Google Cloud’s New AI Chips: What Ironwood and Axion Actually Mean

For years, the AI chip conversation has basically been a one-company story — Nvidia builds it, everyone else buys it. Google’s been quietly working to change that math for over a decade, and its latest move just made that ambition very hard to ignore. The company has rolled out Ironwood, its seventh-generation custom AI chip, alongside a new lineup of Arm-based processors called Axion. And in case anyone doubted whether the hardware could actually deliver, Anthropic — the company behind Claude — signed on for access to up to one million of these chips, in a deal reportedly worth tens of billions of dollars.

That’s not a small validation. Here’s what’s actually in this new hardware, and why it matters well beyond Google’s own data centers.

What Ironwood Actually Is

Ironwood is Google’s seventh-generation Tensor Processing Unit, and unlike some of its predecessors, it’s been built specifically around one job — inference. That’s the industry term for actually running a trained AI model to answer questions, generate images, or power an AI agent, as opposed to the earlier, more resource-intensive process of training that model in the first place. Google’s framing this shift deliberately, describing the current moment in AI as the “age of inference” — the idea being that the hard part now isn’t just building smarter models, it’s serving them fast, cheaply, and reliably to potentially billions of users at once.

The raw numbers behind Ironwood are genuinely eye-catching. A single chip delivers more than four times the performance per chip of its predecessor, Trillium, and roughly ten times the peak performance of the generation before that. Each chip comes packed with 192 GB of high-bandwidth memory, a sixfold jump over the previous generation, with memory bandwidth reaching 7.37 terabytes per second. When Google wires thousands of these chips together into what it calls a superpod, things get genuinely extreme — a single pod can scale up to 9,216 liquid-cooled chips, delivering a combined 42.5 exaflops of compute power. For context, that’s more raw compute than the world’s current fastest supercomputer, packed into infrastructure designed specifically to run AI workloads rather than general-purpose computing.

Efficiency got a serious upgrade too, not just raw speed. Ironwood delivers close to double the performance per watt compared to its immediate predecessor, and Google says it’s roughly 30 times more energy efficient than the very first Cloud TPU the company shipped back in 2018 — a genuinely significant efficiency curve for a chip line that’s been in continuous development for over a decade.

How It Actually Stacks Up Against Nvidia?

Google’s been fairly direct about who this chip is aimed at competitively. By its own figures, Ironwood pods significantly outperform Nvidia’s GB300 NVL72 system on raw FP8 compute — Google claims 42.5 exaflops for its superpod configuration against roughly 0.36 exaflops for Nvidia’s comparable system, along with more total high-bandwidth memory across the pod as well.

It’s worth being clear-eyed about what these numbers actually mean in practice — chip-to-chip benchmark comparisons rarely capture the full real-world picture, and Nvidia remains the dominant player in AI hardware by a wide margin in terms of overall market share. But the fact that Google is now willing to publish direct performance comparisons at all says something about how seriously it’s taking this fight. This isn’t a company quietly building infrastructure for its own internal use anymore — it’s actively positioning Ironwood as a genuine alternative for anyone currently locked into Nvidia’s ecosystem.

The Anthropic Deal Is the Real Headline Here

Impressive spec sheets are one thing. A customer actually betting tens of billions of dollars on the hardware is another entirely, and that’s exactly what happened here. Anthropic committed to accessing up to one million Ironwood TPUs, a deal that industry analysts describe as among the largest cloud infrastructure commitments ever recorded, likely spanning multiple years once networking, power, and cooling infrastructure are factored in alongside the chips themselves.

James Bradbury, Anthropic’s head of compute, connected the decision directly to the inference-first design of the chip, noting that Ironwood’s gains in both inference performance and training scalability would help the company scale efficiently while keeping pace with customer demand for speed and reliability. Given that Anthropic’s Claude models are already among the most widely used AI systems in the world, a commitment at this scale is a pretty strong signal that Ironwood isn’t just a good chip on paper — it’s good enough for one of the most demanding AI workloads currently running anywhere.

Axion: The Quieter Half of the Announcement

Ironwood has understandably grabbed most of the headlines, but Google’s Axion lineup deserves attention too, since it’s aimed at a slightly different problem. Not every workload that touches AI actually needs a specialized accelerator chip — a lot of general-purpose computing tasks that support AI applications run better on efficient, general CPUs instead. That’s exactly what Axion, Google’s custom Arm-based processor family, is built for.

The lineup comes in three configurations. N4A, described as Google’s most cost-effective N-series virtual machine to date, offers up to twice the price-performance of comparable current-generation x86-based virtual machines, with up to 64 vCPUs and 512 GB of memory. C4A steps up from there, offering up to 72 vCPUs and 576 GB of memory, and C4A Metal, Google’s first Arm-based bare-metal offering, pushes that further still to 96 vCPUs and 768 GB of memory for customers who need direct hardware access rather than a virtualized instance.

Both the Ironwood and Axion lines share something else under the hood — Google’s custom Titanium controllers, which offload networking, security, and storage processing away from the main CPU, freeing up more of the chip’s actual processing power for the workload it’s meant to be running rather than background infrastructure tasks.

Why This Matters Beyond Google’s Balance Sheet?

Google Cloud’s revenue has been climbing fast — the division posted $15.15 billion in quarterly revenue recently, up 34% year-over-year, putting it in direct competition with Microsoft Azure’s 40% growth and AWS’s 20% growth over the same stretch. CEO Sundar Pichai has pointed to AI infrastructure demand directly as one of the company’s key growth drivers, and it’s not hard to see why — custom silicon that can genuinely compete with Nvidia gives Google a real differentiator in a cloud market where compute capacity, not just software features, increasingly decides who wins enterprise customers.

For businesses and developers watching from the outside, this launch matters for a simpler reason too — more serious competition in AI chip infrastructure tends to eventually translate into better pricing and more options for anyone building AI-powered products, rather than the market staying dependent on a single dominant supplier indefinitely.

Frequently Asked Questions

What is Google’s Ironwood TPU?

Ironwood is Google’s seventh-generation Tensor Processing Unit, a custom AI accelerator chip built specifically for high-volume AI inference, offering more than four times the performance per chip of its predecessor.

How does Ironwood compare to Nvidia’s chips?

Google claims its Ironwood superpod configuration delivers significantly higher raw compute performance than Nvidia’s comparable GB300 NVL72 system, though Nvidia still holds the larger share of the overall AI chip market.

Why did Anthropic sign a deal for up to 1 million Ironwood chips?

Anthropic’s head of compute cited Ironwood’s inference performance and training scalability as key reasons for the deal, which is aimed at helping the company scale its Claude models efficiently while meeting growing customer demand.

What is Google’s Axion chip used for?

Axion is Google’s custom Arm-based CPU family designed for general-purpose computing workloads that support AI applications without requiring a specialized accelerator chip, offering better price-performance than comparable x86-based options.

When did Ironwood become available to Google Cloud customers?

Ironwood was first introduced for testing in April 2025 and became generally available to Google Cloud customers in the weeks following its official launch announcement in November 2025.