Google reveals 9,600-chip TPU 8t; OpenAI details 3-gen chip roadmap

SumitGup · x · 2026-08-27

At Hot Chips '26, Google unveiled the TPU 8t, a 9,600-chip training system that doubles performance per watt, allowing data centers to train twice as many tokens within the same power budget. Google emphasized separating training and serving chips to eliminate wasted silicon.

Simultaneously, OpenAI disclosed a multi-generational roadmap for its in-house chips: Gen 1 is step one, Gen 2 is nearing tape-out, and Gen 3 is already operational. OpenAI stated this strategy aims to lower infrastructure costs, highlighting partnerships with Broadcom and Celestica.

Original post →

More from Companies & People

Companies & People channel →