At Google Cloud Next 2025, Google Cloud introduced Ironwood, its seventh-generation Tensor Processing Unit (TPU), marking a significant leap in AI hardware innovation. Designed as the company’s most powerful, scalable, and energy-efficient custom AI accelerator to date, Ironwood is notably the first TPU optimized specifically for AI inference—the process by which machine learning models apply learned knowledge to make predictions or decisions on new data

According to Google, Ironwood is a next-generation response to the growing demands of generative AI, where AI agents are not just retrieving data, but generating collaborative insights and actionable intelligence. The TPU is architected to meet the intense computational and communication needs of this emerging inference era.

A cornerstone of Google Cloud’s AI Hypercomputer architecture, Ironwood supports deployment at massive scale, with up to 9,216 liquid-cooled TPUs linked by a high-speed Inter-Chip Interconnect (ICI) network. The system spans nearly 10 megawatts of power and enables developers to unlock unprecedented performance using Google’s Pathways software stack, which allows efficient coordination across tens of thousands of TPUs.

Ironwood introduces key innovations tailored for running advanced AI workloads like Large Language Models (LLMs), Mixture of Experts (MoEs), and complex reasoning systems. These models, which often surpass the capacity of a single chip, benefit from Ironwood’s low-latency, high-bandwidth ICI network, designed to maintain synchronization across large distributed TPU pods.

Google Cloud will roll out Ironwood in two deployment configurations—256-chip and 9,216-chip variants—catering to diverse AI infrastructure needs across enterprises and research labs.