For 100 user load any modern Windows box will more than suffice – even with zero delay. For 5,000 user load (assuming non-zero delay now) – five VM’s recommended so can deploy one BMF agent per VM. Those VM’s would either be on separate hosts or one big host – where 'big' means sufficient CPU, memory and NIC’s. In fact, NIC bandwidth would be major factor for single box – so either 10 gigabit or dual/quad gigabit with bonding/teaming.
Additional Notes:
For non-zero delay: I7-4770 cpu (quad+hyper-threading), 16 GB ram min, dual gigabit NIC bonded. If zero-delay to approx. more users – then small server with dual quad/hexa core with hyper-threading.
For maximum user load 4MB of memory is sufficient for all general tests.