Licensing – more machines or more power
The best way to select hardware is to understand the software you are going to use it for. ArcGIS for Server is licensed and based on processer cores, not users; this means the more cores you have, the more expensive the product becomes. For example, let us say you want to purchase a four-core license. This can give you many hardware solution options: you can get one server with quad-core CPU, or you can get two dual-core servers. If you could even find a single-core server these days, you can get four of those. But the question is which one of those solutions is suitable? According to Moore's law, computer power doubles every 18 months; this can be applied to the price as well, where the cost of the hardware drops. Although this depends on other factors including the clock speed of the core itself, having four single-core servers can theoretically give the same power as a one quad-core server given that they have relatively the same frequency. There is one thing which is important to be noted: not all the ArcGIS for Server tasks run on parallel processing. This means if you have four tasks or requests to be executed, it is more efficient to line them up on four moderately strong servers rather than stacking them up on one powerful server.
Four physical servers have their own dedicated memory and bus routes linked in a high-speed cluster, which gives more availability than one quad-core server and it is definitely cheaper. This way you can save a lot for the same or even more power; the key is not to be tricked by gimmicks of the hardware marketers and to know your requirements before purchasing the hardware. Esri has a dedicated website that compares different hardware and provides recommendations for the best hardware for Server; you can take a look at it at http://www.esri.com/systemdesign.