The high risk population of cardiovascular disease (CVD) is simultaneously at high risk of lung cancer. Given the dominance of low dose computed tomography (LDCT) for lung cancer screening, the feasibility of extracting information on CVD from the same LDCT scan would add major value to patients at no additional radiation dose. However, with strong noise in LDCT images and without electrocardiogram (ECG) gating, CVD risk analysis from LDCT is highly challenging. Here we present an innovative deep learning model to address this challenge. Our deep model was trained with 30,286 LDCT volumes and achieved the state-of-the-art performance (area under the curve (AUC) of 0.869) on 2,085 National Lung Cancer Screening Trial (NLST) subjects, and effectively identified patients with high CVD mortality risks (AUC of 0.768). Our deep model was further calibrated against the clinical gold standard CVD risk scores from ECG-gated dedicated cardiac CT, including coronary artery calcification (CAC) score, CAD-RADS score and MESA 10-year CHD risk score from an independent dataset of 106 subjects. In this validation study, our model achieved AUC of 0.942, 0.809 and 0.817 for CAC, CAD-RADS and MESA scores, respectively. Our deep learning model has the potential to convert LDCT for lung cancer screening into dual-screening quantitative tool for CVD risk estimation.