If you have two or more servers doing the same task, you need load balancing, synchronisation, methods to remove nodes automatically from the custer when they fail, etc. These costs are fixed, i.e. they are approximately the same for a cluster with 5 and a cluster with 1000 nodes. In Google's situation with thousands of servers, these software costs per node are small, for you they are probably the largest part in the equation.

If you must decide between one powerful failsave server or two low-cost nodes, the first will be cheaper.

On the other hand if the choice is between five powerful server or then cheap ones, you should probably go for the latter because you have to invest in the software infrastructure anyway.

One server is also more reliable than 2. But in this example I'm talking about a database server and a webserver. Specifically, you have a more reliable set up running both apache and mysql on the same box. In fact, it should be twice as reliable as running on two boxes.