Our eighth generation TPUs: two chips for the agentic era

https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/eighth-generation-tpu-agentic-era/

Article

TL;DR

Google’s TPU 8t delivers 121 ExaFlops per pod — more than the top 10 supercomputers combined.

Key Takeaways

  • Single TPU 8t pod: 9,600 chips, 2PB HBM, 121 ExaFlops, 2x perf-per-watt vs prior gen
  • Full-stack ownership (silicon + software + datacenter) is a structural cost moat over Nvidia renters
  • Google’s model deprecation is aggressive: 1-year cycles, stricter rate limits than OpenAI

Discussion

Top comments:

  • [Keyframe]: Google quietly gaining strength while others capture news cycle
  • [pmb]: Whole-datacenter design context gives Google permanent cost efficiency edge
  • [WarmWash]: Gemini uses far fewer tokens than rivals — Google not maxing out inference compute
  • [mlmonkey]: One TPU 8t pod exceeds compute of entire top-10 supercomputer list

Discuss on HN


Type Link
Added Apr 22, 2026
Modified Apr 22, 2026
comments 141
hn_id 47862497
score 278
target_url https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/eighth-generation-tpu-agentic-era/