A software launch last month that got little attention in the wider world could have big implications for Google’s burgeoning AI chip business.  Google late last month announced an open-source version of TorchTPU, a software layer for its tensor processing units that represents the biggest leap on the software side of the Google chip business in years.
Nov 17, 2026

Applied AI

Ryan Barwick headshot
Supported by Sponsor Logo

A software launch last month that got little attention in the wider world could have big implications for Google’s burgeoning AI chip business. 

Google late last month announced an open-source version of TorchTPU, a software layer for its tensor processing units that represents the biggest leap on the software side of the Google chip business in years.

The overhaul seeks to make it easier for AI developers to move workflows intended for Nvidia’s graphics processing units to Google’s TPUs. At a high level, it aims to give Google what Nvidia has had with CUDA, the widely used Nvidia software layer around its AI chips that has long been one of that company’s biggest competitive strengths.

Google’s TorchTPU move comes as more companies are looking for chip alternatives to Nvidia, long the dominant force in AI silicon. As my colleague Catherine Perloff and I reported today, AI chip sales for Google, Advanced Micro Devices and Amazon could all benefit from the upheaval at Nvidia, which is dealing with both the departure of founding CEO Jensen Huang and federal scrutiny of its export business.

AI developers have long favored the AI framework PyTorch, which was developed by Meta Platforms and released nearly a decade ago. It was sticky, and Google’s own TPU framework, JAX, was clunky to use and often deemed not worth the effort to implement. 

TorchTPU—which Google developed with Meta, also a TPU customer—effectively bridges the gap, letting developers run their existing PyTorch workflows on TPUs with “minimal code changes,” Google said in an October blog post. “You can now train, serve, and deploy models on TPUs using the tools you already rely on.” it said.

Thomas Kurian, CEO of Google Cloud, the unit that runs the TPU business, said on Google parent Alphabet’s Oct. 28 earnings call that the software update “opens up a much larger market” for TPUs.

“The old objection was ‘rewrite it in JAX,’” Vadim Solovey, the CEO of the cloud software firm DoiT, told The Information. “As of last month that is gone, on paper.”

Google’s TPUs, developed more than a decade ago for internal use, have emerged as a legitimate challenger to Nvidia’s GPUs. Google previously rented its chips to customers such as Anthropic and Apple out of its own data centers, but has since started selling chips directly to companies like Fluidstack, a smaller, independent AI cloud provider.

Last month, Oracle announced plans to host Google TPUs in its cloud—the first major cloud provider other than Google to do so—and The Information reported that both Oracle and OpenAI had agreed to buy TPUs.

TPUs could be a major tailwind for Google’s already booming cloud business, whose revenue doubled in the latest quarter to $30.7 billion. TPU hardware sales alone could surpass $81 billion by next year, according to an October forecast from Morgan Stanley.

Indeed, there’s still plenty of room to grow. Solovey, whose company tracks cloud spending, estimates that only 16 of his about 5,000 clients are running TPUs in “any real volume.”

The rollout of TorchTPU also underscores a bigger issue for Nvidia—the entrenchment of CUDA is threatened by AI’s ability to code, sharply accelerating software development. Tom Brown, Anthropic’s cofounder and chief compute officer, reportedly told investors that the company’s Claude Mythos model was able to cut the time needed to onboard AMD GPUs to three days from more than six months. 

Google is continuing to push the momentum behind its chips. Earlier this month, the company unveiled two new chips in its TPU lineup, one of which was built alongside AMD, calling them more powerful and secure than previous models. 

The turmoil from the market leader could also help give TPUs an opening with smaller AI compute providers, known as neoclouds, that were heavily backed by Nvidia. An executive at one neocloud told The Information that his company was in discussions with Google over its chips, in part because of the growing use of compute for inference—actually running AI models, rather than training them—where Google claims its TPUs are more efficient. 

“We are going to be customer driven…Our customers use TPUs,” he said. “This is an inflection point for the industry, that other silicon might make sense.”

A message from Google Cloud

How to ground your AI agents in trusted business context

AI agents are only as reliable as the context behind them—and critical business logic shouldn't be left to LLM guesswork. Google Cloud’s Knowledge Catalog grounds agents in governed semantic models across your data estate, delivering nearly 4x higher accuracy while using up to 75% fewer tokens.

Explore what's new.

New From Our Reporters

Exclusive

AMD and Google Chip Away at Nvidia as Huang Exit Shakes Customers

By Catherine Perloff and Ryan Barwick

Opportunities

Group subscriptions

Empower your teams to stay ahead of market trends with the most trusted tech journalism.

Learn more


Brand partnerships

Reach The Information’s influential audience with your message.

Connect with our team

About Applied AI

Laura Bratton takes you inside how businesses are using AI to automate work.

Read the archives

Follow us
X LinkedIn Facebook Threads Instagram
Read anywhere
Apple Store Google Play
All our newsletters | Subscribe now
Gift Gift Subscription
Sent to [First name]@gmail.c­om
Manage your preferences or unsubscribe  |  Help
The Information · 251 Rhode Island Street, Suite 107, San Francisco, CA 94103