Background: The impact of alcohol use on dementia risk remains contentious. Methodological limitations of previous studies have made distinguishing causation from con- founding difficult. Mendelian randomization is a quasi- experimental method that can estimate causal effects via genetics. Methods: We estimated dose-response relationships be- tween alcohol and dementia in observational and genetic analyses. First, we pooled and harmonized individual-level phenotype data from two prospective cohort studies: the US Million Veteran Program (MVP) and UK Biobank (UKB; mean follow-up 4 and 12 years respectively). Associations were examined using Cox regression analysis and results combined using random-effects meta-analysis. We compared these findings with genetic associations between alcohol use and dementia (calculated de novo) using data from 2.4 mil- lion participants using Mendelian randomization. Results: 559,559 participants (247,136 from MVP and 312,423 from UKB) aged 56-72 years old at baseline were included in observational analyses. 14,540 developed dementia and 48,034 died during follow-up. Observational as- sociations between alcohol and dementia were U-shaped, meaning that non-drinkers, heavy (> 40 drinks per week) drinkers (hazard ratio = 1.41 [1.15 to 1.74]) and dependent drinkers (1.51 [1.42-1.60]) were all at higher dementia risk compared with light drinkers. In contrast, genetic analyses revealed a monotonically increasing association between alcohol dose and dementia, with no evidence supporting a protective effect of any level of drinking. A two-fold in- crease in genetically-predicted alcohol use disorder prevalence was associated with a 16 % increase in dementia cases (IVW OR = 1.16 [1.03-1.30]), and a one standard deviation in- crease in log-transformed drinks per week was associated with a 15% increase (IVW OR = 1.15[1.03-1.27]). Discussion: Alcohol consumption has a causal role for dementia. Halving the population prevalence of alcohol use disorder would lead to a 16% reduction in dementia cases.