Google Cloud's eighth-generation TPUs and the agentic-era infrastructure race

At Google Cloud Next '26, Google expanded its AI Hypercomputer platform with two new eighth-generation TPUs: the TPU 8t for training, packing 9,600 chips in a single superpod for nearly 3x the compute of the prior generation, and the TPU 8i tuned for low-latency inference and reinforcement learning, delivering 80% better performance per dollar. Around them came the Virgo Network fabric, capable of linking 134,000 TPUs in one data center and over a million across sites, plus native PyTorch support (TorchTPU), GKE nodes that start up to 4x faster, and Managed Lustre storage at 10 TB/s. The connective thread is agentic workloads: fleets of specialized agents that decompose goals, preserve state, and demand both high throughput and very low latency, rather than simple chat. For teams planning cloud and AI roadmaps, the signal is clear: infrastructure is being re-architected around agents, and hardware/software co-design with open frameworks like JAX, PyTorch and vLLM is becoming the competitive lever.

This is a summary by our content curator. Read the original at Google Cloud Blog: https://cloud.google.com/blog/products/compute/ai-infrastructure-at-next26.