Edge Computer Vision - AWS Panorama (Discontinued) and Alternative Setups for On-Site AI Analysis
AWS Panorama reached end of support on May 31, 2026. This article explains alternative setups for analyzing existing IP cameras with AI on site (SageMaker + NVIDIA Jetson-based devices + IoT Greengrass) and how the service worked when it was available.
End of Support and Migration Paths
AWS Panorama reached End of Support on May 31, 2026. Acceptance of new customers ended on May 20, 2025. Because the Panorama Appliance depended on a connection to the cloud service, the device itself also does not work after the service ends. AWS's official guidance offers two alternatives: using ready-made solutions from AWS Partners, or building models with Amazon SageMaker (including the computer vision algorithms in JumpStart) and running them on commercially available NVIDIA Jetson-based devices or edge servers. In the latter setup, edge management is handled with AWS IoT Greengrass or similar tools. This article is kept as a historical record of how the service worked when it was available + a migration guide.
Why Edge Computer Vision Matters
Setups that send IP camera footage to the cloud for ML analysis have challenges in terms of bandwidth cost, latency, and privacy. Each HD camera needs 2 to 5 Mbps of bandwidth, and sending footage from multiple cameras to the cloud makes network costs enormous. In use cases that require real-time performance (safety monitoring, quality inspection), the round-trip latency to the cloud is unacceptable. In addition, because video data contains personal privacy information, sending it to the cloud can itself become a risk. AWS Panorama was a service that deployed ML models to an edge device (Panorama Appliance) and analyzed IP camera footage locally in real time. It was designed so that video data was processed at the edge and only the analysis results (metadata) were sent to the cloud.
Panorama Appliance and Model Deployment
The Panorama Appliance was an edge device equipped with an NVIDIA GPU, and it could analyze up to 8 IP camera streams at the same time. It connected to IP cameras over the RTSP protocol, retrieved video frames, and ran inference with ML models, and for the ML models you could use custom models built with SageMaker or pretrained models (object detection, person detection, and others). Models were packaged as Docker containers and deployed to the edge device from the Panorama console. Application logic (Python) processed the models' inference results and executed actions according to conditions (sending alerts, recording metrics). Because models could also be updated remotely from the cloud, there was no need to travel to the site.
Use Cases and Pricing
The main use cases of Panorama included quality inspection on manufacturing lines (at the time, Lookout for Vision models were deployed to the edge to detect visual defects in real time. That service also ended in October 2025) and visitor analysis in retail stores (counting visitors, analyzing foot traffic, and detecting shelf inventory status). They also included safety monitoring at construction sites (checking whether helmets / safety vests are worn and detecting entry into restricted areas) and detecting parking availability (calculating the number of open spaces in real time from camera footage). The hardware price of the Panorama Appliance was about $4,000 per unit as a one-time purchase, and the monthly service fee was charged per camera stream processed (details in the pricing section below). The communication cost of sending metadata to the cloud was minimal, and no transfer cost for video data was incurred. The only third-party hardware officially certified as a Panorama-compatible device (Compatible device) was one model, the Lenovo ThinkEdge SE70 (equipped with NVIDIA Jetson Xavier NX), and the options other than the Panorama Appliance were limited to this model. Because it was not possible to choose among models to fit procurement or installation conditions, the fact that the hardware supplier was effectively fixed was a constraint that had to be considered at adoption.
Panorama Pricing and Operations
With the end of the service, the following pricing no longer exists (a record from when it was available). The Panorama Appliance was a one-time purchase of about $4,000, and the monthly fee was charged not per device but at $8.33 per month for each active camera stream processed on the device. In addition, storing the application assets deployed to the device (ML models and business logic) in the cloud cost $0.10 per GB per month. Because inference ran at the edge, no cost (bandwidth) for transferring video to the cloud was needed, and in use cases that required real-time performance the total cost was favorable compared with cloud inference. Model updates could be delivered over OTA (Over-the-Air), and you could improve inference logic remotely without traveling to the site. CloudWatch metrics were used to monitor device operating status and inference performance.
Model Development and Deployment Workflow
Image analysis at the edge proceeds by combining development in the cloud with execution on site. First, on the cloud side with its abundant compute resources, you train / evaluate an image recognition model. The flow is to deliver the finished model to devices installed on site and have them run it. Because one device can process multiple camera feeds at the same time, you can monitor multiple points on site all at once. After updating the model, you deliver it to the on-site devices again to apply it. By combining the flexibility of development in the cloud with the immediacy of execution on site, you can achieve both ease of development and low-latency analysis.
Privacy and the Benefits of On-Site Processing
A major benefit of analyzing camera footage on site is that you do not have to send the footage itself to the cloud. Footage often contains people's faces and information that can identify individuals, and sending it outside raises privacy concerns. With a setup that processes at the edge and sends only the necessary analysis results to the cloud, you can make use of sensitive footage while keeping it on site. At the same time, because large volumes of video data are not sent, you can also keep traffic and costs down. Being able to achieve both privacy protection and bandwidth savings is the essential value of performing video analysis at the edge.
Summary - Migration Guidelines
AWS Panorama was a service that enabled edge AI analysis of existing IP cameras with a minimal setup, but it reached end of support on May 31, 2026, and the Appliance also stopped working. To build an equivalent setup today, you run models built with SageMaker on NVIDIA Jetson-based devices or edge servers and distribute / manage them with IoT Greengrass or similar tools. The design of running inference at the edge and not letting footage leave the site remains effective at sites with privacy and bandwidth requirements.
References (Official AWS Resources)
The primary sources for this page are the official AWS website and documentation. Check the official pages below for the latest specifications and pricing.
- AWS Panorama official page
- Amazon SageMaker official page
- AWS IoT Greengrass official page
- AWS Documentation (official)
If this page and the official documentation disagree, treat the official documentation as authoritative.