Google Cloud announced the launch of its eighth-generation Tensor Processing Units (TPUs) and a new $750 million financial initiative during its annual infrastructure summit on April 22, 2026. The hardware release includes two distinct chips: the TPU 8t, optimized for large-scale model training, and the TPU 8i, specifically engineered for high-throughput inference. These developments represent a significant expansion of Google’s custom silicon portfolio as the company seeks to address the increasing computational demands of agentic artificial intelligence.
The TPU 8t delivers a reported 2.8x increase in peak FLOPs compared to the previous generation, utilizing a new 3-nanometer manufacturing process. It features enhanced high-bandwidth memory (HBM4) and utilizes Google’s proprietary Optical Circuit Switching (OCS) technology to scale to pods containing up to 16,384 chips. According to technical specifications released by Google, the TPU 8t reduces training time for trillion-parameter models by approximately 40% while maintaining a 25% improvement in energy efficiency per watt.
Complementing the training hardware, the TPU 8i is designed to handle the real-time execution of complex AI agents. Google Cloud CEO Thomas Kurian stated that the TPU 8i offers a 3.5x improvement in inference throughput for large language models (LLMs) and a 50% reduction in total cost of ownership (TCO) for enterprise customers. The chip is integrated directly into Google Cloud’s Vertex AI platform, allowing developers to deploy autonomous agents that require low-latency responses for multi-step reasoning and tool use.
Alongside the hardware unveiling, Google Cloud introduced the Agentic AI Development Fund, a $750 million capital pool intended to accelerate the creation of autonomous AI systems. The fund is structured to provide technical credits, engineering support, and direct financial grants to partners and startups building on Google’s infrastructure. Eligibility for the fund requires participants to utilize Gemini models and the new TPU 8-series hardware for their primary workloads.
The announcement also detailed updates to the Google Distributed Cloud (GDC) environment, which will now support TPU 8i clusters for edge computing and sovereign cloud requirements. This integration allows regulated industries, such as healthcare and finance, to run agentic workflows locally while benefiting from the performance gains of the eighth-generation silicon. Google confirmed that general availability for TPU 8t and 8i instances will begin in select regions starting in June 2026, with full global rollout expected by the end of the calendar year.