Differentiable ray tracing has recently challenged the status quo in radio propagation modelling and digital twinning. Promising unprecedented speed and the ability to learn from real-world data, it offers a real alternative to conventional deep learning (DL) models. However, no experimental evaluation on production-grade networks has yet validated its assumed scalability or practical benefits. This leaves mobile network operators (MNOs) and the research community without clear guidance on its applicability. In this paper, we fill this gap by employing both differentiable ray tracing and DL models to emulate radio coverage using extensive real-world data collected from the network of a major MNO, covering 13 cities and more than 10,000 antennas. Our results show that, while differentiable ray-tracing simulators have contributed to reducing the efficiency-accuracy gap, they struggle to generalize from real-world data at a large scale, and they remain unsuitable for real-time applications. In contrast, DL models demonstrate higher accuracy and faster adaptation than differentiable ray-tracing simulators across urban, suburban, and rural deployments, achieving accuracy gains of up to 3 dB. Our experimental results aim to provide timely insights into a fundamental open question with direct implications on the wireless ecosystem and future research.