Not true. ConvNets were actually relatively cheap compared to the dominant models of object recognition circa 2007.
They were much smaller than current architectures, of course.

→ View original post on X — @ylecun