[charleshn]

> There's a good reason so much research is done on Nvidia clusters and not TPU clusters.

You are aware that Gemini was trained on TPU, and that most research at Deepmind is done on TPU?