The cubic blue phases of liquid crystals are fascinating and technologically promising examples of hierarchically structured soft materials, comprising ordered networks of defect lines (disclinations) within a liquid crystalline matrix. We present large-scale simulations of their domain growth, starting from a blue phase nucleus within a supercooled isotropic or cholesteric background. The nucleated phase is thermodynamically stable; one expects its slow orderly growth, creating a bulk cubic phase. Instead, we find that the strong propensity to form disclinations drives the rapid disorderly growth of a metastable amorphous defect network. During this process, the original nucleus is destroyed; reemergence of the stable phase may therefore require a second nucleation step. Our findings suggest that blue phases exhibit hierarchical behavior in their ordering dynamics, to match the hierarchy in their structure.kinetic arrest | complex fluids T he blue phases, BPI and BPII, of chiral nematic liquid crystals can each be viewed as an ordered network of topological defects (disclination lines), embedded within a liquid crystalline matrix (a cholesteric) whose local molecular alignment axis rotates helically on moving through the sample. Such phases were long viewed as purely of academic interest, due to their extremely narrow window of thermodynamic stability (of order 1 K) (1). This view has changed completely with the recent development of new compounds showing stable BPs over a 50 K interval, with fast switching between different states (2, 3). Thus BPs now offer a promising device technology not only for displays (4) but also for laser applications (5, 6). To fully realize this potential requires an understanding of how chiral nematic materials switch between different structures. However, theoretical work on BPs has advanced relatively modestly since the late 1980s (1), and so far there is almost no understanding of their phase-change kinetics. This situation is partly because of computational challenges which, despite pioneering progress using small systems (7-9), have prevented the simulation of supraunit cell behavior in the ordered BPI and BPII, and ruled out realistic simulation of a third blue phase, the apparently amorphous (1) BPIII.With the aid of supercomputers and a hybrid lattice Boltzmann algorithm (10) we have recently overcome these difficulties, enabling us to address systems hundreds of times larger than the unit cell volume. This advance has allowed us to address elsewhere the motion of planar interfaces between competing phases (11). By the same methods, we address below the domain growth of the ordered BPs from an isolated nucleus of the stable phase. We find that the growth process is unexpectedly interrupted by the proliferation of a metastable BPIII-like structure. Thus the system seemingly opts for speed rather than efficiency, lowering its free energy rapidly by amorphous defect proliferation even at the expense of the creation of barriers which prevent attainment of the global free-energy...