The Neural Tangent Kernel (NTK) viewpoint represents a valuable approach to examine the training dynamics of Physics-Informed Neural Networks (PINNs) in the infinite width limit. We leverage this perspective and focus on the case of nonlinear Partial Differential Equations (PDEs) solved by PINNs. We provide theoretical results on the different behaviors of the NTK depending on the linearity of the differential operator. Moreover, inspired by our theoretical results, we emphasize the advantage of employing second-order methods for training PINNs. Additionally, we explore the convergence capabilities of second-order methods and address the challenges of spectral bias and slow convergence. Every theoretical result is supported by numerical examples with both linear and nonlinear PDEs, and we validate our training method on benchmark test cases.