Most phosphorus (P) in soils is unavailable for direct biological uptake, as it is locked within primary or secondary mineral particles, adsorbed to mineral surfaces, or immobilized inside of organic material. Deciphering the composition of different P forms in soil is critical for understanding P bioavailability and its underlying dynamics. However, widely used global estimates of different soil P forms are based on a dataset containing few measurements in which many regions or soil types are unrepresented. This poses a major source of uncertainty in assessments that rely on these estimates to quantify soil P constraints on biological activity controlling global food production and terrestrial carbon balance. To address this issue, we consolidated a database of six major soil P “forms” containing 1857 entries from globally distributed (semi-)natural soils and 11 related environmental variables. These six different forms of P (labile inorganic P (Pi), labile organic P (Po), moderately labile Pi, moderately labile Po, primary mineral P, and occluded P) were measured using a sequential P fractionation method. As they do not represent precise forms of specific discrete P compounds in the soil but rather resemble operational pools, we will now refer to them as P pools. In order to quantify the relative importance of 11 soil-forming variables in predicting soil P pool concentrations and then make further predictions at the global scale, we trained random forest regression models for each of the P pools and captured observed variation with R2 higher than 60 %. We identified total soil P concentration as the most important predictor of all soil P pool concentrations, except for primary mineral P concentration, which is primarily controlled by soil pH and only secondarily by total soil P concentration. When expressed in relative values (proportion of total P), the model showed that soil pH is generally the most important predictor for proportions of all soil P pools, alongside the prominent influences of soil organic carbon, total P concentration, soil depth, and biome. These results suggest that, while concentration values of P pools logically strongly depend on soil total P concentration, the relative values of the different pools are modulated by other soil properties and the environmental context. Using the trained random forest models, we predicted soil P pools' distributions in natural systems at a resolution of 0.5∘×0.5∘. Our global maps of different P pools in soils as well as the pools' underlying drivers can inform assessments of the role of natural P availability for ecosystem productivity, climate change mitigation, and the functioning of the Earth system.