PyTorch Foundation announces cross-chip infrastructure at 2026 conference

At the PyTorch Conference North America 2026 in San Jose, the PyTorch Foundation introduced several initiatives aimed at creating portable, open-source AI infrastructure capable of running on any chip or cloud. Key announcements included the launch of TorchTPU, an open-source collaboration between Google and Meta that provides a native PyTorch backend for Google TPUs, and the public beta of TorchNeuron, AWS’s native PyTorch backend for AWS Trainium instances.

These developments are designed to reduce reliance on Nvidia’s CUDA ecosystem by allowing developers to migrate workloads with minimal code changes. Other technical updates included IBM’s torch-spyre compiler work, the introduction of TitanRL for reinforcement learning within the TorchTitan framework, and upgrades to the vLLM plugin.

The conference, which drew over 3,500 attendees, also highlighted enterprise adoption, such as Shopify reporting a 96% reduction in serving costs using PyTorch and vLLM. The Foundation welcomed nine new members, including Shopify, Nebius, and Nscale, as the AI platforms market continues to grow.