There is need for accurate projections of the future dementia burden to prepare healthcare systems and policymakers. Existing projections only account for population ageing, not for observed declines in age-specific dementia incidence of 13% per decade. We developed a dementia microsimulation model that synthesizes population-based data from the Rotterdam Study with changes in demographics between birth cohorts from the early 1900s onwards. We determined dementia prevalence and incidence until 2050 for three different dementia incidence trend scenarios: (1) stable age-specific incidence, (2) linear decline by 13% per decade, (3) nonlinear declines averaging 13% per decade. Assuming a stable age-specific incidence resulted in a 130% increase in incidence and 118% in prevalence between 2020 and 2050. By contrast, the linearly declining trend resulted in substantially smaller increases of 58% in incidence (95%CI: 29-87%), and 43% in prevalence (95%CI: 13-66%), corresponding to 39% lower incidence and 36% lower prevalence by 2050 than in the stable-incidence scenario. Results for various non-linear declines fell between the stable and linear trend. The future burden of dementia is highly susceptible to achievable changes in age-specific incidence. Extension of previously established secular trends globally would reduce widely upheld projections of new dementia cases until 2050 by 39%.