Google Cloud said Thursday that Ironwood, its seventh-generation tensor processing unit, will be generally available to cloud customers in the coming weeks. The chip is designed for large-scale training, reinforcement learning, and high-volume, low-latency inference, with Google claiming more than four times better performance per chip for both training and inference compared to Trillium, the previous generation, and a tenfold peak improvement over TPU v5p.

Ironwood pods scale to 9,216 chips connected by a 9.6 Tb/s inter-chip interconnect, sharing 1.77 petabytes of high-bandwidth memory. Google says these superpods can be joined into clusters of hundreds of thousands of TPUs, with optical circuit switching to route around hardware failures. Tom’s Hardware reported that a single pod delivers 42.5 FP8 exaflops, far exceeding Nvidia’s GB300 NVL72 system at 0.36 exaflops.

Alongside Ironwood, Google expanded its Arm-based Axion lineup with the N4A virtual machine, now in preview, and C4A metal, a bare-metal instance coming soon in preview. Google says N4A offers up to twice the price-performance of comparable x86-based VMs.

Anthropic, an early Ironwood user, plans to access up to 1 million TPUs to serve and train its Claude models. James Bradbury, Anthropic’s head of compute, said demand is growing exponentially and that Ironwood’s improvements will help the company scale efficiently. The announcement lands as the industry’s attention shifts from training runs to the cost and latency of serving models to millions of users.