Story perspectives

Google Cloud launches TPU 8t/8i, 124% efficiency, $750M fund

4/23/2026

Summary

Google Unveils TPUs

Google Cloud announced TPUs—TPU 8t and TPU 8i—on April 22 2026, with 124 % and 117 % more performance per watt.
The eighth-generation TPUs can scale to 9,600-chip pods (~121 exaflops) and, for inference, deliver 80 % better performance-per-dollar with 288 GB HBM and 384 MB SRAM.
Google said the chips enable three-times faster training and five-times lower latency for agentic AI, announced a $750 million fund for firms to adopt AI agents, and will keep selling Nvidia GPUs.

Summary

Google Cuts Nvidia-Tax

Google’s vertical-integration eliminates Nvidia’s high GPU margin fees, avoiding the “Nvidia tax.”
Amin Vahdat says designing all six AI-stack layers together prevents least-common-denominator compromises.
The integrated stack promises lower cost-per-token for training and inference.
Analysts say TPU v8 will shift 2026-27 cloud procurement toward availability windows, latency benchmarks, and HBM sizing.
Vahdat predicts a CPU resurgence with Axion Arm CPUs hosting TPU pods at a 2:1 ratio.

1 of 2

Google Unveils TPUs

Google Cloud announced TPUs—TPU 8t and TPU 8i—on April 22 2026, with 124 % and 117 % more performance per watt.
The eighth-generation TPUs can scale to 9,600-chip pods (~121 exaflops) and, for inference, deliver 80 % better performance-per-dollar with 288 GB HBM and 384 MB SRAM.
Google said the chips enable three-times faster training and five-times lower latency for agentic AI, announced a $750 million fund for firms to adopt AI agents, and will keep selling Nvidia GPUs.

1 / 2