Google has introduced its seventh-generation Tensor Processing Units (TPUs), known as Ironwood, designed specifically for training artificial intelligence (AI) models. This chip marks Google’s first processor built exclusively for developing and deploying AI systems.
According to reports, the performance of the Ironwood chip rivals that of Nvidia’s Blackwell GPUs. Each Ironwood chip delivers up to 4.6 petaflops of computing power (quadrillion operations per second). For comparison, the Nvidia B200 provides 4.5 petaflops, while the GB200 and GB300 reach 5 petaflops.
The Ironwood chips will be available in the coming weeks in two main configurations — as “supermodules” consisting of either 256 or 9,216 chips. AI company Anthropic has announced plans to use one million of Google’s new chips to train its Claude model.
Google first unveiled the Ironwood chip in April 2025 during the Google Cloud Next ’25 conference. This development comes amid intensifying competition in the AI chip market. Notably, Nvidia became the first company in the world to reach a $5 trillion market capitalization on October 29, and CEO Jensen Huang has projected that the combined sales of the upcoming Blackwell and Rubin chip generations could reach $500 billion by 2026.
