Thursday, 22 October 2026
Anawords
Research desk
Chips & Compute

Google open-sources TorchTPU so PyTorch models run natively on its TPU chips

Built with Meta’s PyTorch team, TorchTPU supports eager mode and torch.compile, and Google says training scaled to 8,192 Ironwood TPUs. It arrives the same week as Amazon’s TorchNeuron.

By Mayank Sahu, Founder & Editor
Sources checked by AI · approved for publication by the author
Published
2 min read · 2 sources
Anawords illustration (not a photo): Google TorchTPU. Native PyTorch on TPUs, open source. Confirmed by official source.Illustration · not a photo
Built by Anawords from the facts in this story. Not an official company image. Graphic: Anawords

In short

Google announced on 20 October 2026 that TorchTPU, a native PyTorch stack for its Tensor Processing Units, is now open source. Co-developed with Meta’s PyTorch team, it lets existing PyTorch workloads run on TPUs with minimal code changes, and Google says training scaled across up to 8,192 Ironwood TPUs.

Google open-sourced TorchTPU on 20 October 2026, co-developed with Meta’s PyTorch team.1
PyTorch workloads run on TPUs with minimal code changes; eager mode and torch.compile are supported.1

What Google released

Google Cloud executives Mark Lohmeyer and Bill Jia announced TorchTPU at PyTorch Conference 2026 on 20 October.12 It is a native PyTorch stack for Google’s TPUs, co-developed with the PyTorch team at Meta and published on GitHub.1 Google says TPUs now join GPUs and CPUs as a native PyTorch option.1

How it works

  • Few code changes: existing PyTorch workloads run on TPUs “with minimal code changes”, Google says; SiliconANGLE describes a few configuration changes.12
  • Eager and compiled: developers can debug in eager mode, then use `torch.compile` for production speed.1
  • Libraries: Hugging Face Transformers can place a model on a TPU with `device_map="tpu"`, and there are backends for vLLM and SGLang.1
  • Migration help: a TPU agent suite, in early preview, helps move apps and fix errors automatically.12

Scale results

Google says TorchTitan training was validated on up to 8,000 Ironwood (TPU v7x) chips, and a scaling test reached up to 52% model FLOPs utilisation on Qwen 3 8B across up to 8,192 Ironwood TPUs.1 These are Google’s own figures.

Who is using it

Meta’s Jana van Greunen said TorchTPU lets Meta “deploy our PyTorch workloads to Google TPUs”.1 Hugging Face, Inferact (vLLM), RadixArk (SGLang), Anyscale and India’s Zenteiq.ai are quoted as partners.1 Zenteiq.ai’s founder reports more than four times higher inference throughput across 1,000 TPU v6e chips in its setup.1

What it means

Two cloud providers released native PyTorch backends for their own chips in the same week: Google’s TorchTPU and Amazon’s TorchNeuron. Both aim at the same thing: making it cheap to try hardware other than Nvidia GPUs without rewriting code. Analyst Holger Mueller told SiliconANGLE it “remains to be seen” how much this changes Nvidia’s position.2

For developers and startups, including in India where Zenteiq.ai is already a user, more real choices of hardware can mean better prices and availability. The practical test is simple: run your own model on each and compare cost per result.

What to watch

  • Independent benchmarks comparing TorchTPU, TorchNeuron and Nvidia GPUs on the same models.
  • When components marked preview, such as the TPU agent suite, become generally available.

Related: Intel’s GPU D270 inference card and Meta’s Prometheus supercluster.

Keep reading on this topic

Chips & ComputeAmazon opens TorchNeuron beta so PyTorch code runs on Trainium chips with a one-line device changeThe open-source backend supports Trainium2 and Trainium3, works with Hugging Face and vLLM plugins, and starts public beta on 20 October. Some PyTorch features are not supported yet.Chips & ComputeMeta details Prometheus, its Ohio AI supercluster heading for about 1 million GPUsMeta’s engineers say the New Albany site trained its Muse models, now runs hundreds of thousands of GPUs, and cut training interruptions by 99% from a 2025 peak.Chips & ComputeASML says AI demand means its 2027 and 2028 supply will fall short, after record €11.7 billion quarterThe chip-equipment maker raised its 2026 sales outlook to €44–45 billion and is looking at adding capacity beyond the 30% rise already planned. China’s share of its machine sales jumped to 23%.Chips & ComputeNvidia opens its Vera Rubin rack designs to the Open Compute Project at OCP Global SummitThe updated MGX 1.2 specification supports racks up to 250kW. Nvidia also announced Meta’s adoption of Spectrum-6 switches and a new optics format with Cisco. More in Chips & Compute →

How this story was checked

  1. ResearchedWith AI assistance · 2 sources
  2. AI source checkEach sentence matched to its source
  3. Editor approvalMayank Sahu · 22 Oct, 18:14
  4. Published22 Oct, 18:14
Read the research behind this story3 sources read · 2 used · what we could not verify · published 22 Oct, 18:04 →

Verify this story

2 confirmed1 reported2 company claim1 not known

How this was made: researched, drafted and source-checked by AI (Claude, made by Anthropic), then published under the editor’s standing approval. A person did not read it line by line before it went live. AI policy

Fact ledger

How labels work
Confirmed · 2
Google open-sourced TorchTPU on 20 October 2026, co-developed with Meta’s PyTorch team.Source 1
PyTorch workloads run on TPUs with minimal code changes; eager mode and torch.compile are supported.Source 1
Reported · 1
Partners include Meta, Hugging Face, Inferact and RadixArk.Source 1,2
Company claim · 2
Google reports up to 52% MFU on Qwen 3 8B across up to 8,192 Ironwood TPUs.Source 1
Zenteiq.ai reports over 4x higher inference throughput on 1,000 TPU v6e chips.Source 1
Not known · 1
Licence terms.Not announced

Questions readers ask

What is TorchTPU?

An open-source, native PyTorch stack from Google that lets PyTorch models run on Google’s TPU chips with minimal code changes.

Did Meta help build TorchTPU?

Yes. Google says it was co-developed with the PyTorch team at Meta.

How is TorchTPU different from TorchNeuron?

TorchTPU targets Google TPUs; Amazon’s TorchNeuron targets AWS Trainium chips. Both bring native PyTorch support to non-Nvidia hardware.

Sources

  1. 1Google Cloud: TorchTPU is now open source, run native PyTorch on TPUs
    Checked 22 Oct 2026 · Archived copy (22 Oct 2026)    OFFICIAL
  2. 2SiliconANGLE: Google unveils TorchTPU
    Checked 22 Oct 2026 · Archived copy (22 Oct 2026)    OUTLET

Updates and corrections

v122 Oct 2026, 18:14First published. Research note published first. Sources checked by AI; approved for publication by the editor.

Written and approved by

Mayank Sahu, Founder & Editor

Mayank Sahu is an SEO expert who loves using AI and learning how it works. He started Anawords because he felt the market was missing the best AI articles: clear, sourced and honest about what is confirmed and what is only claimed. He sets the rules every Anawords story must pass, runs its AI-assisted research desk and gives the approval its stories are published under. Each story says how it was checked.

ProfileLinkedInFact-check policyAI policy
See a mistake, or named in this story? We reply within 24 hours. Report an error