There isn't an explicit mathematical concept of a cluster.

Here's what K-means does. Say, K is 3.

You try all the possible ways to partition your data points into three groups. You pick the partition that minimizes the sum of squared differences within each group.
Then you iterate the procedure.