Organizations Shift to TorchTPU for Native PyTorch on Cloud TPUs, Ditching PyTorch/XLA for New Projects
A new open-source backend replaces PyTorch/XLA, allowing organizations to run AI models on Google Cloud TPUs with minimal code changes.
Key points
- Google and Meta open-sourced TorchTPU to serve as the native PyTorch backend for Cloud TPUs, replacing PyTorch/XLA for new projects.
- The software stack supports eager mode for debugging and torch.compile for production throughput, integrating with vLLM and SGLang serving engines.
- Self-reported metrics show TorchTitan training reaching 52% model FLOPs utilization on Qwen 3 8B across 8,192 Ironwood chips.
Google and Meta announced the open-source release of TorchTPU, a native backend for PyTorch on Google Cloud TPUs, on October 20, 2026, according to MarkTechPost. This software stack positions TPUs alongside GPUs and CPUs as standard devices in the PyTorch ecosystem. Organizations using PyTorch for artificial intelligence development can now target these custom chips without relying on the older PyTorch/XLA infrastructure. The update affects teams building large language models, recommendation systems, and generative applications.
The risk
The primary operational shift involves replacing PyTorch/XLA with TorchTPU for new workloads, a change recommended by Google to simplify development, according to MarkTechPost. TorchTPU uses standard PyTorch tensors without wrappers, supporting eager mode for debugging and torch.compile for optimized production graphs. It handles divergent execution modes better than its predecessor, which required pure SPMD code. This reduces the engineering effort needed to port models from GPUs to TPUs.
Who is exposed
Teams that currently run PyTorch models on GPUs have the most to gain from this update, according to Help Net Security. Google states that these workloads can move to TPUs with minimal code changes. Meta already runs large language model, ads, and recommendation workloads on TPUs through TorchTPU in production. The release also adds TPU support to the vLLM and SGLang serving engines, Ray, and Hugging Face libraries.
Cost and compliance
Performance figures remain self-reported by Google and its partners, according to MarkTechPost. TorchTitan training achieved 52% model FLOPs utilization on the Qwen 3 8B model. Scaling tests showed near-linear throughput growth up to 8,192 Ironwood chips. Zenteiq.ai reported a fourfold increase in daily token output using llm-d on Google Kubernetes Engine, though this result involves a separate routing tool.
What happens next
Managers should evaluate whether current GPU-bound workloads can migrate to TPUs to diversify hardware dependencies. Verify that internal development pipelines support the new TorchTPU versions and that serving engines like vLLM are configured for TPU inference. Monitor self-reported performance metrics against internal benchmarks to validate cost-efficiency claims before committing large-scale training jobs to the new backend.
Questions to ask your suppliers
Ask infrastructure providers if their managed AI services support the new native backend and how they handle backward compatibility with existing legacy frameworks. Determine which internal teams need training on the updated compilation workflows and debugging tools to avoid productivity losses. Check whether current service-level agreements cover performance guarantees for workloads running on this specific hardware configuration. Identify who is responsible for monitoring resource utilization and ensuring that model deployments meet compliance standards during the transition period.
What to do and how to stay safe: Google
- Review current PyTorch codebases to identify workloads that can migrate to the new native backend without significant refactoring.
- Test the new backend in non-production environments to compare performance and stability against existing GPU-based deployments.
- Verify that serving engines and training frameworks in your stack have updated support for the new TPU integration.
- Monitor official documentation for updates on supported chip generations and any changes to distributed training primitives.
Step-by-step guide: Implement MFA for Cloud Accounts: A Step-by-Step Rollout Plan
General security guidance from the CloudSec Pulse newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Does TorchTPU replace PyTorch/XLA for all existing projects?
Google recommends TorchTPU for all new workloads, but existing projects using PyTorch/XLA may continue to function; the new backend is positioned as the successor.
What hardware does TorchTPU support?
It runs on Google Cloud TPUs, with testing reported on generations ranging from v5e up to 8,192 Ironwood chips, according to MarkTechPost.
Is TorchTPU open source?
Yes, Google and Meta open-sourced TorchTPU under the Apache 2.0 license, releasing four repositories including the core backend and serving engines.
