Google’s investment in custom silicon has grown alongside the increasing demands of AI infrastructure. Its Tensor Processing Units, originally developed for internal use, have evolved into core components of large-scale cloud services. Designed to handle both training and inference, these chips offer a specialized alternative to general-purpose processors for machine learning workloads.
The latest generation, called Ironwood, builds on this foundation with improvements in processing speed and system bandwidth. These changes enable support for larger models and reduce delays during inference and training.
Available through Google Cloud, Ironwood has been introduced to meet the demands of larger models, increased enterprise use, and the growing presence of AI in production systems.
Why It Matters: AI workloads place heavy demands on compute resources and memory access. General-purpose chips can support many of these needs, but hardware designed for specific tasks often performs more efficiently. Google’s internal chip development gives it greater control over infrastructure behavior, allowing the company to manage costs and maintain consistent performance for AI models in production.
- Ironwood Adds Capacity for Larger AI Workloads: Ironwood is the seventh generation of Google’s TPU line, designed to support growing model sizes and increasing system demands. Each superpod built with Ironwood can link over 9,000 processors through a high-bandwidth interconnect that reaches 9.6 terabits per second. The system also includes 1.77 petabytes of shared memory, allowing data to move efficiently across chips during training and inference. These upgrades help reduce processing delays in large-scale deployments and allow models to run at higher throughput without additional wait times. Ironwood is also built to scale across clusters, enabling workloads to expand without restructuring infrastructure.
- Cloud Revenue Tied to AI Infrastructure Growth: Google Cloud reported $15.15 billion in revenue during its most recent quarter, an increase influenced in part by growing interest in AI-related infrastructure. Although the company does not separate TPU earnings from overall cloud results, higher customer usage and a revised infrastructure spending plan point to broader chip deployment. Google has raised its capital investment target to $93 billion, with funding directed toward expanding data centers, increasing compute access, and maintaining support for customers deploying AI systems at scale.
- TPUs Running in Production Across Large Deployments: Several companies now rely on TPUs for full-scale production workloads. One partnership includes plans to use up to one million chips for foundation model development and ongoing inference tasks. These deployments require consistent performance and infrastructure that can scale without major changes to software architecture. Customers using TPUs have cited gains in model responsiveness and training speed, with deployment flexibility allowing them to continue expanding without moving to a different platform.
- Access to TPUs Managed Through Google Cloud: Google does not sell TPUs as hardware units. Instead, it provides access through managed services on its cloud platform. This model allows the chips to be integrated with orchestration tools, automated maintenance, and scheduling systems that coordinate usage across data centers. Features like load balancing and fast recovery from hardware faults are built into the service layer, reducing downtime and improving reliability. Customers can concentrate on building and serving models while the underlying infrastructure handles distribution and system performance.
- Testing TPU Systems in Satellite Environments: Google is also preparing to evaluate TPUs beyond ground-based data centers. A research effort now underway includes plans to launch two solar-powered satellites equipped with TPU systems. Scheduled for 2027, these prototypes will test how well the chips perform in orbit without conventional power or cooling. The project is intended to explore new approaches to compute infrastructure, particularly in environments with limited physical resources or where long-term energy availability is a concern.
Trusted insights for technology leaders
Our readers are CIOs, CTOs, and senior IT executives who rely on The National CIO Review for smart, curated takes on the trends shaping the enterprise, from GenAI to cybersecurity and beyond.
Subscribe to our 4x a week newsletter to keep up with the insights that matter.
