42.5 ExaFLOPS: Google Launches Ironwood TPU to Power Next-Gen Agentic AI

At the Next ’25 conference, Google unveiled Ironwood, its seventh-generation Tensor Processing Unit (TPU) designed to power large-scale model training and high-throughput, low-latency AI inference workloads, with production to begin in the coming weeks. According to Google, Ironwood delivers over four times the performance of its predecessor, the sixth-generation Trillium TPU, while remaining the company’s most powerful and energy-efficient chip to date for AI reasoning and insight generation.

The technical brief reveals that a single Ironwood pod contains 9,216 liquid-cooled chips connected by an Inter-Chip Interconnect network, delivering 42.5 exaFLOPS of compute performance, and each chip peak of 4,614 TFLOPS. Ironwood uses up to 9.6 Tb/s ICI bandwidth to coordinate thousands of chips as a single system, with up to 1.77 petabytes of shared HB Memory and Optical Circuit Switching for dynamic, uninterrupted operation.

Google also announced a multi-billion-dollar, multi-year deal with Anthropic to provide up to one million TPU units for training and running Claude models, underscoring Ironwood as a cornerstone of the AI infrastructure.