Google is doubling down on custom AI hardware with the launch of two specialized 8th-generation Tensor Processing Units (TPUs). The new chips, referred to as TPU v8t and v8i, are purpose-built for the distinct demands of AI model training and inference, according to a post on Google's official blog.
A Specialized Strategy for a New AI Era
Google is moving beyond a one-size-fits-all approach to AI acceleration. In the announcement, the company detailed a split strategy for its new TPU family: the v8t is built for the computationally intense task of training large-scale AI models, while the v8i is engineered for the high-volume, low-latency demands of real-time inference.
This bifurcation addresses a long-standing tension in AI infrastructure, where the hardware needs for building a model differ substantially from those for deploying it.
Under the Hood: Performance and Efficiency Gains
According to Google, the new chips represent a leap in performance and efficiency over their predecessors, with the company tailoring silicon to specific workloads to maximize output and minimize costs for Cloud customers.
- TPU v8t: Optimized for large-scale training, with Google citing up to 2x the raw compute performance per chip over the previous generation. It uses liquid cooling and can scale into large pods for training foundation models.
- TPU v8i: Designed specifically for inference, focused on low latency and performance-per-watt, intended to make AI agents and real-time services more cost-effective to run at scale.
Building the Foundation for AI Agents
Google frames the launch as foundational for what it calls the "agentic era," in which AI systems perform complex, multi-step tasks on behalf of users — requiring both heavy training power and efficient, low-cost inference.
Why it matters
The launch signals that the AI infrastructure market is splitting into distinct hardware paths for training and inference. It also sharpens Google's competition with Nvidia and other chipmakers by offering a more specialized, end-to-end hardware ecosystem for its cloud customers.