A manufacturing company with remote facilities and limited internet has a large labeled image dataset on-premises for defect detection. They need a solution that lowers compute costs, scales training efficiently, and lets the trained model run in low-connectivity facilities for real-time conveyor-belt inference. Which solution meets these requirements?

The correct solution addresses all requirements. Moving training data to Amazon S3 and using SageMaker for training leverages cloud scalability and cost efficiency. SageMaker Neo optimizes the model for edge devices, reducing its size and improving inference speed, which is crucial for real-time performance on conveyor belts. Deploying with AWS IoT Greengrass allows the optimized model to run locally on edge devices in low-connectivity environments, enabling real-time inference without constant internet access. The first incorrect option fails to address the low-connectivity requirement for inference, as a SageMaker hosting endpoint requires continuous internet access. The second and fourth incorrect options involve on-premises training, which doesn't leverage cloud scalability or cost optimization for training, and the fourth also misses the crucial optimization step with SageMaker Neo for efficient edge deployment.