Google open-sources TorchTPU to challenge Nvidia's dominance in AI development

Google and Meta co-developed and open-sourced TorchTPU on October 20, 2026, allowing PyTorch workloads to run natively on Google Cloud TPUs by replacing device calls from CUDA to TPU. Arriving one day after Amazon unveiled TorchNeuron, the release challenges Nvidia's CUDA dominance among four million developers. TorchTPU passed compatibility tests on 98% of 4,917 Hugging Face models and achieved 52% model FLOPs utilization across 8,192 Ironwood TPU chips.

TorchTPU release and licensing

• ▪TorchTPU replaces PyTorch/XLA and is released under the Apache 2.0 open-source license across four GitHub repositories.
• ▪Google open-sourced TorchTPU, a native PyTorch backend for Cloud TPUs co-developed with Meta, at PyTorch Conference 2026 on October 20, 2026.

Developer onboarding and code integration

• ▪TorchTPU registers Google TPUs as standard hardware devices using PyTorch's PrivateUse1 interface, enabling eager execution without source code modifications or wrappers.
• ▪Google launched the TPU Agent Suite in early preview to automate onboarding, diagnostics, and performance tuning for developers migrating to Google Cloud.
• ▪Migrating existing PyTorch workloads from Nvidia GPUs to Google TPUs using TorchTPU typically requires changing `.to("cuda")` to `.to("tpu")`.

Model validation and framework integration

• ▪Google integrated TorchTPU with major AI development frameworks, including Hugging Face, vLLM, SGLang, and LiteRT.
• ▪Google validated 4,917 Hugging Face open-source models on TorchTPU, achieving a 98% pass rate and 97% numerical parity relative to GPU baselines.

TorchTPU performance benchmarks

• ▪TorchTPU achieved 52% model FLOPs utilization training Qwen 3 8B using TorchTitan across up to 8,192 Ironwood TPU v7x chips.
• ▪TorchTPU's DeferAndFuse execution mode yields over 50% speed improvement compared to Strict Eager mode by fusing eager operators before dispatch to TPUs.
• ▪Applying TorchTPU optimizations to FLUX.1-dev reduced image generation time from 11.49 seconds to 2.78 seconds per step.

Nvidia market position in AI

• ▪Nvidia's market capitalization surpassed $5.1 trillion on October 19, 2026.
• ▪More than four million developers work within Nvidia's CUDA toolchain, and approximately 85% of AI applications are built on PyTorch.

Amazon hardware-neutrality initiatives

• ▪Amazon released TorchNeuron for its Trainium chips on October 19, 2026, one day before Google open-sourced TorchTPU.
• ▪Amazon is submitting code upstream to PyTorch to make the framework more hardware-neutral by removing Nvidia-specific assumptions.