TPU

A Tensor Processing Unit (TPU) is a custom AI accelerator chip designed by Google to handle the matrix-heavy computations at the core of machine learning training and inference. Unlike general-purpose CPUs or graphics-focused GPUs, TPUs are purpose-built to run AI frameworks efficiently at scale, making them central to the infrastructure behind Google's AI services, research, and cloud offerings.

Google develops TPUs for its own data centers and makes them available to external developers through Google Cloud. Each generation targets improvements in throughput, memory bandwidth, and energy efficiency for AI workloads. The chips compete in a rapidly growing market alongside GPU-based solutions from Nvidia and custom silicon from other large cloud providers.

Manufacturing TPUs requires leading-edge semiconductor fabrication, and Google has historically depended on TSMC as its primary foundry partner. As AI chip demand has intensified, Google has entered talks to bring Samsung into production of its next-generation TPU—codenamed "Icefish"—to diversify its supply chain and secure additional capacity. This reflects a broader industry trend of hyperscalers actively managing foundry relationships to keep pace with surging AI infrastructure demand.