At Google Cloud Next on April 22, Google announced Virgo Network, a new data center fabric built on a flat two-layer non-blocking topology that can link 134,000 TPU 8t chips within a single data center and over one million TPUs across sites into a unified training cluster. It delivers 47 petabits per second of bisectional bandwidth—four times the bandwidth per accelerator and 40% lower latency than the prior generation—and also supports up to 80,000 NVIDIA Vera Rubin GPUs per site. Google frames the design as a campus-as-a-computer architecture that turns geographically distributed infrastructure into one seamless supercomputer.