Google Cloud today announced the launch of its eighth-generation Tensor Processing Units (TPUs) and a $750 million investment fund at the Google Cloud Next '26 conference in Las Vegas. The new hardware suite, consisting of the TPU 8t for training and the TPU 8i for inference, is specifically engineered to support agentic AI—autonomous systems capable of multi-step reasoning and task execution. The accompanying $750 million Agentic AI Development Fund is designed to accelerate the deployment of these technologies across Google’s global partner ecosystem.
The TPU 8t is positioned as the company’s primary training engine for frontier-scale models. According to technical specifications released by Google, a single TPU 8t superpod can scale to 9,600 chips, providing 121 ExaFlops of compute power and two petabytes of shared high-bandwidth memory. This architecture delivers nearly three times the compute performance of the previous generation, which Google referred to as the Ironwood series. The system utilizes doubled inter-chip interconnect (ICI) bandwidth and a new Virgo Network fabric, allowing for clusters of more than one million TPUs to be orchestrated as a single seamless supercomputer.
For inference and reinforcement learning, Google introduced the TPU 8i. This chip is optimized for the ultra-low latency requirements of agentic workflows and Mixture of Experts (MoE) models. Technical improvements include a tripling of on-chip SRAM to 384 MB and an increase in high-bandwidth memory to 288 GB. The TPU 8i also features a dedicated Collectives Acceleration Engine (CAE) that reduces on-chip latency by up to five times. Google claims the 8i delivers an 80 percent improvement in performance-per-dollar for inference compared to its predecessor. The hardware utilizes a Boardfly topology to directly connect 1,152 TPUs in a single pod.
The $750 million Agentic AI Development Fund aims to provide financial resources, technical credits, and engineering support to systems integrators, consulting firms, and software partners. Google Cloud CEO Thomas Kurian stated that the fund will support AI value identification, prototyping, and the embedding of Google forward-deployed engineers into partner teams. The initiative is part of a broader strategy to establish the Agentic Enterprise, supported by the newly launched Gemini Enterprise Agent Platform, which provides tools for building, governing, and optimizing autonomous agents.
Additional infrastructure updates include the integration of eighth-generation TPUs with Google’s custom Axion Arm-based processors and the Virgo data center fabric. Sundar Pichai, CEO of Google and Alphabet, noted that the infrastructure is designed to handle the shift from reactive systems to proactive systems of action. The TPU 8i is currently available in private preview, with the TPU 8t expected to reach early access later in 2026.