Google Ironwood: A Leap Forward in AI Inference with Seventh-Generation TPU

• For AI reasoning, Google’s seventh-gen-made Ironwood TPU delivers 42.5 exaflops in bulk configurations, eclipsing the world’s swiftest supercomputers.
• Google Cloud users will be able to leverage this service in 2025 with the hope of minimising the operation costs of AI, though pricing in pounds sterling is still unclear.

Ironwood, an inference chip, was the seventh-generation TPU announcement by Google and the showstopper at the Google Cloud Next ’25 event on April 09, 2025. Ironwood is a leading AI chip designed for inference, enabling trained AI models to provide predictions or responses for applications such as chatbots and recommendation systems powered by AI. It brings Google into the race with AI hardware when looking at the processor of a formidable player challenging the industry behemoths, such as Nvidia.

What is Ironwood?

Ironwood is the progression of a milestone strategy by Google during its decade-long quest for specialised AI accelerators. Unlike having bilateral resource support for training and inference, Ironwood is the first TPU pre-optimised for purpose-built inference. This perspective stems from the reality of having to deploy intricate AI models, dubbed “thinking models”, such as Large Language Models (LLMs), a Mixture of Experts (MoE), and advanced reasoning systems such as Google’s Gemini 2.5.

“Ironwood is our most powerful, capable, and energy-efficient TPU yet,” explained Amin Vahdat, Vice President of Google Cloud and General Manager of ML, Systems, and Cloud AI, to the world. “And it’s purpose-built to power thinking, inferential AI models at scale.”

Technical Features

Ironwood’s technical features make it a leading player in AI hardware. The details of this power are as shown below:

• Processing capability: Ironwood has a processing rate of 4,614 TFLOPs per Ironwood chip at FP8 precision, shaping the supercomputer with its 25.5 Exaflops, about 24 times faster than the fastest existing El Capitan computer.
• Memory and bandwidth: Provided with 192GB of HBM memory per chip, Ironwood provides six times more memory than Trillium and 4.5 times more memory bandwidth, leading to high levels of data processing. Even though all devices have been liquid-cooled, Ironwood ensures perpetual improvement of 2-fold efficiency against what Trillium ever mastered, with power usage effectiveness enhanced by 30 times against TPU v2 from 2018.
• SparseCore: The enhanced SparseCore is well-leveraged for outsize embeddings in Ironwood, perfectly tailored for the purposes of ranking, recommendation, finance, and scientific workloads.
• Inter-Chip Link (ICL): The ICL has a bidirectional capacity of 1.2 terabytes per second, surpassing Trillium by 1.5 times and thereby enabling low-latency communication across thousands of chips.

Ironwood partners with Google’s AI supercomputing architecture and debuts using Pathways, a software stack from Google’s DeepMind, which helps manage complex AI workloads efficiently.

Economic and Competitive Impact

The past quarter or so has seen an inevitable economic challenge manifest in the gap between the existing AI silicon available for inference and its skyrocketing cost, caused by the broad AI model deployment transitioning from R&D to production. Ironwood is set to cut down the costs of making AI models efficient by optimising them, which cuts down on the very expensive data centre electrical costs. By offering Ironwood through Google Cloud, Google aims to capture a significant portion of the growing AI inference market, projected to reach USD 106.15 billion in 2025.

Ironwood helps reduce Google’s reliance on third-party chip providers like Nvidia, Intel, and AMD, which supply a significant portion of processors for Google’s cloud computing infrastructure. While TPUs currently represent a small fraction of Google’s cloud computing resources, Ironwood’s inference optimisation could increase their adoption for AI workloads.

“Ironwood is built to support this next phase of generative AI and its tremendous computational and communication requirements,” Vahdat explained. While it is no surprise that Ironwood proves itself against training, more interestingly, present industry jurisdiction suggests inference is a much higher-volume market compared to the much less-travelled-by-training segment.

Comparison with the Previous Generations

The Ironwood has succeeded Trillium, the sixth-generation TPU by Google that was announced in 2024 and was made generally available in December. Unlike Trillium, the design is for inference only, a significant strategic divergence. Performance per watt is doubled while the memory and bandwidth are significantly increased, making Ironwood a significant upgrade.

Ironwood was 30 times more efficient than TPUv2 in 2018, reflecting a decade of AI HW advancements by Google. The liquid-cooled architecture and a better ICI network give it another unique perspective over previous models, allowing for smooth scaling to tens of thousands of chips.

Corporate Context

With some big names like Amazon and Microsoft both offering their entries to forays into providing AI accelerators, Ironwood finds itself in a highly competitive market. Just as Nvidia leads the development of AI acceleration technologies, Ironwood would target a niche in inferencing, leveraging the value of integration with Google Cloud on an enterprise scale.

Google TPUs have always been reserved exclusively for Google engineers or accessible through Google Cloud, an advantage that is in-house while developing AI, such as for powering Gemini 2.5 and AlphaFold. Ironwood certainly promotes this form of compounding advantage by bringing the focus more precisely to allow developers to engage with demanding applications of AI while providing nearly unrivalled performance.

Availability and Future Prospects

Later in the year 2025, Ironwood will be available for Google Cloud customers, possibly in two configurations: 256-chip servers for smaller-scale usage or 9,216-chip pods for very large, high-performance jobs (TechCrunch). Pricing considerations in British pounds are still kept secret; however, Google is likely to save money for its cloud users through an emphasis on efficiency.

Due to the capacity to work on large models such as Gemini 2.5, Google identified Ironwood as a key player for the latter in their AI strategy in the ‘inference era, a time when AI systems proactively operate on behalf of users. With advances in the field of AI, Ironwood is expected to bring access to high-performance computing for all, thus allowing firms worldwide the opportunity to afford such powerful AI applications.

Conclusion

The Ironwood TPU by Google has shaken the turf of AI hardware, rendering stunning supremacy in power, efficiency, and scalability for inference tasks. Each component of Ironwood is created to fight the heavy costs of AI deployment and reduce dependency on external chip suppliers, making Ironwood the pillar upholding Google’s AI ecosystem. Once it gets out in late 2025, Ironwood promises to expedite the development and deployment of the next generation of AI, opening the door to an even more intelligent and more accessible future.