We study how advances in artificial intelligence reshape scientific knowledge production. We first develop a model of knowledge acquisition on a metric space in which experiments generate local spillovers, laboratories face setup costs, and AlphaFold2 (AF2) arrives as an exogenous signal that lowers baseline information costs unevenly across proteins. The model predicts reallocation rather than simple expansion: substitution away from experiments AF2 already informs, alongside increased search on frontier proteins and shifts in follow-up work. We test these predictions using a new dataset linking AF2 outputs to roughly 553,000 proteins and Protein Data Bank activity. Exploiting cross-protein variation in AF2 prediction quality after the July 2021 release, we find that AF2 substantially reduces the probability that a well-predicted protein receives any experimental attention, yet total depositions and follow-up work increase sharply---effort narrows onto fewer proteins rather than disappearing. The composition of new depositions shifts toward first-ever structure determinations for previously unresolved proteins, and within follow-up work, toward resolving previously uncharacterised regions and away from pure corroboration. The net effect is a reallocation of effort toward proteins where AI predictions and prior experimental knowledge are jointly available, driven primarily by academic rather than industry laboratories. AI thus reorganises where, how, and by whom science is produced.