Google Unveils Ironwood: Its Most Powerful AI Chip Yet

Google has revealed its seventh-generation Tensor Processing Unit (TPU), Ironwood. It is designed specifically to power and launch the next wave of generative AI technologies. This announcement was made during Google Cloud Next '25, the company's annual conference focused on cloud computing and AI innovations.

"Our seventh generation TPU, Ironwood, represents our largest and most powerful TPU to date, a more than 10x improvement from our most recent high-performance TPU," said Thomas Kurian, CEO of Google Cloud, in a blog post about the event.

The new Ironwood TPU boasts impressive specifications. It features over 9,000 chips per pod and delivers 42.5 exaflops of computing power per pod. For perspective, one exaflop equals a billion billion (10¹⁸) calculations per second. This massive computational capability is designed to support increasingly sophisticated AI models like Gemini 2.5.

According to Amin Vahdat, VP & GM of ML, Systems, and Cloud AI, Ironwood marks the first TPU designed specifically for inference rather than training. This focus enables Google and its customers to deploy more advanced inferential AI models at scale - models that use existing data to make predictions and decisions without requiring new training.

Alongside Ironwood, Google announced Gemini 2.5 Flash, a new AI model optimized for low latency and cost efficiency. Available on Vertex AI, Google's platform for building and managing AI applications, Flash is intended for everyday use cases requiring fast responses.

"Flash is ideal for everyday use cases like providing fast responses during high-volume customer interactions, where real-time summaries or quick access to documents are needed," Kurian explained. The model can adjust its reasoning depth based on prompt complexity and allows performance control based on customer budgets.

Google also unveiled Gemini 2.5 Pro, which is now available for public preview on Vertex AI. This model is designed for complex tasks requiring deep reasoning and coding expertise. It can process up to one million words while maintaining context.

Apart from Ironwood, Google also made several other significant announcements, including:

• Agent2Agent: An open protocol for secure communication between diverse AI agents across enterprise platforms, developed with over 50 industry partners, including Accenture, Salesforce, and SAP.
• AI Agent Marketplace: A dedicated section within Google Cloud Marketplace for browsing, purchasing, and managing AI agents built by partners.
• Google Distributed Cloud (GDC): A partnership with NVIDIA to bring Gemini to NVIDIA Blackwell systems, with Dell as a key partner, enabling local AI model use in air-gapped and connected environments.

"The opportunity presented by AI is unlike anything we have ever witnessed. It holds the power to improve lives, enhance productivity, and reimagine processes on a scale previously unimaginable," Kurian concluded.