Have you considered the option of using ready-made server? For example:
Cluster Compute Instance on Amazon Web Services
( 2xNVIDIA Tesla M2050 "Fermi" GPUs., 2x quad-core Intel "Nehalem" X5570, 22 GB of RAM )
Hourly payment: 2.10$ per hour.
Advantages — you can start to use right now, scales as needed.
Disadvantages — price ( compared to a private server )