Evolutionary theory hypothesizes that intermediate virulence maximizes pathogen
fitness as a result of a trade-off between virulence and transmission, but empirical evidence
remains scarce. We bridge this gap using data from a large and long-standing HIV-1 prospective
cohort, in Uganda. We use an epidemiological-evolutionary model parameterised with this data to
derive evolutionary predictions based on analysis and detailed individual-based simulations. We
robustly predict stabilising selection towards a low level of virulence, and rapid attenuation of the
virus. Accordingly, set-point viral load, the most common measure of virulence, has declined in the
last 20 years. Our model also predicts that subtype A is slowly outcompeting subtype D, with both
subtypes becoming less virulent, as observed in the data. Reduction of set-point viral loads should
have resulted in a 20% reduction in incidence, and a three years extension of untreated
asymptomatic infection, increasing opportunities for timely treatment of infected individuals.