What is the intuition behind support vector sparsity?

Only a subset of training points define the SVM model.

SVM solutions are sparse because only points on or within margin constraints have non-zero Lagrange multipliers. This makes prediction efficient since only support vectors are needed.