Google splits its eighth-generation TPUs into training and inference chips

At Google Cloud Next in Las Vegas, Google announced its eighth generation of TPUs, split for the first time into distinct chips for training and serving: TPU 8t for training and TPU 8i for inference and reinforcement learning. TPU 8t connects up to 9,600 chips and was said to improve training price-performance by up to 2.8x over the seventh-generation Ironwood. TPU 8i puts 1,152 chips in a pod with triple Ironwood's on-chip SRAM, optimized for running autonomous AI agents in parallel. Both target TSMC's 2nm process, with availability to customers outside Google aimed at late 2027. With Anthropic contracting for up to a million TPUs and Google investing up to $40 billion in that company, TPUs had become not just the substrate for Google's own models but a business set against NVIDIA.