A factory had a pilot that ran image inference using AWS IoT Greengrass on an industrial PC, uploading images to S3 and invoking a SageMaker endpoint. When scaling to thousands of machines, internet bandwidth became saturated and latency is unacceptable. What is the most cost-effective way to reduce latency and resolve the network capacity issue?

The most cost-effective solution is to move the inference logic to the edge. Deploying the inference Lambda and model directly on the AWS IoT Greengrass core on each industrial PC eliminates the need to upload large image files to S3 and invoke a remote SageMaker endpoint for every inference. This drastically reduces internet bandwidth usage and network latency, as inference happens locally. Results can then be forwarded directly to an on-site web service, bypassing the cloud for immediate actions. Incorrect options: Provisioning a 10 Gbps Direct Connect link and increasing SageMaker instance sizes is expensive and doesn't fully address latency if inference still requires cloud round trips for every image. Compressing images helps with bandwidth but still requires uploading and downloading, introducing latency and additional processing steps. Enabling SageMaker autoscaling and a Direct Connect link still relies on cloud-based inference, which is not ideal for low-latency, high-volume edge processing and doesn't fully resolve network saturation from image uploads.