Google has announced the general availability of its latest Ironwood TPU, the 7th-gen offering from the Tensor Processing Unit family, alongside new Arm-based Axion compute instances, designed to handle both high-performance AI inference and everyday workloads.
Ironwood delivers a tenfold peak performance improvement over TPU v5p and more than 4x the performance per chip compared to TPU v6e (Trillium), making it Google’s most powerful and energy-efficient TPU yet, capable of scaling up to 9,216 chips in a superpod with 1.77 petabytes of shared High Bandwidth Memory. This level of connectivity, combined with Optical Circuit Switching and advanced Inter-Chip Interconnect networking, ensures uninterrupted, high-speed AI inference at planet-scale.
Meanwhile, the new Axion portfolio complements specialized accelerators with efficient general-purpose compute. N4A, the cost-effective Arm-based virtual machine now in preview, provides up to two times better price-performance than comparable x86 VMs, while the forthcoming C4A metal bare-metal instance offers dedicated physical servers for specialized workloads. Customers such as Vimeo, ZoomInfo, and Rise report measurable gains in performance, efficiency, and cost reduction, with improved throughput for transcoding, data pipelines, and API services.
Ironwood and Axion form part of Google’s AI Hypercomputer strategy, which integrates custom silicon, networking, storage, and software to maximize system-level efficiency. Software enhancements like Cluster Director in Google Kubernetes Engine, MaxText for LLM training, vLLM, and GKE Inference Gateway help developers optimize workflows, reduce serving costs, and accelerate time-to-first-token by up to 96%. This tightly co-designed stack demonstrates how Google is equipping organizations to scale AI infrastructure efficiently while supporting both cutting-edge model training and routine computing tasks.
Test applications are now open for Ironwood TPUs, Axion N4A, and C4A metal for early experience of Google’s next-gen yet flexible high-performance environment.
