Switching to an ASIC resistant PoW coin doesn't solve this problem but merely delays the inevitable. As interest and hash power grows ASICS will be developed within time regardless.
I believe it is possible to design a memory hard PoW that is not electrically more efficient on an ASIC, but it will be very slow. I originally didn't think so, but have since realized I had a mistake in my 2013/4 research on memory hard hashes. It is possible that Cuckoo Hash already achieves this, but it is more difficult to be certain and it is very slow when DRAM economics are maximized (although it adds asymmetric validation which is important for DDoS rejection if the transaction signatures are ECC and not Winternitz and for verification when PoW share difficulty can't be high because each PoW trial is so slow).
Cryptonote's memory hard hash can't possibly be ASIC resistant, because by my computation it could not possibly have 100 hashes/second on Intel CPUs and be ASIC resistant.
See also Zcash's analysis thus far.
Correction follows.
It will be impossible to design a memory hard PoW that is not electrically more efficient on an ASIC, unless the hash function employed (for randomizing the read/writes over the memory space) is insignificant w.r.t. the RAM power consumption, which is probably not going to be the case in any design where that hash function has sufficient diffusion to be secure.
The only way to make an ASIC resistant PoW is for the proving computation to be memory latency bound, because DRAM latency can't be improved much in general (whereas hardwired arithmetic computation and memory bandwidth can be accelerated with custom hardware):
http://community.cadence.com/cadence_blogs_8/b/ii/archive/2011/11/17/arm-techcon-paper-why-dram-latency-is-getting-worse
http://www.chipestimate.com/techtalk.php?d=2011-11-22
However, what a GPU (which starts with 4 - 10X worse main memory latency than CPUs) and especially an ASIC will do to get better DRAM amortization (if not also lower electricity consumption due to less latency) is run dozens or hundreds of instances of the proving algorithm with the memory spaces interleaved such that the latencies are combined and amortized over all instances, so that the effective latency drops (because reading from the same memory bank of DRAM is latency free if multiple accesses within the same bank are combined into the same transaction). This can even be done in software as interleaved memory spaces without needing a custom memory controller. More exotic optimizations might have custom memory controllers and larger memory banks (note I am not expert on this hardware issue). This is probably why Cryptonote includes also AES-NI instructions because GPUs have only at best at parity in performance per watt on AES, but that won't be enough to stop ASICs.
However that optimization for ASICs will bump into memory bandwidth limit so the amortization will have a limit. Theoretically memory bandwidth can be increased with duplicated memory banks for reads but not for writes!
Using larger memory spaces in a properly designed memory hard PoW hash function (not Scrypt) can decrease the probability of that instances will hit the same memory bank within a sufficiently small window of time necessary to reduce the latency. Also using wider hash functions (e.g. my Shazam at 2048 to 4096-bits) reduces the number of instances that can be interleaved in the same memory bank (and standard DRAM I think has bank/page size of 4KB?). The ASIC can respond by designing custom DRAM with larger memory banks and run more instances, but that not only raises the investment required but the memory bandwidth limit for writes seems to be an insurmountable upper bound.
So although I think a memory hard PoW hash can be made which is more ASIC resistant than current ones, I think it will be impossible to sustain parity in hashes/Watt and hashes/$hardware. Perhaps the best will be within 1 to 2 orders-of-magnitude on those.
So all profitably mined PoW coins (with sufficient market caps) are destined to be centralized into ASIC mining farms running on cheap or free electricity, but the scale and rate at which this happens can be drastically improved over SHA256 (Bitcoin, etc).
My design of unprofitably mined PoW will only require that the difficulty from the PoW shares sent with transactions is sufficient to making ASIC mining unprofitable for the level of block reward offered. Keeping the CPU implementation of the PoW prover within 1 to 2 orders-of-magnitude of an ASIC implementation reduces the level of such aforementioned difficulty needed.
I hope I didn't make another error in this corrected statement. It is late and I am rushing.