The sizing of aircraft structures is based on design allowables (e.g., strength) which contribute to the achievement of positive margins of safety, in particular with regards to the variability of material properties. Nowadays, most of the design allowables are derived from expensive and time consuming physical test campaigns, which could be replaced to some extent by numerical simulations. As design allowables consist of an estimation of a quantile with a given confidence level (e.g., B-value is estimator of the 10th percentile with 95% confidence level), we aim to estimate such statistical quantity for complex non-linear finite element. Many algorithms have been developed for quantile estimation, including the crude Monte-Carlo approach computationally intractable in practice. Sequential methods called ``active learning'' strategies have also been proposed. Such techniques progressively enrich an initial surrogate model with new simulations given an acquisition criterion. Parallelized versions of active learning strategies have been proposed for finding optima, allowing a significant reduction which could be suited for heavy non-linear finite element simulations but applications to quantile estimation are rarer. Viana et al attempted to use multiple surrogates but this approach is severely limited: the number of simulations added at each iteration is bounded by the number of surrogates that appropriately fit the model, but not by the parallel computing resource. This paper aims to implement a parallelized version of active learning dedicated to quantile estimation, compatible with an arbitrary number of simulations added at the different steps of the method. Quantiles of benchmark functions are estimated with the developed approach, which was then applied to estimate the quantile of failure force of an L-shape unfolding test. Our results show that the quantile estimated in a parallelized fashion are not only slightly more accurate but also converge with much fewer iterations compared to the standard one-by-one method in all test cases, demonstrating the interest to parallelized calculations.