Over the past few million years, the Earth descended from the relatively warm and stable climate of the Pliocene into the increasingly dramatic ice age cycles of the Pleistocene. The influences of orbital forcing and atmospheric CO 2 on land-based ice sheets have long been considered as the key drivers of the ice ages, but less attention has been paid to their direct influences on the circulation of the deep ocean. Here we provide a broad view on the influences of CO 2 , orbital forcing and ice sheet size according to a comprehensive Earth system model, by integrating the model to equilibrium under 40 different combinations of the three external forcings. We find that the volume contribution of Antarctic (AABW) vs. North Atlantic (NADW) waters to the deep ocean varies widely among the simulations, and can be predicted from the difference between the surface densities at AABW and NADW deep water formation sites. Minima of both the AABW-NADW density difference and the AABW volume occur near interglacial CO 2 (270–400 ppm). At low CO 2 , abundant formation and northward export of sea ice in the Southern Ocean contributes to very salty and dense Antarctic waters that dominate the global deep ocean. Furthermore, when the Earth is cold, low obliquity (i.e. a reduced tilt of Earth’s rotational axis) enhances the Antarctic water volume by expanding sea ice further. At high CO 2 , AABW dominance is favoured due to relatively warm subpolar North Atlantic waters, with more dependence on precession. Meanwhile, a large Laurentide ice sheet steers atmospheric circulation as to strengthen the Atlantic Meridional Overturning Circulation, but cools the Southern Ocean remotely, enhancing Antarctic sea ice export and leading to very salty and expanded AABW. Together, these results suggest that a ‘sweet spot’ of low CO 2 , low obliquity and relatively small ice sheets would have poised the AMOC for interruption, promoting Dansgaard–Oeschger-type abrupt change. The deep ocean temperature and salinity simulated under the most representative ‘glacial’ state agree very well with reconstructions from the Last Glacial Maximum (LGM), which lends confidence in the ability of the model to estimate large-scale changes in water-mass geometry. The model also simulates a circulation-driven increase of preformed radiocarbon reservoir age, which could explain most of the reconstructed LGM-preindustrial ocean radiocarbon change. However, the radiocarbon content of the simulated glacial ocean is still higher than reconstructed for the LGM, and the model does not reproduce reconstructed LGM deep ocean oxygen depletions. These ventilation-related disagreements probably reflect unresolved physical aspects of ventilation and ecosystem processes, but also raise the possibility that the LGM ocean circulation was not in equilibrium. Finally, the simulations display an increased sensitivity of both surface air temperature and AABW volume to orbital forcing under low CO 2 . We suggest that this enhanced orbital sensitivity contributed to the development of the ice age cycles by amplifying the responses of climate and the carbon cycle to orbital forcing, following a gradual downward trend of CO 2 .