It's too bad we can't make the POW somehow dynamic so that any specialized hardware is impossible, and only GPU / FPGA is possible.

Maybe a variant of Keccak where the size of the sponge is increased along with additional zero bits required. Seems like this would either have to resist specialized hardware or imply sha3 is compromised such that the size of the sponge does not incerase the number of possible output bits as expected.