We introduce TimePoint (TP), a self-supervised model for fast and robust time series alignment, trained entirely on synthetic data. Unlike prior foundation models that focus on forecasting, TP targets the critical but underexplored task of alignment by learning generalizable temporal landmarks (keypoints and descriptors).To support this, we propose SynthAlign, a synthetic data generation pipeline using structured waveform composition and CPAB-based time warping with ground-truth correspondences. Despite no access to real data during training, TP achieves zero-shot generalization on 100+ real-world datasets, outperforming DTW variants in both alignment accuracy and efficiency.We position TP as an alignment-centric foundation model, showing that high-quality synthetic data and task-specific pretraining can unlock scalable and transferable representations for time series analysis beyond forecasting. Disclaimer: This work is based on an accepted paper from the conference main track.