We develop methods to calculate the curvature power spectrum in models where features in the inflaton potential nonlinearly excite modes and generate high frequency features in the spectrum. The first nontrivial effect of excitations generating further excitations arises at third order in deviations from slow roll. If these further excitations are contemporaneous, the series can be resummed, showing the exponential sensitivity of the curvature spectrum to potential features. More generally, this exponential approximation provides a power spectrum template which nonlinearly obeys relations between excitation coefficients and whose parameters may be appropriately adjusted. For a large sharp step in the potential, it greatly improves the analytic power spectrum template and its dependence on potential parameters. For axionic oscillations in the potential, it corrects the mapping between the potential and the amplitude, phase and zero point of the curvature oscillations, which might otherwise cause erroneous inferences in for example the tensor-scalar ratio, formally even when that amplitude is $10^3$ times larger than the slow roll power spectrum. It also estimates when terms that produce double frequency oscillations that are usually omitted when analyzing data should be included. These techniques should allow future studies of high frequency features in the CMB and large scale structure to extend to higher amplitude and/or higher precision.