Forget one chip to rule them all: With TPU 8, Google has an AI arms race to win

Google unveiled dual-optimized TPU 8 accelerators at Cloud Next: the TPU 8t for training (2.8x faster than Ironwood) and TPU 8i for inference (80% higher performance per dollar). The chips employ specialized architectures—TPU 8t with optical-circuit switches supporting up to 9,600 accelerators per pod, TPU 8i trading compute for larger SRAM and a collective acceleration engine to reduce MoE latency. Google also replaced x86 with Arm-based Axion CPUs and deployed custom network topologies (Virgo and Boardfly) to minimize scaling losses, signaling a vertical integration strategy in response to intensifying AI hardware competition.