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 with 2PB shared HBM, targeting training; 8i targets inference.

Key Takeaways

  • One TPU 8t pod = 121 ExaFLOPS, dwarfing top 10 supercomputers combined at 11,487 PetaFLOPS
  • Vertical integration from silicon to model gives Google a structural cost advantage over NVIDIA-dependent rivals
  • Gemini uses fewer tokens than rivals — unclear if intentional efficiency or compute constraint

Discussion

Top comments:

  • [Keyframe]: Google has been quietly gaining market share while others grabbed headlines
  • [WarmWash]: Gemini uses far fewer tokens than Claude/GPT — either efficiency or compute rationing
  • [pmb]: At scale, Google’s whole-datacenter chip design beats anything NVIDIA can centralize
  • [mlmonkey]: One pod alone exceeds all top 10 supercomputers combined in raw flops

Discuss on HN


Type Link
Added Apr 22, 2026
Modified Apr 22, 2026
comments 167
hn_id 47862497
score 342
target_url https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/eighth-generation-tpu-agentic-era/