What name did Google give to its new TPU chip built for AI inference, introduced in 2025?

Google's latest advancement in specialized hardware for artificial intelligence, introduced in 2025, is called Ironwood. This name designates the company's seventh-generation Tensor Processing Unit (TPU), a custom-built chip engineered to accelerate machine learning tasks. While previous TPU generations often focused on both training and inference, Ironwood marks a significant shift by being specifically designed for AI inference at scale, meaning it excels at using trained AI models to generate responses and insights efficiently.

Tensor Processing Units, or TPUs, are Application-Specific Integrated Circuits (ASICs) developed by Google to handle the demanding computations of neural networks more efficiently than general-purpose CPUs or GPUs. The introduction of Ironwood highlights the growing importance of inference workloads as AI models become larger and more complex. As the industry moves from primarily training frontier models to deploying them for useful, responsive interactions, hardware like Ironwood becomes crucial for delivering fast, low-latency AI services.

Ironwood boasts impressive capabilities, including the ability to scale up to 9,216 chips in a "superpod," offering immense computational power for demanding AI workloads such as large language models (LLMs) and mixture-of-experts models. This seventh-generation TPU integrates increased compute density, memory capacity with 192 GB of high-bandwidth memory per chip, and enhanced interconnect technology, all while offering significant gains in power efficiency. Its design aims to support the "age of inference," where AI agents proactively retrieve and generate data to collaboratively deliver insights, making it a cornerstone for the next phase of generative AI.