Google Cloud Launches 8th-Gen TPUs With 121 Exaflops Power and Million-TPU Training Clusters at Next '26

Google Cloud unveils its eighth-generation TPUs at Next '26, delivering 121 exaflops of power in a single superpod, alongside a revolutionary Virgo Network fabric capable of connecting over one million TPUs into a single training cluster, plus major AI infrastructure upgrades including NVIDIA Vera Rubin-powered instances and an AI Inference Gateway slashing latency by over 70%.

• Google Cloud unveils its eighth-generation TPUs at Next '26, featuring two specialized chips: TPU 8t for high-throughput training delivering 121 exaflops in a single superpod, and TPU 8i for ultra-low latency inference offering 80% better performance per dollar than the previous generation.
• New infrastructure announcements include A5X bare metal instances powered by NVIDIA Vera Rubin NVL72, Axion N4A VMs offering up to 30% better price-performance for agent workloads, and the Virgo Network fabric capable of connecting over one million TPUs across multiple data centers into a single training cluster.
• Google Cloud is expanding its AI storage and orchestration capabilities with Google Cloud Managed Lustre delivering 10 TB/s bandwidth, GKE nodes starting up to 4x faster with pod startup times cut by 80%, and a new AI-powered Inference Gateway that reduces time-to-first-token latency by more than 70%.