At the Google Cloud Next conference on April 20, 2026, Google Cloud announced the general availability of its seventh-generation Tensor Processing Unit, designated as TPU v7 and codenamed Ironwood. This latest iteration of Google’s custom-designed AI accelerator is specifically engineered to address the requirements of the inference phase of artificial intelligence, where the computational focus has shifted from training foundational models to the high-volume deployment of agentic AI workflows. The TPU v7 is designed to handle models exceeding two trillion parameters with significantly reduced latency and operational overhead.
Technically, the TPU v7 represents a substantial leap in performance density. Built on a 3-nanometer manufacturing process, each TPU v7 chip delivers 4.6 petaflops of peak compute at FP8 precision. The hardware incorporates 192GB of HBM3e high-bandwidth memory, providing the necessary data throughput for real-time multi-modal reasoning. A key architectural advancement is the expansion of the superpod configuration, which now allows for the synchronous interconnection of up to 9,216 chips. This scaling is facilitated by the third generation of Google’s proprietary Optical Circuit Switching technology, which enables dynamic reconfiguration of the cluster topology to bypass hardware failures without interrupting active workloads.
Energy efficiency remains a central metric for the v7 generation. Google reported that the TPU v7 achieves a 50% improvement in performance-per-watt compared to the TPU v6 series released in 2025. This efficiency is partly attributed to the integration of liquid cooling at the rack level and the use of the Titanium offload system, which handles networking and security tasks to free up the main processor for AI computations. According to internal benchmarks released by Google, the TPU v7 provides a 44% lower total cost of ownership for large-scale inference tasks compared to equivalent third-party GPU configurations.
The conference also saw the introduction of the second-generation Axion ARM-based CPU. These processors are designed to work in tandem with TPU v7 clusters, managing general-purpose computing and data preprocessing. Google confirmed that Anthropic has committed to utilizing over one million TPUs to serve its Claude model family, citing the hardware's price-to-performance ratio for long-context window processing.
TPU v7 instances are available in preview starting today in select North American and European data center regions. General availability for enterprise customers is scheduled for the third quarter of 2026. This release solidifies Google’s strategy of vertical integration, offering a complete AI infrastructure stack that includes custom silicon, specialized networking, and the Vertex AI software platform.