Google unveils 8th-generation TPU chips: two variants for the agentic AI era

On April 22, 2026, at its Cloud Next ‘26 conference, Google introduced the eighth generation of TPU (Tensor Processing Unit) chips. For the first time in the line’s history, Google is offering two specialized variants — one for training, one for inference of autonomous AI agents.

What two TPU variants does Google offer?

Google named the new chips TPU 8t and TPU 8i, where the letter suffixes denote the primary workloads. TPU 8t (training) is focused on training complex models and features a large unified memory pool, allowing large models to be hosted without splitting across multiple nodes.

TPU 8i (inference), on the other hand, is designed for what Google calls the “agentic era” — a world in which autonomous AI agents must perform multi-step reasoning, planning, and task execution in real time. Specialization means inference workloads are no longer run on chips optimized for training, which should reduce latency and improve cost-efficiency for AI agent products.

Google emphasizes that this is the first generation of TPUs purpose-designed with agentic AI as the primary use case, not as a byproduct of training hardware.

Why does agentic AI require different hardware?

A classic chatbot or generative AI model generates a response in a single pass through the neural network. Agentic systems, however, operate very differently — an agent plans a task, calls tools, processes results, reasons again, and repeats, sometimes dozens of times per user query.

This working pattern means that the latency of each individual inference call becomes critical. If every agent step takes 2 seconds, ten steps mean 20 seconds of waiting — unacceptable for interactive applications. Google claims that TPU 8i is optimized precisely for this pattern: fast, responsive inference for multi-step workflows.

The second difference is economic. In agentic use, the number of inference calls can be 10–50 times higher than in classic generative use, making cost per token far more important. A dedicated inference chip theoretically consumes less energy per inference call than a chip designed for training.

What does this mean for Google Cloud and competitors?

The announcement comes at a time when Nvidia dominates the AI hardware market with Blackwell and upcoming generations, while AMD aggressively pushes its MI series. Google’s approach of two specialized variants is an attempt to differentiate on use case optimization, not just raw performance.

Google emphasizes a “full-stack purpose-built infrastructure” — a combination of chips, networking, data centers, and energy efficiency that should deliver “responsive agentic AI at massive scale.” TPU 8t and 8i are available exclusively through Google Cloud, meaning clients such as Anthropic (which uses TPUs for Claude models) and Google’s own Gemini models benefit directly.

It remains to be seen whether Google will publish concrete benchmark numbers and per-inference pricing. For now, the announcement is more strategic — a signal that agentic AI has matured to the point where it justifies its own generation of hardware, rather than just a new software layer on top of existing chips.