Something that I don’t think has been emphasised enough in the discussion so far:
The Equihash parameters that Zcash selected for launch {n=200, k=9} were suboptimal.
What do I mean by this? The peak memory required for optimally solving Equihash with parameters {n=200, k=9} was estimated to be much higher than it turned out to require in practice. Just before launch, the open-source mining competition successfully produced miners that required much less memory than previous implementations. However, there was insufficient time to modify the Equihash parameters prior to the launch itself.
(The particular Equihash parameters used also produced relatively large proofs compared with other possible parameters, but that’s besides the point.)
In order to achieve some level of ASIC resistance, I think the choice of Equihash parameters should be revisited, and parameters that require much higher peak memory should be used. It will still be possible to develop an ASIC for adjusted parameters, but it will be far less economical versus GPUs, e.g. if 20× the memory is required.
Furthermore, note that changing the Equihash parameters is relatively straightforward for Zcash, and its mining ecosystem (pools and miners) will be able to adapt relatively quickly.