The compute bill for one epoch
lean
rich
model grows with data
data
model
GPUs
≤ 1 year
> 1 year
your setting