Ocean tides generate electromagnetic (EM) signals that are emitted into space and can be recorded with low-Earth-orbiting satellites. Observations of oceanic EM signals contain aggregated information about global transports of water, heat, and salinity. We utilize an artificial neural network (ANN) as a non-linear inversion scheme and demonstrate how to infer ocean heat content (OHC) estimates from magnetic signals of the lunar semi-diurnal (M2) tide. The ANN is trained using monthly OHC estimates based on oceanographic in-situ data from 1990-2015 and the corresponding computed tidal magnetic fields at satellite altitude. We show that the ANN can closely recover inter-annual and decadal OHC variations from simulated tidal magnetic signals. Using the trained ANN, we present the first OHC estimates from recently extracted tidal magnetic satellite observations. Such space-borne OHC estimates can complement the already existing in-situ measurements of upper ocean temperature and can also allow insights into abyssal OHC, where in-situ data are still very scarce.