Google Unveils 8th-Gen Dual TPUs: Separate Training (8t) and Inference (8i) Chips Challenge NVIDIA

CNBC·high signal

Google announced TPU 8t and TPU 8i at Cloud Next 2026, splitting its eighth-generation TPU into purpose-built training and inference architectures for the first time. TPU 8t delivers 2.8x performance over seventh-gen Ironwood for the same price, while 8i boosts inference performance 80%, both achieving 2x better performance-per-watt. Training superpods scale to 9,600 TPUs with 2 petabytes of shared high-bandwidth memory — a direct shot at NVIDIA's data center dominance.