It's not cheap. If you want to compete wit cpu/fpga/gpu ,you'll need to use the latest manufacturing processes and do complex designs,and now we're talking about more than 100 million USD investment.
However, algorithms are so rarely the stumbling block for applications; data storage, management and communications are. The major exception is graphics, hence the GPU.
The money is a huge barrier to entry for hobbyist types, but if we're talking about commercial stuff, it'll probably cost you a lot more to pay for the engineers who are competent enough to implement something that will actually work than it will to do the fabrication.