1 comment

[ 0.24 ms ] story [ 12.9 ms ] thread
No distinction between learning and inference? Makes quite a difference in the compute needed.

Also "well over 1 million GPUs online by the end of 2025" is 1 GPU per ~8k world citizens. Not that crazy a number when talking about arguably #1 player in this field.

(And those GPUs will mostly be doing inference as that's where cash comes in. Training cost is more like R&D budget)