Assessing soil organic carbon (SOC) stocks is a methodological issue for SOC monitoring at regional scale but crucial for global agendas of SOC sequestration to mitigate climate change and reduce food insecurity. The ‘4 per mille Initiative: Soils for Food Security and Climate’, highlighted agricultural soil as a major lever for climate action and the need to assess SOC stock at different spatial and temporal scale. Infrared spectroscopy appeared as a promising tool to address this methodological issue. This work aimed to evaluate the potential of visible-near-infrared (VNIR) and mid-infrared (MIR) spectroscopic measurement methods to predict SOC stock and its variables (SOC content and bulk density) in tropical volcanic soils of ‘La Réunion’ island. The diversity of agricultural soils of ‘La Réunion’ was captured in the sample set (n = 95) with soil orders such as Andosols, Cambisols and Ferralsols. Partial least squares regressions (PLSR) with leave-one-out cross validation were used to build prediction models. With RPD higher than 2, the present study showed good prediction accuracy of models by MIR and VNIR spectroscopy of SOC content, bulk density and SOC stock for measurements in the laboratory or in the field. Accurate and direct SOC stock predictions were achieved on dried and sieved soil samples with MIR spectroscopy (RPD = 2.25; R2cv = 0.80; RMSEcv = 0.69 KgC m−2) and VNIR spectroscopy (RPD = 2.74; R2cv = 0.87; RMSEcv = 0.61 KgC m−2) but also directly on cores in the field with VNIR spectroscopy (RPD = 3.29; R2cv = 0.91; RMSEcv = 0.51 KgC m−2). This unexpected ability to predict directly SOC stocks by infrared spectroscopy can be partly explained by the high SOC content coupled with the large variation of SOC content and bulk density, providing a large range for those variables, and then a higher predictability. Yet these results questioned the underlying drivers of the bulk density and SOC stock, both being largely physical parameters supposed to be hardly predictable by infrared spectroscopy. Analyses of mean spectra and regression coefficients, combined with amorphous product (Alo, Feo, Sio, Alp, Fep) prediction models, demonstrated that these were detected in the spectra and were the drivers of SOC stock accurate predictions by infrared spectroscopy. Short-range ordered minerals, especially allophanes, appeared as proxy of the bulk density, and amorphous products such as Alp and Fep indicated the presence of organo-mineral complexes involved in SOC storage. Such promising results for SOC stock predictions from near- and mid-infrared volcanic soil spectra, confirmed that VNIR and MIR diffuse reflectance spectroscopy are an appropriate tool, rapid, low cost and non-destructive, to study SOC stocks in tropical volcanic soils. Upscaling of SOC stocks across the agricultural soils of the island is now just a step ahead, following external validation with additional data to validate the robustness of the prediction models.