TPUs: Google's home advantage

(Image credit: Future)

published 12 December 2025

In the race to train and deploy generative AI models, companies have poured hundreds of billions of dollars into GPUs, chips that have become essential for the parallel processing needs of large language models.

Nvidia alone has forecast $500 billion in sales across 2025 and 2026, driven largely by Jensen Huang, founder and CEO at Nvidia, recently stated that “inference has become the most compute-intensive phase of AI — demanding real-time reasoning at planetary scale”.

Google is meeting these demands in its own way. Unlike other firms reliant on chips by Nvidia, AMD, and others, Google has long used its in-house ‘tensor processing units’ (TPUs) for AI training and inference.

What are the benefits and drawbacks of Google’s reliance on TPUs? And how do its chips stack up against the competition?

In this episode, Jane and Rory discuss TPUs – Google’s specialized processors for AI and ML – and how they could help the hyperscaler outcompete its rivals.

Highlights

“So one of the main advantages of TPUs for Google is that it designs them, it has them manufactured through TSMC, which means it has sole dibs on them. This avoids a lot of the supply chain bottlenecks we've seen in recent years where companies have been queuing up for Nvidia chips. Nation states have been queuing up for Nvidia chips, or they've been forced to buy them in massive quantities up front and then slowly figure out where they're going to deploy them all.”

“Nvidia really is setting the standard on enterprise AI hardware. Like you said, Intel, AMD are also major players in the space and they have their chunk of the market, but Nvidia is the one that the major AI developers come back to again and again. It's making the hundred billion dollar announcements, all of this investment that seems to be announced every month.”

“It's a trade off for Google, because what you're doing is throwing a lot more compute at one workload, but potentially you're sacrificing some latency there. So all of that is to say that it's not as simple as just bigger numbers are better here. It really depends on the workload. But it's clear that, particularly with the latest iteration of TPUs, Google is seriously contending with the raw performance that companies like Nvidia can offer.”

“I'd say at this point it's clear that we haven't quite hit the wall that maybe some people were predicting earlier in the year, where people were saying it's incremental returns, if you just increase the size of the model – and maybe we do need to go back to the drawing board. I think Gemini 3 Pro shows that, at least for now, there is some life in just scaling things and seeing how much performance we can squeeze out of them.”

Footnotes

TOPICS

Rory Bathgate is Features and Multimedia Editor at ITPro, overseeing all in-depth content and case studies. He can also be found co-hosting the ITPro Podcast with Jane McCallion, swapping a keyboard for a microphone to discuss the latest learnings with thought leaders from across the tech sector.

In his free time, Rory enjoys photography, video editing, and good science fiction. After graduating from the University of Kent with a BA in English and American Literature, Rory undertook an MA in Eighteenth-Century Studies at King’s College London. He joined ITPro in 2022 as a graduate, following four years in student journalism. You can contact Rory at rory.bathgate@futurenet.com or on LinkedIn.

Highlights

Footnotes

Subscribe