Our eighth generation TPUs: two chips for the agentic era
Article
TL;DR
Google’s TPU 8t delivers 121 ExaFlops per pod with 2PB shared HBM — 10x all top supercomputers combined.
Key Takeaways
- Single TPU 8t pod = 121 ExaFlops; top 10 supercomputers total only 11,487 PetaFlops
- Google owns full stack — silicon, software, data centers — structural cost advantage over NVidia renters
- Gemini produces fewer tokens than rivals despite having most compute; reasoning depth unclear
Discussion
Top comments:
- [TheMrZZ]: 9,600 chips + 2PB shared HBM per superpod is a genuine competitive moat
-
[pmb]: At scale, whole-datacenter design lets Google always be more cost-efficient than NVidia
when you are doing big AI you basically have to buy it from NVidia or rent it from Google
- [WarmWash]: Gemini uses drastically fewer tokens than ChatGPT/Claude — intentional budget constraint?
- [jjice]: Owning your hardware and full stack is huge; Google clowned on early but positioned well