← All transcripts

Inside Google’s new AI hardware Transcript, AI Summary & Key Points

Google Cloud Tech · Jul 08, 2026 · Science & Technology · 05:00 · EN-US

🧠 AI Summary

Google Cloud infrastructure is presented as supporting large-scale AI training with Ironwood TPUs, standard compute with Google Axion processors, physical AI deployments at the edge, and on-premises AI workloads with Google Distributed Cloud. Ironwood can scale to 9,216 chips, with 144 cubes forming a pod and scaling up to a million TPUs. Google Axion provides ARM-based C4A and N4A processors for performance-effective and cost-sensitive workloads. Physical AI models can be trained in parallel simulated environments, then deployed through Google Distributed Cloud to robots. Google Distributed Cloud brings Gemini and NVIDIA B300 Blackwell systems on premises so customers can retain control of their data.

🔑 Key Points

  • Ironwood is Google's seventh generation TPU for large-scale training.
  • Ironwood can scale up to 9,216 chips.
  • One pod consists of 144 cubes, with each cube being one rack.
  • Ironwood infrastructure can scale up to a million TPUs for large-scale training workloads.
  • ICI, or Inter-Chip Interconnect, provides high-speed, low-latency connectivity between chips.
  • Google Axion consists of ARM-based chips and processors, including C4A and N4A chips.
  • N4A is intended for cost-effective and cost-sensitive workloads, while C4A is intended for performance-effective workloads.
  • Physical AI gives robots an AI brain so they can perform manual tasks autonomously.

💡 Business ideas

Physical AI robotics for automating manual tasks in warehouses, logistics, and manufacturing01:54

Train AI models to control robots that perform specific manual tasks autonomously, such as picking up a box and moving it to another location. Train the robots in simulated environments in the cloud, then deploy the trained model at the edge and on the robot.

For
Organizations in warehouses, logistics, and manufacturing that still rely on manual tasks and want to automate them.
Solves
Manual work in warehouses, logistics, and manufacturing requires human effort; physical AI gives robots the ability to perform defined tasks autonomously.
  • A robot instructed to grab a box and bring it elsewhere; the claimed outcome is that it performs the task autonomously.
Soon you can unlock the full business plan.

Behind this: 9 build steps · 3 tools and how each is used · how to validate demand · 1 more real example · 7 things the video never answers.

Inquire for details

🤖 AI in practice

Used for

Scale large-scale AI training across up to 9,216 Ironwood TPU chips, with the option to scale to up to a million TPUs. 00:27
Run databases, web servers, applications, and other standard compute workloads that provide data for AI systems. 01:43
Train robots to perform manual warehouse, logistics, and manufacturing tasks autonomously. 01:54
Deploy trained physical-AI models at edge locations and on robots. 02:51
Run Gemini on premises for customers that need innovation while retaining control of their data. 02:58
The rest of this analysis is held back.

Behind this: 4 advice · 4 ai usage · 4 lessons · 2 limitations · 2 risks · 3 tools.

Inquire for details

From this video

7 products

Gemini Google Axion Google Distributed Cloud Ironwood NVIDIA Blackwell NVIDIA GeForce RTX NVIDIA RTX graphics cards

1 business idea

📄 Transcript

Searchable transcript of Inside Google’s new AI hardware — Google Cloud Tech (05:00). Search for a phrase, then click its timestamp to jump straight to that moment in the video.

Captions sourced from the original video on YouTube, published by Google Cloud Tech. The video, its captions and all related intellectual property remain the property of their respective owners; AINotes claims no ownership. Provided for research, accessibility and search — see the Transcript Notice and Copyright Policy.

00:00 [MUSIC PLAYING] CAMERON PERON: So you're building the next innovation for your business with Google Cloud infrastructure. But what does this infrastructure look like? Here at the AI Infra Zone, for the fourth year we're sporting the very latest in things like TPUs, Google Axion, and Google Distributed Cloud. Where should we start? Chelsie, can you tell us more about large scale training?

00:26 CHELSIE CZOP: I would love to. So here we have Ironwood, our seventh generation TPU. The really cool thing about Ironwood is it can scale up to 9,216 chips. Behind me is one rack, or one cube. And it takes 144 of those to make up a single pod. And it doesn't stop there. We can scale up to a million TPUs to support your large scale training workload.

