My naive assumption would be, takes 2^36 cpu operations

2^36 arithmetic operations is what is claimed. That's not quite the same as CPU operations, because we only have 64-bit CPUs with up to 512-bit vector instructions, but we're talking about factoring 800-bit numbers. So we need to allow for several CPU instructions to implement each of the required arithmetic operations.

If so, 2^36 ~ 7 x 10^10, so a few seconds on GHz processors.

Yeah I would be great if they could translate that into core-years to match the references they listed

I'm guessing it's a shorthand for the order of units of work. log2(8.4E10) = 36.3 bits of operations