alt Hacker NewsI think data centers and power are the dark-fiber of this current "bubble".
Inference is 2/3rds of the power used in AI, and I expect the majority of inference to go local in the next 2-3 years, about the time these data centers will come online.
The other 3rd is training, and the major labs are training 6-7 times a year right now. But I suspect we will see a few things. 1) training iterations will slow as the improvements between training runs gets lower and lower. Is it worth spending an extra $500M(ish) for the small improvement in quality? Who is paying for that?
2) Training improvements and improvements in GPUs will require less power. Along with fewer training runs, as we add more and more tokens, the amount of power needed for a training run will likely asymptote at some point.
The dark fiber comparison is very close. I remember contracting bandwidth with Global Crossing in the late 90s/early 2000s when they were HOT and two years later they were bankrupt.