Meet Ironwood: Google’s most powerful AI chip, which is 24 times faster than world’s top supercomputer

At Cloud Next 2025, Google took the wraps off Ironwood, its seventh-generation Tensor Processing Unit — and a serious leap forward in AI computing muscle. Billed as the company’s most powerful, scalable, and energy-efficient AI chip to date, Ironwood, as per Google represents a significant shift in the development of AI and the infrastructure that powers its progress

Unlike previous TPUs optimised primarily for training AI models, Ironwood is the first purpose-built for inference — the phase where AI systems actually generate answers, make decisions, and reason in real time. That means it’s designed to handle the heavy compute demands of large language models, mixture-of-experts architectures, and complex reasoning tasks with speed and efficiency.

Each Ironwood chip packs a staggering 4,614 TFLOPs of peak compute, and Google Cloud customers can scale up to massive 9,216-chip pods. At full scale, that’s 42.5 exaflops of raw compute — more than 24 times the processing power of the world’s current largest supercomputer, El Capitan.

Google also says Ironwood doubles performance-per-watt over last year’s Trillium TPUs, while boasting 192GB of ultra-fast memory per chip — six times more than before. By minimizing data movement and latency on-chip, Ironwood is engineered to keep up with the increasingly “thinking” nature of today’s AI models.