Ironwood

Ironwood is a computer chip made by Google, built specifically for artificial intelligence. It is the seventh generation of a chip family that Google has been developing in-house for about ten years. This family is called TPU, short for Tensor Processing Unit. Such chips can do only a single kind of task, but do it extremely well: perform vast numbers of multiplications and additions simultaneously. This is exactly what the computational work behind every AI model consists of. A regular processor in a laptop can do almost anything, but is hundreds of times slower at this particular task.

Why Google builds its own chips

The market for AI chips has so far been dominated by a single company: Nvidia. Its graphics cards are scarce and expensive, and delivery times of months are normal. Anyone operating huge data centers, as Google does, ends up in an uncomfortable position of dependency. Building its own chips is one way out of this situation.

Then there’s power consumption. A large AI data center uses as much energy as a small city. The electricity bill has by now become a bigger cost factor than the purchase price of the hardware. Google states that Ironwood delivers roughly twice the computing performance per watt compared to the previous generation. Whether this figure holds up in practice is difficult to verify from the outside — such manufacturer claims usually refer to ideal test conditions.

For investors, Ironwood is therefore above all a signal. If major cloud providers successfully deploy their own chips, their long-term demand for purchased hardware decreases. This is exactly what turns such announcements into news for the finance section, not just the tech section.

Computing as a network rather than solo

A single Ironwood chip is not enough for a modern AI model. The models are too large to fit into the memory of one chip. That’s why thousands of chips are wired together into a network, referred to in industry jargon as a pod. Google states that Ironwood pods can consist of up to around 9,000 chips working jointly on a single task.

The real difficulty lies not in the computing itself, but in the exchange of data. Every chip must continuously pass intermediate results to its neighbors. If this connection is too slow, expensive chips sit idle waiting for data. That’s why Google connects its TPUs via a proprietary network in which every chip has direct links to several neighbors. Think of it like a grid of streets rather than a single main road where everything gets jammed.

Ironwood is also clearly tailored to one of the two AI phases. During training, a model learns from example data over weeks. During inference, the finished model answers real requests from users. Ironwood is optimized primarily for inference, since that’s where billions of requests occur every day and each one costs money.

Where the chips show up in everyday life

You can’t buy Ironwood chips. Google doesn’t sell them as a component, only as computing time in its own cloud. Anyone who wants to use them rents them by the hour through Google Cloud. This is a clear difference from Nvidia, whose cards any operator can install in their own data center.

Still, almost everyone has probably already interacted with them in some way. Google’s own AI models in the Gemini series run on TPUs. Anyone reading the AI summary above search results or using the chatbot has their request processed on such chips. Other companies also rent this hardware, including AI labs that don’t want to run their own data centers.

In news coverage, the name usually comes up in connection with the race for AI infrastructure. Amazon is developing comparable chips with Trainium and Inferentia, Microsoft with Maia. A common misconception is to view such chips as a direct replacement for Nvidia hardware. They tend to complement it instead: specialized chips are cost-effective for in-house needs, while Nvidia remains the standard for the open market for now.