Scientific output in resource-intensive fields is concentrated in a small number of elite universities. We study whether an open-access artificial intelligence tool reduces this concentration, using the public releases of AlphaFold, the deep-learning system that predicts protein structures. In the two years after AlphaFold was released openly, the share of the average low-ranked university in top-journal protein research rose by 50 percent relative to that of the average top-ranked university. The gains are absent outside protein research and in lower-tier journals. Papers from lower-ranked universities are also more likely to redirect subsequent research. Lower-ranked universities also co-author less with top-ranked universities, and the decline grows with their share of in-house structural biologists. The evidence suggests that open-access artificial intelligence lowers the human and financial resources required to produce frontier science.