We want to run AI inference right on the production line (edge AI): should we choose an NVIDIA Jetson-based machine or an x86 machine with a discrete GPU?

The two approaches solve different problems, so decide by the number of video streams and the software you already have. Jetson-based machines: compact, drawing roughly 15–60 W depending on the module generation, able to run fanless, and a good fit for 1–4 cameras with models optimised in TensorRT. The operating system is Linux (JetPack) on ARM, so software written for Windows or libraries only built for x86 will not run as-is. x86 machines with a discrete GPU: run both Windows and Linux, make it easy to reuse existing machine-vision software, and handle many streams or large models. The trade-off is a PCIe x16 slot, a power supply with enough capacity (common GPUs draw 70–250 W) and usually a fan, so placement and dust must be planned. Quick method: count the cameras, resolution and inference frames per second you need, then test the model on a sample machine before committing to quantities.