Google Unveils Ironwood, Its Chip Built for the Age of Inference

Google has unveiled Ironwood, the seventh generation of its TPUs (Tensor Processing Units — the chips the company designs specifically for AI workloads instead of relying on general-purpose processors). What matters here isn't just the horsepower on offer, but the target: as Google announced on its official blog, this is the first TPU designed specifically for inference — that is, for running already-trained models, not training them.

That distinction matters because for years the conversation around AI chips revolved almost exclusively around training: the phase in which a model is built by devouring data across huge farms of processors. Ironwood shifts the focus to the next phase — serving those models to millions of users and agents — a territory where Nvidia has built much of its market dominance with its GPUs.

From models that "answer" to models that "reason"

Google frames the launch within what it calls the "age of inference": the shift from AI models that provide real-time information for people to interpret, to models that proactively generate insights and interpretations. In this scenario, the company explains, AI agents will proactively retrieve and generate data to collaboratively deliver insights and answers, not just raw data.

Ironwood is designed to sustain that leap: large language models, mixture-of-experts architectures (which split the work across several specialized subsystems), and advanced reasoning tasks — everything Google groups under the label of "thinking models." These kinds of workloads demand massive parallel processing and highly efficient memory access, and Ironwood was built, according to the company, to minimize data movement and latency on the chip itself while handling large-scale tensor manipulations.

Numbers that top the world's largest supercomputer

Google offers Ironwood in two configurations depending on the workload: one with 256 chips and another that scales up to 9,216 liquid-cooled chips linked by a low-latency, high-bandwidth Inter-Chip Interconnect (ICI) network that, at full deployment, spans nearly 10 megawatts.

In that 9,216-chip configuration, a single pod reaches 42.5 exaflops of compute — more than 24 times the power of El Capitan, the world's largest supercomputer, which delivers 1.7 exaflops per pod, according to Google's figures. Each individual chip peaks at 4,614 teraflops. These are numbers built to train and serve the most demanding dense or mixture-of-experts models, the ones with reasoning capabilities.

Memory and bandwidth, the real bottleneck

Beyond raw compute power, Google is putting the emphasis on memory, which tends to be the real limiting factor when serving large models. Ironwood packs 192 GB of high-bandwidth memory (HBM) per chip, six times more than Trillium, its sixth-generation predecessor announced last year. That memory's bandwidth reaches 7.37 terabytes per second per chip — 4.5 times Trillium's — and chip-to-chip interconnect has been expanded to 1.2 terabytes per second bidirectional, 1.5 times more than the previous generation.

On energy efficiency, Google says Ironwood delivers twice the performance per watt of Trillium, and nearly 30 times the efficiency of its first TPU from 2018. The company stresses that its liquid-cooling system reliably sustains up to twice the performance of conventional air cooling even under continuous, heavy AI workloads — a point that matters at a time when power availability is one of the real constraints on deploying AI capacity.

Ironwood also debuts an upgraded version of SparseCore, an accelerator specialized in processing the large embeddings (numerical representations of data) common in recommendation and ranking systems, which Google says can now also be applied outside traditional AI territory, in financial and scientific domains.

Pathways: coordinating tens of thousands of chips

To make the most of this scale, Google is giving developers access to Pathways, its own machine learning runtime developed by Google DeepMind, which efficiently coordinates distributed computing across multiple TPU chips. According to the company, Pathways makes it possible to go beyond a single Ironwood pod and combine hundreds of thousands of chips to accelerate work at the frontier of generative AI. Ironwood is also part of what Google calls its Google Cloud AI Hypercomputer architecture, which combines jointly optimized hardware and software for the most demanding AI workloads.

As an example of what's already running on TPUs, Google points to models from the Gemini 2.5 family and AlphaFold, the Nobel Prize-winning protein structure prediction system, both of which run on this infrastructure today.

What this means for the AI chip market

Over the past few years, any company looking to train or serve large models has depended heavily on Nvidia's GPUs — a bottleneck that has sent Nvidia's market value soaring and left many cloud customers tied to the availability and pricing of a single hardware vendor. Ironwood is Google's most ambitious bet yet at offering its own, vertically integrated alternative — not just in raw power, but specifically in inference, the territory where the bulk of compute spending is likely to concentrate as already-trained models get deployed at massive scale in products and agents.

Google hasn't given a price for Ironwood. The company says it will be available to Google Cloud customers later this year, and directs anyone wanting more information to request details directly through Google Cloud.