Any ASIC is possible.
Let me give an oversimplified example, to demonstrate the clear theoretical (though not necessarily practical) feasibility of doing it at all. Intel CPUs have floating point hardware. CryptoNight doesn't use floating point at all, so you could build a clone of an Intel CPU, get rid of the floating point, and have a slightly smaller/simpler chip that would cost less to manufacture, and probably use a bit less electricity.
You could call this an ASIC, in the sense that no modern computer can really function without floating point any more, so it would be almost useless for regular programs, at least Windows certainly wouldn't run on it (someone ambitious could still make a Linux soft-float build). You wouldn't really get much performance increase, but you could, in theory, get a price-performance increase and some power reduction.
There are certainly other components of the CPU that aren't needed for CryptoNight, and could be removed as well, to similar effect. I estimate that something in the neighborhood of 30% of the chip area of a desktop CPU (more for server CPUs that don't have a GPU) is directly being used for CryptoNight. So at best in this (somewhat unrealistic) model, you could reduce the chip size by 2/3. Or possibly you could build a specialized chip of similar size that would hash 3x as fast. A while back dga (early XMR miner who did most of the optimizations of the CryptoNight code, and computer science professor at CMU) once estimated there would be roughly a 5x gain from ASICs.
You have to trade that off against the R&D costs, economies of scale (massive manufacturing volume) of just using standard CPUs.
At a 3 billion market cap it is plausible there would still be no ASICs for CryptoNight. But it is possible there would be. It is difficult to say. How much is that hypothetical 5x gain worth relative to one time R&D costs and the expected useful life of such a chip?
We can be pretty sure that the performance and price-performance gain from a CryptoNight ASIC would be much smaller than for SHA256 or even Scrypt. We don't know the exact parameters until someone tries to build it.