00:49 And in between each chip is ICI to be able to connect it. That's Inter-Chip Interconnect, which is high speed, low latency to be able to get you the fastest training that you need to support your large scale training workloads. CAMERON PERON: Amazing. Thank you, Chelsie. CHELSIE CZOP: You're so welcome. CAMERON PERON: Behind all these incredible AI workloads, we need to talk about standard compute.

01:08 So for use cases like the databases, the web servers, the applications, all that data is fueling your AI. How can we run those workloads in a way that is cost effective, and we're achieving the performance that we're looking for? This is Google Axion. Google Axion is our ARM based chips and processors, featuring both our C4A and our N4A chips. N for cost effective and cost sensitive workloads.

01:34 And C for performance effective workloads. Incredible. So we've talked about large scale training and we talked about the standard compute workloads, but what about those workloads where we train them on the cloud, but we need to deploy them in Edge locations. Let's explore the very latest in physical AI. Robert-- ROBERT RIEMER: Hey, how's it going?

01:53 CAMERON PERON: All good, man. Tell us more about physical AI. ROBERT RIEMER: Yeah, so the idea behind physical AI is really to overcome the challenges that we have today in warehouses, and logistics, and manufacturing, where there are still a lot of manual tasks, right? And we want to automate those tasks, where [? human aids ?] comes to the game, robotics, and so on.

02:12 And physical AI is the idea to bring a brain to the robot. And the brain is the AI, right? So that the robot is able to do those specific manual tasks autonomous. So we say to the robot, hey, go and grab-- I don't know-- this box. And bring it over there. And it does all the work autonomously. And for this, we need to have a piece of simulation. Because you need to train the robot.

02:34 And by training in a physical world, you would need to have thousands of those robots simultaneously training. What we can do in the cloud is we can abstract that and do all this training in parallel scenarios in a simulated world, running on our G4 and latest NVIDIA RTX graphics card, right? CAMERON PERON: Right. ROBERT RIEMER: And then once we have trained those things with all the data that we are gathering, we can then bring this model to the edge on our Google Distributed Cloud, and then, from there, basically to

03:01 the robot itself. We are solving basically those manual tasks in warehouses, logistics, manufacturing to really make them autonomous. CAMERON PERON: Amazing. Thank you so much, Robert. ROBERT RIEMER: Have a good one. CAMERON PERON: Great. So we learned about large scale training with TPUs. We learned about how to enable everyday workloads with Google Axion.

03:19 And we learned a little bit about physical AI. But what about running Gemini and agents on premises? Brian! BRIAN KRACIK: Hey, Cameron. How are you doing? CAMERON PERON: All good. Tell us about Google Distributed Cloud. BRIAN KRACIK: I tell you what, Cameron. A lot of our customers have regulatory requirements. And they can't do everything in the Cloud.

03:38 So they need control. But we also want to give them innovation. So when you're looking for innovation and control, we're bringing that on premises with Google Distributed Cloud. And just today at the show we announced Google Distributed Cloud and Gemini running on that with support of NVIDIA B300s and those Blackwell systems. Isn't that cool? CAMERON PERON: Incredible stuff.

04:01 Amazing. BRIAN KRACIK: So you can do all of that on your premises without the data ever leaving your site. We take these B300s right here, put them into this chassis. And we'll load them up into our racks. CAMERON PERON: OK. BRIAN KRACIK: And when we look at this, this is how you get that innovation and control. We have compute. We have networking. We have storage, and everything you require to run Gemini on premises.

04:26 So if you need that innovation, you need the control, you think of Google Distributed Cloud and all the magic of Google Cloud. CAMERON PERON: Great. Wonderful. Thank you so much, Brian. BRIAN KRACIK: All right, my friend. CAMERON PERON: Great. So we learned about large scale training with TPUs. We learned about those everyday workloads with Google Axion.

04:43 And we learned about all the incredible physical AI and regulated data that you can do with Google Distributed Cloud. You've got everything that you need to build your next innovation. And I can't wait to see what you build next. [MUSIC PLAYING]