Orbital angular momentum (OAM) can be employed as information carrier in optical communications, which include OAM multicasting where separate OAM beams are produced to construct one-to-many links, and OAM shift keying where different OAM states are used for symbol coding. However, it remains a significant challenge to achieve a hyperscale OAM multicasting with OAM shift keying simultaneously, due to the deviation of complex OAM spectral components at the transmitter. Here, we demonstrate the concept of OAM neural communication by developing a high-precision, noniterative and ultra-multi-user optical communication coding scheme based on a deep-learning U-net neural network. The neural network is trained with only small datasets (there are 10 10 datasets in total, but only 10 3 datasets are used) to predict the OAM spectrum of ultra-high precision, with the mean square error (MSE) evaluated as 0.02 in all OAM channels. As a proof of concept, we experimentally demonstrate 1-to-40 multicasting with 16-ary shift keying in OAM communication. After a 3-meter free space transmission, we realize the encoding and decoding of 40 images with 16 gray values with a BER of 0. Moreover, our artificial intelligence-based methodology can be extended to manipulate three-dimensional OAM pixel arrays for holographic display and optical trapping.