Owing to the high concentrations of Al2O3 from dissolved fuel cladding and Na2O from the neutralization of waste, about half of the tank waste (by volume), managed as high-level waste (HLW), at Hanford, Washington, USA, is rich in Na and Al. Formulations of high waste-loading glasses result in lower amounts of SiO2, which often leads to undesirable precipitation of nepheline (NaAlSiO4), thus, severely impacting the chemical durability of the final waste form. The last two decades have witnessed the development of several empirical models, including the nepheline discriminator, optical basicity model, sub-mixture model, and the most recent machine learning-based models, to predict the crystallization of nepheline in nuclear waste glasses. While most of these models perform well within the compositional domain encompassed by their database, all the models have three main drawbacks: (1) they are conservative in terms of waste loading; (2) there are always some glass compositions that do not follow the prediction of these models; and (3) owing to their empirical nature, these models do not help in understanding the underlying science controlling the crystallization in nuclear waste glasses. The work discussed in the presentation is a decade-long effort dedicated to unearthing the underlying compositional and structural drivers controlling the crystallization of nepheline-based aluminosilicate phases in sodium and alumina-rich borosilicate glasses. It has been shown that the formation of Si–O–Al linkages in the glass structure is crucial for the crystallization of nepheline. If the Si–O–Al linkages can be broken or replaced by other linkages, for example, Si–O–B or Al–O–P, the crystallization in the glasses can be suppressed. The technological implications of the results beyond nuclear waste vitrification will also be discussed.