Semi-supervised Learning (SSL) has received increasing attention in autonomous driving to relieve enormous burden for 3D annotation. In this paper, we propose UpCycling, a novel SSL framework for 3D object detection with zero additional raw-level point cloud: learning from unlabeled de-identified intermediate features (i.e., smashed data) for privacy preservation. The intermediate features do not require additional computation on autonomous vehicles since they are naturally produced by the inference pipeline. However, augmenting 3D scenes at a feature level turns out to be a critical issue: applying the augmentation methods in the latest semi-supervised 3D object detectors distorts intermediate features, which causes the pseudo-labels to suffer from significant noise. To solve the distortion problem while achieving highly effective SSL, we introduce hybrid pseudo labels, feature-level Ground Truth sampling (F-GT) and Rotation (F-RoT), which safely augment unlabeled multi-type 3D scene features and provide high-quality supervision. We implement UpCycling on two representative 3D object detection models, SECOND-IoU and PV-RCNN, and perform experiments on widely-used datasets (Waymo, KITTI, and Lyft). While preserving privacy with zero raw-point scene, UpCycling significantly outperforms the state-of-the-art SSL methods that utilize raw-point scenes, in both domain adaptation and partial-label scenarios.