The most expensive part of a software project is the people, not the licenses. People who pay for MS believe that they're saving money in the long run.
The increased engineering burden of having to consider software license costs when designing systems hardly seems worth it. Imagine if you had to pay a per-node fee for your Hadoop cluster? How much extra effort would be put into optimizing the individual performance of each node and job if there were an additional 50% per-node cost to run it?
But you already go through the exercise; you compare the cost of adding a server versus the cost of engineering a solution. The only difference is server cost is different. My time's worth $200/hr, what's yours?
Hence the use of the phrase "extra effort." In a theoretical world where Microsoft offered "Microsoft Map Reduce" as a Microsoft-sanctioned way to run map reduce jobs, one could imagine the licensing cost for a single box could well exceed the cost of hardware, ops, etc, if SQL server license fees are any indication. How much extra engineering effort does this translate into? Is it worth it?