Google introduced the eighth generation of its custom tensor processor unit, with two distinct, purpose-built architectures for training and inference: TPU 8t and TPU 8i. The new chips are intended to power Google’s custom-built supercomputers and support workloads spanning model training, agent development, and large-scale inference.
Google said its tensor processor units have powered leading foundation models, including Gemini, for years. TPU 8t and TPU 8i are positioned to deliver scale, efficiency and capabilities across training, serving and agentic workloads, reflecting a broader push to support increasingly demanding Artificial Intelligence systems.
Google framed the launch around the growing needs of Artificial Intelligence agents, which are expected to reason through problems, execute multistep workflows and learn from their own actions in continuous loops. That shift is creating new infrastructure demands, and Google said TPU 8t and TPU 8i were designed in partnership with Google DeepMind to handle the most demanding Artificial Intelligence workloads and adapt to evolving model architectures at scale.