's variant, TorchTPU, enables AI applications to operate on Google's own AI chips, known as Tensor Processing Units (TPUs). It also simplifies the process of migrating existing PyTorch applications to TPUs and, consequently, to Google Cloud. Developers can accomplish this through a straightforward configuration adjustment without needing to rewrite their applications.

Amazon made a comparable announcement on Monday, releasing its own PyTorch variant called TorchNeuron. TorchNeuron enables AI applications to be ported to Amazon's proprietary AI chip, Trainium, and to AWS without rewriting. Amazon also stated it is going further in its push to diminish Nvidia's dominance over PyTorch. The company is contributing code to the main PyTorch project that separates various features from Nvidia's chips, making the platform more compatible with any hardware by default.

Google may be surpassing Amazon in several respects, however. It claims to have collaborated directly with Meta, Hugging Face, Inferact (creator of vLLM), RadixArk (creator of SGLang), and others to guarantee that TorchTPU's features and code libraries function properly with these popular AI development tools, just as the original PyTorch does. Additionally, the search giant has introduced a specialized tool, the TPU agent suite, to assist developers in migrating their applications to Google's chips and cloud while resolving errors along the way. It even upgraded its widely-used LiteRT tool, which runs AI on phones, laptops, and embedded devices, to integrate better with PyTorch AI applications.

None of this guarantees that AI application developers will immediately abandon Nvidia's hardware en masse simply because they can now easily transition to alternative chips and clouds. But Amazon and Google are evidently striving to eliminate Nvidia's competitive edge.