Open-source software development has facilitated the “reuse” of many digital applications across diverse operational contexts. However, a similar level of reuse may be more difficult to achieve for artificial intelligence (AI). Unlike traditional software, AI often requires more complex algorithmic architectures, significant and costly processing capacity, and extensive training data. Yet, one of the greater promises of AI is its potential to be reused in related contexts, or even dissimilar ones. To investigate how open-source code reuse is amenable to high-performance AI applications, we explore AlphaFold, a system developed by DeepMind (Alphabet), which has been considered one of the most important breakthroughs in AI. AlphaFold has enabled solving the structures of more than 200 million proteins, promising to transform and accelerate basic biological, medicinal, and pharmaceutical research. In an unexpected move, AlphaFold’s code and model weights were open-sourced by DeepMind to the research community. Although open-sourcing AlphaFold was intended to facilitate its reuse, it also raised new challenges: barriers to modification, limited repurposability, and inscrutability. Our analysis and conceptual development offer a new understanding of open-source AI reuse as managing the scales across its constituent physical, logical, and data layers. Our insights on AI reuse contribute to researchers concerned about AI’s increased opacity, restrictive sharing by companies, regulatory risks, and equitable application.