In this work we explored design options for a future Atlantic-scale observational network enabling the release of carbon system estimates by combining data streams from various platforms. We used outputs of a physical–biogeochemical global ocean model at sites of real-world observations to reconstruct surface ocean pCO₂ by applying a non-linear feed-forward neural network. The results provide important information for future BGC-Argo deployment, i.e. important regions and the number of floats.