What is the role of Gram matrix in kernel SVM?

Gram matrix stores all pairwise kernel similarities between training points.

The Gram matrix K(i,j)=K(x_i,x_j) replaces explicit feature mapping in kernel SVM. It is central to dual optimization but has O(n²) memory cost, making it a scalability bottleneck.