Google steps up AI computing with a faster, cheaper Ironwood TPU

• Google is launching Ironwood, its new TPU.
• The chips could ease GPU shortages.

Many companies training AI models have found themselves stuck — GPUs are expensive, in short supply, and power-hungry. That might soon change. Google is preparing to add a new version of its Tensor Processing Unit, called Ironwood, to its cloud service. The chip is built for speed and efficiency, aimed at helping customers run large AI models at a lower cost.

Analysts say Ironwood’s performance could be on par with GPUs from Nvidia and AMDwhich are the main tools most enterprises rely on today. That could make a real difference for businesses waiting months for GPU access or paying high premiums for electricity to keep their systems running.

Myron Xie, a research analyst at SemiAnalysis, said Google could still face limits of its own. Taiwan Semiconductor Manufacturing Company (TSMC), which makes the chips, is struggling with demand for advanced packaging — a step required to connect many small chips into one large processor. Those constraints could slow Ironwood’s rollout, Xie said.

Built for TensorFlow

Ironwood is the seventh generation of Google’s TPU line. It was designed to work hand in hand with TensorFlow, Google’s open-source AI framework that powers many research and enterprise projects. Omdia principal analyst Alexander Harrowell said TPUs have an edge when training or running models built in TensorFlow.

“Many AI models, especially in research and enterprise scenarios, are built using TensorFlow,” he said. “The TPUs are highly optimised for such operations, while general-purpose GPUs that support multiple frameworks aren’t as specialised.”

Each Ironwood chip delivers 4,614 FP8 teraflops of performance and includes 192 GB of HBM3E memory with a bandwidth of 7.37 terabytes per second. Those numbers show just how powerful the system is — and how fast data can move through it. A single Ironwood pod can scale up to 9,216 accelerators, offering 42.5 FP8 exaflops of computing power.

For comparison, Nvidia’s new GB300 NVL72 system reaches 0.36 exaflops. The pods link together through a 9.6 terabit-per-second optical interconnect, with a total of 1.77 petabytes of HBM3E memory across the system.

These pods can also be grouped into much larger clusters, which Google calls its AI Hypercomputer. The system combines compute, storage, and networking under one control layer so companies can train and deploy models across thousands of processors as if they were a single machine. To keep things running smoothly, Google uses optical circuit switching, which automatically routes data around any hardware issue without stopping the job.

IDC estimates that customers using Google’s Hypercomputer setup have seen an average 353% return on investment over three years, along with 28% lower IT costs and 55% higher operational efficiency compared with traditional infrastructure.

Who’s using it

Anthropic, one of the fastest-growing AI labs, plans to use as many as one million TPUs to train and serve its Claude models. The company said Ironwood offers strong cost-to-performance benefits, helping it handle more workloads for less money. Other firms are also starting to adopt Ironwood. Lightricks, known for its creative tools, is using the hardware to train its LTX-2 multimodal model, which combines text and image inputs.

Google itself has been increasing its TPU orders each year to meet demand — both for customers and for its own services, which rely heavily on AI. According to Harrowell, Google will buy $9.8 billion worth of TPUs from Broadcom in 2025, up from $6.2 billion in 2024 and $2 billion in 2023. That puts Google’s TPU program just behind Nvidia in scale, with around 5% of the AI chip market, while Nvidia still holds about 78%.

The challenge ahead

Despite the performance gains, many enterprises may not rush to adopt Ironwood. IDC research director Brandon Hoff said most organisations have already built their systems around Nvidia’s CUDA software, which has been around since 2007. “For enterprise customers who are writing their own inferencing, they will be tied into Nvidia’s software platform,” Hoff said. TensorFlow, by contrast, was only released in 2015, so many legacy systems still depend on CUDA.

That software gap could slow TPU adoption even as Google closes the hardware gap. Still, Ironwood’s efficiency, scale, and integration with TensorFlow may appeal to research groups and newer AI firms that don’t have as much existing code to rewrite.

Completing Google’s chip stack

Ironwood is part of a bigger story. Over the past decade, Google has developed its own chips across different products — from the mobile Tensor processors in its phones to the Titanium controllers used in its data centres. Now it’s also rolling out Axion, its first Arm-based general-purpose CPU for cloud servers. Axion is built on the Arm Neoverse v2 platform and offers up to 50% better performance and 60% higher energy efficiency than current x86 chips, Google says.

Together, Axion CPUs, Ironwood TPUs, and Titanium controllers give Google a complete in-house chip lineup for running AI and cloud services — one that could help it rely less on outside suppliers and compete more directly with Nvidia and AMD.

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology eventsclick here for more information.

AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.