Google TPU 8th Gen: Scaling AI Training with TPU 8t & TPU 8i

Google introduces its 8th generation TPU with TPU 8t optimized for training and TPU 8i for inference, scaling to 9k units in a single cluster.

Google Cloud has unveiled its 8th generation of Tensor Processing Units (TPUs), introducing a bifurcated architecture with the TPU 8t and TPU 8i. The TPU 8t (Training) is engineered for the massive compute requirements of foundation model pre-training, while the TPU 8i (Inference) is optimized for high-throughput, low-latency deployment in production environments.

The new TPUs can scale to clusters of 9,000 units, providing unprecedented compute density. Google claims the 8th gen architecture offers a 3.5x performance-per-watt improvement over the previous generation, addressing the critical energy consumption challenges facing modern data centers. The TPU 8i specifically targets the growing demand for real-time agentic reasoning, where fast inference is the primary bottleneck.

This story is part of our May 05, 2026 Tech Pulse briefing. We are deep-diving into the architectural shifts, security implications, and market impact of this development.