COVID-19 has overloaded national health services worldwide. Thus, early identification of patients at risk of poor outcomes is critical. Our objective was to analyse SARS-CoV-2 RNA detection in serum as a severity biomarker in COVID-19. Retrospective observational study including 193 patients admitted for COVID-19. Detection of SARS-CoV-2 RNA in serum (viremia) was performed with samples collected at 48-72 h of admission by two techniques from Roche and Thermo Fischer Scientific (TFS). Main outcome variables were mortality and need for ICU admission during hospitalization for COVID-19. Viremia was detected in 50-60% of patients depending on technique. The correlation of Ct in serum between both techniques was good (intraclass correlation coefficient: 0.612; p < 0.001). Patients with viremia were older (p = 0.006), had poorer baseline oxygenation (PaO2/FiO2; p < 0.001), more severe lymphopenia (p < 0.001) and higher LDH (p < 0.001), IL-6 (p = 0.021), C-reactive protein (CRP; p = 0.022) and procalcitonin (p = 0.002) serum levels. We defined "relevant viremia" when detection Ct was < 34 with Roche and < 31 for TFS. These thresholds had 95% sensitivity and 35% specificity. Relevant viremia predicted death during hospitalization (OR 9.2 [3.8-22.6] for Roche, OR 10.3 [3.6-29.3] for TFS; p < 0.001). Cox regression models, adjusted by age, sex and Charlson index, identified increased LDH serum levels and relevant viremia (HR = 9.87 [4.13-23.57] for TFS viremia and HR = 7.09 [3.3-14.82] for Roche viremia) as the best markers to predict mortality. Viremia assessment at admission is the most useful biomarker for predicting mortality in COVID-19 patients. Viremia is highly reproducible with two different techniques (TFS and Roche), has a good consistency with other severity biomarkers for COVID-19 and better predictive accuracy.