There is a long-established, remarkable correspondence between the nitrogen-to-phosphorus ratio N:P~15 of deep ocean water and the ''Redfield ratio'' of N:P~16 required by phytoplankton. Redfield and subsequent workers have suggested that it is due to N-fixing organisms being selected when N:P<16 but being out-competed when N:P>16. Models have shown this mechanism can work, but recent observations reveal that the real system is more complex. First, the C:N:P stoichiometry of phytoplankton varies with growth rate, nutrient and light limitation, species and phylum. Second, although N-fixation is sometimes P-limited and suppressed by N-addition, there is also evidence for Fe-limitation, light-limitation and P and Fe co-limitation of N-fixers. Here we adapt recent models to include non-Redfieldian stoichiometry of phytoplankton and limitation of N-fixers by resources other than P. We show that the deep ocean N:P is set by the N:P threshold that triggers N-fixation, and is not directly related to the N:P ratio of sinking material. However, assuming competitive dynamics set the N:P threshold for N-fixation, it will be close to the N:P requirement of non-fixers (rather than that of N-fixers) and consequently so will the deep ocean N:P ratio. Theoretical limits on the N:P requirements of phytoplankton suggest that since the deep ocean became well oxygenated, its N:P has remained within the range 7.7–32.3. Decreases in phytoplankton C:P and N:P ratios over the past ~1 Gyr are predicted to have driven a decrease in deep ocean N:P, probably via increasing PO₄. Even if Fe or light limitation restrict N-fixers to a fraction of the surface ocean, they reach higher densities there, minimising variations in deep ocean N:P. Thus Redfield's mechanism is robust and we expand it to suggest that phytoplankton C:N:P and deep ocean N:P have co-evolved.