Cutaneous squamous cell carcinoma (cSCC) is the most common skin cancer with metastatic potential and development of metastases carries a poor prognosis. To address the need for reliable risk stratification, we developed cSCCNet, a deep learning model using digital pathology of primary cSCC to predict metastatic risk. A retrospective cohort of 227 primary cSCC from four centres is used for model development. cSCCNet automatically selects the tumour area in standard histopathological slides and then stratifies primary cSCC into high- vs. low-risk categories, with heatmaps indicating most predictive tiles contributing to explainability. On a 20% hold-out testing cohort, cSCCNet achieves an area under the curve (AUC) of 0.95 and 95% accuracy in predicting risk of metastasis, outperforming gene expression-based tools and clinicopathologic classifications. Multivariate analysis including common clinicopathologic classifications confirms cSCCNet as an independent predictor for metastasis, implying it identifies predictive features beyond known clinicopathologic risk factors. Histopathological analysis including multiplex immunohistochemistry suggests that tumour differentiation, acantholysis, desmoplasia, and the spatial localisation of lymphocytes relative to tumour tissue may be important in predicting risk of developing metastasis. Although further validation including prospective evaluation is required, cSCCNet has potential as a reliable and accurate tool for metastatic risk prediction that could be easily integrated into existing histopathology workflows.