Google deploys its 7th Generation “Ironwood” TPU, and it bests Nvidia Blackwell
Google has unveiled its most powerful Tensor Processing Unit to date, and “AI Hypercomputer” ambitions
Google Cloud have unveiled its new 7th Generation Tensor Processing Unit, Ironwood. These new TPUs will be available to customers later this month and will enable efficient scaling into the “AI Hypercomputer” domain.
With Ironwood, alongside its new Axion CPUs, Google Cloud aims to deliver excellent training and inference performance to customers. With their new “Ironwood pods” (Ironwood Superpod), Google aims to scale up to 9,216 AI accelerators within a single domain. These chips are connected using a 9.6 Tb/s interconnect, with the superpod containing 1.77 PB (PetaByte) of memory. If we compare this to Nvidia’s Blackwell GB300 NVL72 systems, Google’s Superpods have much more memory, much more compute, and a faster interconnect.
🧵 Ironwood, our most powerful and energy-efficient Tensor Processing Unit (TPU) yet, will become generally available in the coming weeks. pic.twitter.com/6EuvT8ABUm
— Google (@Google) November 6, 2025
Google’s using Ironwood to create its “AI Hypercomputer”
Google’s Ironwood scaling does not end with their 9,216 accelerator-strong Superpods. These pods can be joined to create clusters with hundreds of thousands of TPUs. This is what creates Google’s AI Hypercomputer. This AI Hypercomputer unites compute, storage, and networking under a single management layer, delivering a high level of reliability.
Using its reconfigurable fabric (Optical Circuit Switching) tech, this large-scale network can switch and re-route around hardware interruptions to sustain high levels of performance and continuous operation. This prevents hardware failure and other issues from taking down the entire network. Thanks to this tech, maintenance is easier and less obstructive.
Anthropic has confirmed plans to use as many as one million TPUs to expand its Claude model family. Why, because Anthropic sees major cost-to-performance gains by using Google’s tech. Furthermore, Lightricks has also opted to deploy Google’s TPU hardware to train and serve its AI models.
Google has proven that the AI market doesn’t need to revolve around GPU manufacturers like AMD and Nvidia. Dedicated Tensor processing units have their place.
You can join the discussion on Google’s Ironwood TPU on the OC3D Forums.
