Google Unveils 8th-Gen TPUs, Axion CPUs at Next ’26

At the Google Cloud Next ’26 conference in Las Vegas, Google has unveiled 8th-gen TPUs, Axion CPUs, and NVIDIA Blackwell integration to power the new era of autonomous AI agents and massive model training.

The biggest news from the event is the debut of Google’s eighth-generation Tensor Processing Units (TPUs). For the first time, Google is splitting its TPU architecture into two specialized chips: the TPU 8t for training and the TPU 8i for inference.

The TPU 8t is a beast of a processor, designed to slash the time it takes to train frontier models from months to just weeks. A single Superpod of these chips now scales to 9,600 units, delivering a mind-blowing 121 ExaFlops of compute.

On the other side, the TPU 8i is optimized for the swarming behaviour of AI agents. In this new era, complex tasks are often handled by multiple specialized agents working together. The TPU 8i features three times more on-chip SRAM than previous generations, allowing it to handle these collaborative reasoning tasks with near-zero latency.

Google isn’t just focusing on accelerators; it is also optimizing the host processors. Both the new TPUs now run on Google’s custom Axion Arm-based CPUs. By designing the full stack, Google claims it has achieved energy efficiency levels that off-the-shelf components simply cannot match.

Google also announced the general availability of NVIDIA Blackwell GPUs across its cloud regions. This includes the A4X Max VMs, which use the NVIDIA GB300 NVL72 architecture to provide a 2x speed increase in training and serving compared to last year’s hardware.

Looking even further ahead, Google teased the A5X instances powered by NVIDIA’s Vera Rubin architecture. These AI factories are designed to scale up to nearly one million GPUs in a multisite cluster, offering a glimpse into the future of industrial-scale AI.