Google Cloud’s introduction of a new generation of artificial intelligence (AI)-focused computing built around its proprietary Axion CPUs and Ironwood TPUs represents its latest bid to challenge NVIDIA Corp.’s dominance in the AI hardware market.
Highlighting the announcement is Ironwood, Google’s seventh-generation tensor processing unit. Ironwood TPU delivers 4,614 FP8 TFLOPS of performance and features 192 GB of HBM3E memory with bandwidth reaching 7.37 TB/s, according to Google executives in a blog post. The chips can scale into massive pods containing 9,216 AI accelerators, delivering a combined 42.5 ExaFLOPS of FP8 computing power for training and inference workloads.
The performance exceeds NVIDIA’s GB300 NVL72 system, which provides 0.36 ExaFLOPS of FP8 performance. Google’s Ironwood pods also carry approximately 1.77 petabytes of HBM3E memory, surpassing those of NVIDIA, Google claims. The pods connect via a proprietary 9.6 Tb/s Inter-Chip Interconnect network and can expand into clusters running hundreds of thousands of TPUs as part of Google’s AI Hypercomputer platform.
To ensure reliability at large scale, Google employs Optical Circuit Switching, a reconfigurable fabric that automatically routes around hardware failures to maintain continuous operation. The AI Hypercomputer architecture delivers an average 353% three-year return on investment, 28% lower IT spending, and 55% higher operational efficiency for enterprise customers, according to IDC data.
Major AI companies are already committing to the platform. Anthropic plans to deploy up to one million TPUs to operate and expand its Claude model family, citing significant cost-to-performance advantages. Lightricks has begun using Ironwood to train and serve its LTX-2 multimodal system.
Alongside the TPUs, Google introduced Axion, its first Armv9-based general-purpose processor. Built on Arm’s Neoverse v2 platform, Axion is designed to deliver up to 50% better performance and 60% higher energy efficiency compared to modern x86 processors, plus 30% better performance than existing Arm-based cloud instances, Google said in its post.
Google hasn’t disclosed full specifications for Axion, though reports suggest the chip features 2 MB of L2 cache per core, 80 MB of L3 cache, and supports DDR5-5600 memory.
There are three Axion configurations: the C4A with up to 72 vCPUs and 576 GB of memory, the N4A with 64 vCPUs and 512 GB of RAM, and the C4A Metal bare-metal configuration exposing up to 96 vCPUs and 768 GB of memory.
Both Axion and Ironwood systems incorporate Google’s custom Titanium controllers, which offload networking, security, and storage processing from the host CPU to improve overall performance.
The launches cap a decade of Google’s custom silicon development, which began with the original TPU and continued through various specialized chips for YouTube, mobile devices, and infrastructure management.
