"We haven't found the 'right' algorithm yet." seems like the obvious answer, but the numbers in the paper all make sense and I'm interested in some more exotic explanations why it could actually be some orders of magnitude more than a 5090.

Although that's not looking at memory, and I am also interested in some explanation of what... a 5090 has 32GB which, a human brain has more like a petabyte of memory assuming 1 byte/synapse. Which is to say 1 million GB in which case even a large cluster of H100s has an absurd amount of TOPS but nowhere near enough high-speed memory.