AWS IoT Greengrass is AWS's open-source edge runtime and cloud service for deploying and managing software, including machine learning models, on edge devices. In 2026 edge AI on AWS means Greengrass V2 (or the lightweight nucleus lite), models compiled with SageMaker Neo or exported to ONNX, and your own monitoring, because Greengrass V1 support ended on 7 October 2026 and SageMaker Edge Manager was discontinued in 2024.
Updated October 2026
Who this is for: IoT platform owners, edge and OT engineering leads, and CTOs in manufacturing, energy, retail and logistics running computer vision or sensor models on AWS-connected edge devices.
The business problem: an edge AI stack that is aging out
Many edge AI projects built between 2019 and 2022 followed the same AWS recipe: train in SageMaker, compile with Neo, manage with SageMaker Edge Manager and deploy with Greengrass V1. Several of those pieces have now reached end of life. If your devices still depend on them, you face security exposure, failed deployments and models you can no longer monitor, on equipment that runs your production lines and stores. Migration is no longer optional; the question is how to do it without downtime.
What changed in 2026
- Greengrass V1 support ended. AWS's maintenance policy states that on 7 October 2026 AWS discontinues support for AWS IoT Greengrass Version 1, after which V1 resources are no longer accessible. V1 had been in its extended life phase, with no security patches, since 30 June 2023 (AWS docs).
- Greengrass nucleus lite, introduced with release 2.14.0 in December 2024, is a C-based runtime that needs less than 5 MB of RAM and less than 5 MB of storage, bringing V2 to low-cost devices (AWS IoT blog).
- SageMaker Edge Manager was discontinued on 26 April 2024. AWS recommends ONNX for a cross-platform runtime and Greengrass V2 for deployment and monitoring (AWS docs).
- SageMaker Model Monitor, one suggested telemetry destination, moved to maintenance and is closed to new customers from 30 July 2026 (AWS). New projects need another monitoring path.
- AWS also announced end of support for AWS Panorama, AWS IoT Analytics and AWS IoT Events in 2025 (AWS).
What still works: the 2026 AWS edge AI toolkit
| Component | Status in 2026 | Role | Action |
|---|---|---|---|
| Greengrass V1 | End of support 7 Oct 2026 | Legacy runtime with Lambda functions | Migrate now |
| Greengrass V2 (Java nucleus) | Active | Components, deployments, IPC, MQTT | Default for gateways and edge servers |
| Greengrass nucleus lite | Active (since Dec 2024) | Same V2 cloud API on constrained devices | Use for small Linux devices |
| SageMaker Neo | Documented and available | Compiles models for ARM, Intel, NVIDIA, NXP, Qualcomm and other targets | Use where your target is supported |
| ONNX Runtime | AWS-recommended | Portable inference runtime | Default for portability |
| SageMaker Edge Manager | Discontinued Apr 2024 | Model fleet management | Replace |
According to the SageMaker Neo documentation, Neo supports frameworks including PyTorch, TensorFlow, TensorFlow Lite and ONNX, and compiled models can be deployed to an AWS IoT Greengrass device.
Reference architecture: edge AI with Greengrass V2 in 2026
SageMaker or any ML platform
Neo compile or ONNX export, quantise
Greengrass V2 component (model + inference app)
Fleet deployment; local inference on camera or sensor data
MQTT telemetry, sampled data to S3, drift checks
Figure: Edge AI on AWS in 2026. Models are trained, compiled or exported to ONNX, packaged as Greengrass V2 components, deployed to device fleets for local inference, and monitored through MQTT telemetry and sampled data that feed retraining.
The monitoring step is where most teams now need new tooling. Our guides on MLOps for edge AI and edge analytics cover drift detection and telemetry patterns.
Migration plan: Greengrass V1 to V2 in 30 days
- Week 1 – inventory. List every V1 core, Lambda function, connector, ML resource and Edge Manager dependency. Rank devices by business criticality.
- Week 2 – rebuild as components. Convert Lambda functions to V2 components (AWS provides a V1-to-V2 migration guide), export models to ONNX or recompile with Neo, and replace Edge Manager agents.
- Week 3 – pilot. Deploy to a test group of devices; validate inference accuracy, latency and resource use against the old stack.
- Week 4 – staged rollout. Roll out by site with rollback plans, and switch monitoring to your new telemetry path.
Still running Greengrass V1 or Edge Manager? Get a free edge AI migration assessment. We inventory your devices and models and give you a sequenced plan to V2 and a supported monitoring stack. Request the assessment.
Objection: "Should we stay all-in on AWS at the edge?"
Greengrass V2 is a solid runtime, but the past three years show that edge services change. Keep models in portable formats (ONNX), keep inference apps containerised where possible, and keep monitoring and governance in a layer you control. That way a service retirement becomes a configuration change, not a rebuild. See edge computing platforms for alternatives, and AWS Inferentia for the cloud side of AWS inference.
Where it's heading: 2026 → 2030
| Period | Edge AI on AWS |
|---|---|
| ~2020 | Greengrass V1, Neo and Edge Manager; Lambda at the edge (history). |
| 2026 | Greengrass V2 and nucleus lite, ONNX, Neo; V1 and Edge Manager retired; teams own monitoring. |
| 2028 | Expected: small language and vision language models running on edge devices alongside classic vision models. |
| 2030 | Our view: edge fleets run agents, not just models, managed through portable, multi-cloud control planes. |
How NexaStack helps
NexaStack is built for running AI at the data source. Its Edge AI capabilities focus on deploying, managing and updating models across edge nodes, keeping data inside enterprise boundaries, and continuing to operate with limited or no cloud connectivity. Its observability and evaluation layer gives you the monitoring that Edge Manager used to provide, and its composable agent framework lets edge devices move from detection to action. It can sit alongside AWS IoT services or run on private infrastructure.
FAQ
What is AWS IoT Greengrass?
It is an open-source edge runtime and cloud service from AWS for building, deploying and managing software, including ML inference, on edge devices.
What is the difference between AWS IoT Greengrass and IoT Core?
IoT Core is the cloud service that connects devices and routes messages. Greengrass runs on the device or gateway to execute code and models locally and talks to IoT Core.
Is SageMaker Neo still available in 2026?
Yes. Neo remains documented for compiling models for cloud instances and edge devices, including deployment to Greengrass devices. SageMaker Edge Manager, which was often paired with it, was discontinued in 2024.
What are the latest edge AI trends on AWS in 2026?
Migration from Greengrass V1, nucleus lite for small devices, ONNX portability, on-device small language models and independent model monitoring.
What happens to Greengrass V1 devices now?
According to AWS, after 7 October 2026 V1 resources are no longer accessible, so V1 devices can no longer rely on AWS cloud services. Plan an immediate migration to V2.
Keep your edge AI running, and future-proof it. Start a 30-day edge AI migration pilot with NexaStack on one device group. Book your pilot.
