Option 1: Design a High Memory Bandwidth FPGA/ASIC Board
Looking at the RX 470, we can do some quick math ($175 / (211 GB/s) ) to see that it costs $0.82 per GB/s.  Compared to a single GDDR5 chip (e.g. Micron EDW4032BABG), which costs $6.83 and has a bandwidth of 24 GB/s, we can do better at $0.28 per GB/s.  So, if we can build custom chip (either ASIC or FPGA) than can interface with 9 GDDR5 chips, we’ll have a memory bandwidth of 216 GB/s at a price of $61.47.  This still isn’t a finished mining product, since we need a FPGA or ASIC memory controller, circuit board, & support electronics.  If the shipped final assembly (adding additional parts, processes, testing, & logistics) costs less than the RX 470 (only $175), then the custom board will beat the GPU based card.  That is, until a faster, more efficient, cheaper graphics card comes to market.  For instance HBM graphics cards cards are already available.  But, if you find low cost, off-the-shelf, FPGA or ASIC chips with 5-10 DDR or HBM memory controllers, or your company is experienced in building custom high memory bandwidth ASIC solutions, you may be able to beat out off-the-shelf hardware.  But, if in this situation, you should probably change your business model and build graphics cards instead, since that’s already a huge market.