Maxime Keutgen De Greef, lead author of this preprint , is part of InMOS, another Schmidt Sciences project. This work presents SamudraBGC, a new machine-learning emulator of ocean biogeochemistry that captures key regional and seasonal patterns in a high-resolution North Atlantic configuration, including contrasts between the oligotrophic subtropical gyre, productive subpolar gyre, and energetic jet. The model also reproduces seasonal-to-interannual variability and the spread and growth rates of a 50-member dynamical ensemble, at a tiny fraction of the computational cost: ~0.1 GPU-hours per model year compared with ~3,600 CPU-hours for the dynamical model. These results pave the way toward ensemble-based data assimilation for ocean biogeochemistry.