In a move that radiates the same energy as showing up to your ex’s party looking incredible: Google just dropped not one but TWO new AI chips at Cloud Next, and Nvidia’s stock briefly flinched.

Meet the TPU 8t (for training) and TPU 8i (for inference) — because apparently, the era of one chip to rule them all is over. Now you need a specialized chip for learning things and a different specialized chip for pretending you already knew them. Relatable, honestly.

🤚 The Training Chip (TPU 8t)

Nearly 3x compute per pod over the previous generation. Scales to 9,600 chips. Delivers 121 ExaFLOPS, which is a number so large it has the same emotional impact as your unread email count — you know it’s big, you just can’t process what it means anymore.

👐 The Inference Chip (TPU 8i)

80% better performance-per-dollar, 288GB of HBM, and specifically optimized for mixture-of-experts models. For those keeping score at home, “mixture of experts” is when your AI model has the computational equivalent of a group project where only two people actually do the work. Just like real life.

👑 The Gold-Leaf Analysis

Both chips support open frameworks like JAX, PyTorch, and vLLM, which is Google’s way of saying “please stop giving all your money to Nvidia, we also make things.” The split into dedicated training and inference hardware reflects a mature industry realization: the thing that learns and the thing that answers are fundamentally different beasts.

Sort of like how the person who writes the company handbook is never the person who actually follows it.

Nvidia declined to comment but was last seen doing 121 ExaFLOPS worth of deep breathing exercises.