I study the problem of renormalizing a non-renormalizable theory with a reduced, eventually finite, set of independent couplings. The idea is to look for special relations that express the coefficients of the non-renormalizable terms as unique functions of a reduced set of independent couplings λ, such that the divergences are removed by means of field redefinitions plus renormalization constants for the λs. I consider non-renormalizable theories whose renormalizable subsector R is interacting. The "infinite" reduction is determined by i) perturbative meromorphy around the free-field limit of R, or ii) analyticity around the interacting fixed point of R. In general, prescriptions i) and ii) mutually exclude each other.When the reduction is formulated using i), the number of independent couplings remains finite or slowly grows together with the order of the expansion. The growth is slow in the sense that a reasonably small set of parameters is sufficient to make predictions up to very high orders. Instead, in case ii) the number of couplings generically remains finite. The infinite reduction is a tool to classify the non-renormalizable interactions and address the problem of their physical selection.