What if Google had just created the chip that makes your favorite ChatGPT as fast as a phone call? Ironwood, its new technological gem, promises miracles... but what is it really for?
Google has unveiled Ironwood, its latest chip dedicated to AI, at the Google Cloud Next 2025 conference. According to the American giant, this 7th generation TPU surpasses the power of the world's fastest supercomputer by 24 times. A monster of 42.5 exaflops when deployed on a large scale. But in concrete terms, what is such power used for?
Unlike previous generations, Ironwood specializes in inference, the stage where AI applies its knowledge to answer questions or generate content. Out with model training, in with efficient execution. "We are entering the age of inference, where AI agents must reason, not just spit out data," explains Amin Vahdat, vice president of Google.
Also read – Your Google Pixel is more efficient since the last update
Performance, economy... and complex thinking
With 192 GB of high-speed memory per chip and 7.2 Tbps of bandwidth, Ironwood can handle much larger and more complex AI models. Most importantly, it consumes half the energy per operation of its predecessor. A key advantage in the era of power-hungry data centers.
This energy efficiency allows Google to offer services like Gemini 2.5, which can adapt its depth of reasoning according to the complexity of the task. Imagine: the same model can discover medicines in the morning and compose music for your evening, all without overheating.
Google isn't stopping at hardware. The firm is also announcing a multi-agent development kit, allowing different AIs to collaborate. An A2A protocol even ensures compatibility between competing agents, such as Salesforce and SAP. The goal? For your HR assistant and your accounting software to communicate seamlessly.
For businesses, Ironwood could reduce AI operating costs by 40%, according to Google. Enough to convince reluctant sectors, such as healthcare and finance, to take the plunge. It remains to be seen whether Microsoft and Amazon will follow suit with their own chips... or whether they will let Google dominate the age of inference.
0 Comments