The recently identified Prompt Elasto-Gravity Signals (PEGS), generated by large earthquakes, propagate at the speed of light and are sensitive to the earthquake magnitude and focal mechanism. These characteristics make PEGS potentially very advantageous for earthquake and tsunami early warning. PEGS-based early warning does not suffer from the saturation of magnitude estimations problem that P-wave based early warning algorithms have, and could be faster than Global Navigation Satellite Systems (GNSS)-based systems while not requiring a priori assumptions on slip distribution. We use a deep learning model called PEGSNet to track the temporal evolution of the magnitude of the 2010 $\textrm{M}_{\textrm{w}}$ 8.8 Maule, Chile earthquake. The model is a Convolutional Neural Network (CNN) trained on a database of synthetic PEGS – simulated for an exhaustive set of possible earthquakes distributed along the Chilean subduction zone – augmented with empirical noise. The approach is multi-station and leverages the information recorded by the seismic network to estimate as fast as possible the magnitude and location of an ongoing earthquake. Our results indicate that PEGSNet could have estimated that the magnitude of the Maule earthquake was above 8.7, 90 seconds after origin time. Our offline simulations using real data and noise recordings further support the instantaneous tracking of the source time function of the earthquake and show that deploying seismic stations in optimal locations could improve the performance of the algorithm.