We’re launching two specialized TPUs for the agentic era.

Source: Google Date: 2026-04-22 URL: https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/tpus-8t-8i-cloud-next/

Summary

Google announced the TPU 8i for enabling AI agents to rapidly execute multi-step reasoning, planning, and action tasks, and the TPU 8t for training complex models using a unified large memory pool. Together the chips are positioned to “deliver highly responsive agentic AI to the masses” as part of Google’s AI Hypercomputer infrastructure stack.

Implications

This is the same TPU 8t/8i announcement covered in the “Our eighth generation TPUs” post — this version is the shorter, consumer-facing framing while the other post covers technical specs. The “agentic era” label applied to the chip names is deliberate positioning: Google is framing hardware generation around a workload paradigm shift, not just performance increments.

The “to the masses” language is worth flagging — it implies Google intends to make TPU 8i accessible via pricing that enables broad adoption, not just top-tier enterprise customers. If that holds, it changes the inference cost floor for anyone running agents on Google Cloud relative to GPU-based alternatives.

Low standalone analytical weight relative to the detailed TPU 8t/8i technical post — refer there for depth. This signal is the narrative wrapper.

Watch:

• Whether “to the masses” translates to a mid-tier Cloud pricing tier specifically for agent workloads
• GA timing relative to major Gemini model releases — the chip schedule determines when the next generation of Gemini agents can ship