Google unveils Ironwood TPU for faster AI performance

USA – At the Google Cloud Next 2025 event, Google introduced a powerful new chip called Ironwood, the seventh version of its Tensor Processing Unit (TPU).

This new chip is built mainly for AI inference, which is when trained AI models are used to make decisions or predictions using new data.

Unlike earlier TPUs that focused on training AI models, Ironwood is designed to make these models work faster and more efficiently in real-world tasks—like chatbots, voice assistants, and real-time analytics.

Faster, smarter, and greener

Google says Ironwood is twice as fast and uses half the energy compared to the last version.

This helps businesses save money on power while getting better performance, especially important for today’s growing AI needs.

Ironwood is also highly scalable, meaning it can work with anywhere from 256 to over 9,000 chips at once, all connected using a high-speed link called the Inter-Chip Interconnect (ICI).

Together, they can reach up to 42.5 exaflops of computing power, which is more than some of the world’s top supercomputers.

Built for big AI jobs

Ironwood combines the best features of earlier TPUs and improves memory, making it better for running large AI models, like chatbots, image generators, and tools that answer questions. Its design makes sure all parts work smoothly together, even in large systems.

Ironwood will be available through Google Cloud, so developers and companies can start using it right away to power their AI tools and services.

Google also announced it will invest $75 billion in its AI and cloud systems this year. This move shows that Google is serious about staying ahead in the AI industry.

Tough competition in AI chips

Google is not alone in this race. Other big tech companies—like Nvidia, Intel, AMD, Meta, and Amazon—are also building powerful AI chips.

Nvidia, still the market leader, reported a staggering 78% year-over-year revenue growth, reaching US $39.3 billion in its latest quarter, according to Forbes.

Its data center revenue alone surged 93% to US $35.6 billion, driven by rapid adoption of its Blackwell AI chips, which are optimized for massive training workloads.

Meta and Amazon are developing custom silicon, such as MTIA (Meta Training and Inference Accelerator) and Trainium/Inferentia, respectively, to reduce reliance on third-party chips and lower latency for in-house AI systems.

Meanwhile, AMD’s MI300X GPU and Intel’s Gaudi 3 AI accelerator are also gaining traction, especially among cloud service providers and enterprises building private AI infrastructure.

According to a report by TechTarget, all these companies are racing to make chips that are faster, use less power, and work well in the cloud or on devices at the edge. Because of this, new chips are now being released every year to keep up.