Orthogonal Frequency Division Multiplexing (OFDM) is the workhorse of current 5G deployments due to its robustness in quasi-static channels and efficient spectrum use. However, in high-mobility scenarios, OFDM suffers from inter-carrier interference (ICI), and its reliance on dense pilot patterns and cyclic prefixes reduces spectral efficiency significantly. In this work, we propose Deep-OFDM : a learnable modulation framework that augments traditional OFDM by incorporating neural parameterization. Instead of mapping each symbol to a fixed resource element, Deep-OFDM spreads information across the OFDM grid using a convolutional neural network modulator. This modulator is jointly optimized with a neural receiver through end-to-end training, enabling the system to adapt to time-varying channels without relying on explicit channel estimation. Deep-OFDM outperforms conventional OFDM when paired with neural receiver baselines, particularly in pilot-sparse and pilotless regimes, achieving substantial gains in BLER and goodput, particularly at high Doppler. In the pilotless setting, the neural modulator learns a low-rank structure that resembles a superimposed pilot, effectively enabling reliable communication without explicit overhead.These results highlight the potential of transmitter–receiver co-design for robust, resource-efficient communication in challenging channel conditions, paving the way for AI-native PHY designs in next-generation wireless systems.