Google Announces TPU v8t Sunfish and TPU v8i Zebrafish
StorageReview.com, May 1,2026
Google announced TPU v8t and v8i chips optimized for training and inference in the agentic era.
Google unveiled its next-generation AI accelerators at Google Cloud Next: the TPU v8t "Sunfish" for training and TPU v8i "Zebrafish" for inference, along with the Virgo data center fabric.
The v8t scales to 9,600 chips with 121 EFLOPS of FP4 compute, while the v8i delivers 80% better performance-per-dollar than its predecessor for LLM inference with enhanced on-chip SRAM. These architecturally distinct chips share a common platform and fabric but are optimized for different workloads: the v8t for dense matrix multiplication at scale and the v8i for KV cache operations and low-latency serving.
The Virgo fabric connects over 134,000 v8t chips with 47 Pb/s of non-blocking bisection bandwidth, providing 4x better per-accelerator bandwidth and 40% lower latency than prior generations.