In the suburbs, you pay the gas bill, the water bill, the electric bill, the lawn guy, the property taxes and so on. In the city, you pay the rent, period.
Beyond the simple costs of server purchase and maintenance, there are licensing and software support costs. Linux may be free-as-in-beer, but most server OSs cost money. Your management suite probably has a per-server license, as does your antivirus software. Add in the cost of support contracts for the operating systems and service contracts for the servers, and you're spending a good pile of cash on each machine. Fewer machines, fewer costs.
Server consolidation comes in many forms and flavors, but the most popular methods are virtual servers, 1U consolidation and blade servers.
Virtual servers: This method is intriguing, but frightening. Software like VMware and Connectix (recently purchased by Microsoft) let you buy one big server and run many instances of the needed OS--thus the "virtual servers" moniker. Unisys, for example, makes a huge 32-processor Intel machine, called the ES7000 Series. Using VMware or Connectix, you could replace as many as 32 single-processor servers with virtual machines. Personally, this approach gives us the "eggs all in one basket" willies, but many companies are buying in.
Consolidation to a rack: Stacking identical 1U servers is a proven method, but it's a bit old school. On the plus side, it generally gets a company to a single hardware platform, but you may not appreciably reduce the overall number of servers in the rack as you would with blade servers.
Blade servers: The last method, and the one that we think is conceptually closer to the utility data center ideal is the blade server. Blade servers are special chassis made up of multiple processing blades. They have centralized control centers and can be teamed, repurposed and redeployed in a rapid, centralized fashion--well, that's the idea, anyway. To see how well theory matches reality, check out "Pitching Blades".