Google's TPU v8 splits into separate training and inference chips

At Cloud Next 2026 Google introduced its eighth-generation TPU as a two-chip family for the first time: TPU 8t for large-scale model training and TPU 8i optimized for low-latency inference and reasoning. The split is designed to push inference economics below NVIDIA Blackwell on long-context, reasoning-heavy workloads and is paired with new DeepMind Decoupled DiLoCo work for resilient distributed training.