A lazy data structure exposes access methods that exhibit lazy-evaluator behavior. For example, you have a "rectangle" structure that contains position, width and height, and the structure has methods that let you apply matrix transformations to scale, rotate, and skew the rectangle's corners. The matrix methods perform the computation, and then cache that result in case it's needed again.
That is more about functional programming than laziness, but maybe I am splitting hairs. OpenGL, for example, implements the matrix stack functionality, and then it only works for one class. A lot of work then to get ky everywhere you need it.
Clojure is doing a lot of work with fusion and sequences to transfer laziness from arrays up to the other objects that consume them, but it is harder as you go farther away.
