Our eighth generation TPUs: two chips for the agentic era

Google has unveiled the eighth generation of its custom Tensor Processing Units (TPUs), introducing two distinct chips purpose-built for the emerging 'agentic era' of AI. The TPU 8t is engineered as a training powerhouse for developing complex foundation models, while the TPU 8i specializes in high-speed, low-latency inference to support responsive and collaborative AI agents. This architectural split addresses the divergent computational demands of training massive models versus running them in production, a design informed by years of anticipation of rising inference needs. Developed over a decade in partnership with Google DeepMind, these chips represent a culmination of custom hardware innovation, integrating advancements in numerics, liquid cooling, and interconnects to deliver superior power efficiency and absolute performance for the most demanding AI workloads.

These new TPUs are designed to power the next generation of AI supercomputing, enabling models to reason through problems, execute multi-step workflows, and learn from continuous feedback loops—core requirements for advanced AI agents. By co-designing silicon with software and model architectures, Google aims to provide the scale and capabilities needed for cutting-edge model development and massive inference deployments. Early adopters like Citadel Securities are already leveraging TPUs for frontier AI workloads. The TPU 8t and 8i will be generally available later this year, offering organizations specialized hardware to efficiently scale both training and serving phases as agentic AI becomes central to enterprise strategy.