A pure liquid can reach metastable equilibrium with its subsaturated vapor across an appropriate membrane. This situation is analogous to osmotic equilibrium: the reduced chemical potential of the dilute phase (the subsaturated vapor) is compensated by a difference in pressure between the phases. To equilibrate with subsaturated vapor, the liquid phase assumes a pressure that is lower than its standard vapor pressure, such that the liquid phase is metastable with respect to the vapor phase. For sufficiently subsaturated vapors, the liquid phase can even assume negative pressures. The appropriate membrane for this metastable equilibrium must provide the necessary mechanical support to sustain the difference in pressure between the two phases, limit nonhomogeneous mechanisms of cavitation, and resist the entry of the dilutant (gases) into the pure phase (liquid). In this article, we present a study of the limit of stability of liquid water--the degree of subsaturation at which the liquid cavitates--in this metastable state within microscale voids embedded in hydrogel membranes. We refer to these structures as vapor-coupled voids (VCVs). In these VCVs, we observed that liquid water cavitated when placed in equilibrium with vapors of activity aw,vapair<or=0.85 (relative humidity<or=85%). When expressed as a pressure in the liquid based on thermodynamic calculations, the liquid cavitated at pressures P<or=-22 MPa (-220 bar). This limiting pressure is smaller in magnitude than the limit predicted by homogeneous nucleation theory or molecular simulations (Pcav=-140 to -180 MPa). To determine the cause of the disparity between the observed and predicted stability limit, we examine experimentally the likelihood of several nonhomogeneous mechanisms of nucleation: (i) heterogeneous nucleation caused by hydrophobic patches on void walls, (ii) nucleation caused by the presence of dissolved solute, (iii) nucleation caused by the presence of pre-existing vapor nuclei, and (iv) invasion of air through the hydrogel membrane into the voids. We conclude that, of these possibilities, (i) and (ii) cannot be discounted, whereas (iii) and (iv) are unlikely to play a role in determining the stability limit.