Edge AI computer vision runs image and video models on hardware next to the camera, so inspections, safety checks and counts happen in milliseconds and images stay on site. The 2026 stack has five layers: cameras, edge compute, an open runtime, vision models (now including vision language models) and a central control plane that deploys, monitors and retrains models across every site.
Updated October 2026
What changed in 2026
- The first wave of all-in-one edge vision kits is gone. Microsoft retired the Azure Percept DK on 30 March 2023 (Microsoft Learn), and AWS ended support for AWS Panorama on 31 May 2026, recommending off-the-shelf hardware managed with tools such as AWS IoT Greengrass and EKS Anywhere (AWS docs).
- Edge hardware got much bigger. NVIDIA Jetson Thor offers up to 2070 FP4 TFLOPS (sparse) and 128 GB of memory, aimed at physical AI (NVIDIA).
- Vision models learned to reason. NVIDIA describes Cosmos as usable "as a vision language model (VLM) to reason over objects, interactions, and intent" (NVIDIA).
- Kubernetes became the edge standard. Azure Arc-enabled Kubernetes lets you "attach Kubernetes clusters running anywhere" and manage them centrally (Microsoft Learn).
Who this is for: heads of quality, plant engineering, retail operations and IT architecture who are scaling computer vision from one pilot line to many sites, or replacing a retired edge vision kit.
The business problem: pilots work, rollouts stall
Most companies have proved computer vision works. A defect model hits its accuracy target on one line. A shelf model counts stock in one store. Then the rollout stalls. Each site has different cameras, lighting and network rules. Models drift. Updating 200 devices by hand is not possible. And if the pilot was built on a vendor kit that has since been retired, the team has to start again.
The fix is to treat edge vision as a platform, not a project: standard layers, open runtimes and one control plane.
The 2026 edge computer vision stack
IP, industrial, thermal
embedded GPU or edge server
containers, Kubernetes, inference server
detectors, segmenters, VLMs
reject part, alert, count
Figure: The five-layer edge computer vision stack: cameras feed edge compute running an open runtime and models that act locally, while a central control plane and data platform manage rollout and retraining.
Layer by layer
- Cameras: reuse what you have where image quality allows. Lighting and lens choice matter more than model choice for inspection.
- Edge compute: embedded GPUs for one to a few cameras; edge servers for many streams or vision language models.
- Runtime: containers on Kubernetes or a managed edge runtime, so the same model runs on any site. AWS calls Greengrass "an open-source edge runtime and cloud service" (AWS).
- Models: fast detectors and segmenters for every frame; vision language models for verification, descriptions and rare cases. See vision transformers.
- Control plane and data: push models to sites, watch accuracy per site, send hard examples back for labelling and retraining. Our guide to image and video analytics on Databricks covers the data side.
Then vs now: edge vision in 2020 and 2026
| Area | Around 2020 | 2026 |
|---|---|---|
| Hardware | Vendor dev kits and appliances (for example Azure Percept, AWS Panorama) | Off-the-shelf edge GPUs and servers; kits retired |
| Runtime | Vendor-specific SDKs | Containers and Kubernetes managed centrally |
| Models | One CNN per task, trained on many labels | Detectors plus VLMs with zero-shot reasoning |
| Data loop | Manual export and relabel | Automatic hard-example capture into a data platform |
| Output | Alerts to people | Alerts plus agents that take allowed actions |
Stuck between a successful pilot and a multi-site rollout? Book a free edge vision scale-up review. We map your pilot onto a stack that can run on every site.
Migrating from a retired edge vision kit
- Export your trained models in an open format such as ONNX, or retrain from your labelled data.
- Pick off-the-shelf edge hardware that meets your throughput needs, tested on your own footage.
- Package models and pre/post-processing as containers.
- Manage devices and rollouts from one control plane rather than per-device scripts.
Related reading: AWS IoT Greengrass for edge AI and computer vision solutions.
How to start in 30 days
- Week 1: pick one use case with a known cost, such as a defect that escapes to customers. Collect a few hundred sample images from the real line.
- Week 2: fine-tune a detector or test a vision language model zero-shot; containerise it.
- Week 3: deploy to one edge device through your control plane and run in shadow mode.
- Week 4: measure detection and false-alarm rates against human inspection and capture hard examples for the next training round.
Where it's heading: 2026 → 2030
| Period | Edge computer vision |
|---|---|
| Around 2020 (history) | Vendor kits, single-task CNNs, manual updates. |
| 2026 (now) | Open runtimes, edge GPUs that run VLMs, central fleet management, kits retired. |
| 2028 (expected) | Our view: fewer task-specific models, more general vision models adapted with small amounts of data. |
| 2030 (our view) | Vision is one input to physical AI agents that perceive, reason and act across whole sites. |
How NexaStack helps
NexaStack provides the control-plane layer for edge vision. Its Vision AI offering analyses streams in real time at the edge and supports deployment "from edge devices to cloud platforms," while its Edge AI solution lets you train and fine-tune custom vision models for settings such as factory floors or retail aisles, then deploy and update them "across thousands of edge nodes with centralized orchestration." Self-hosted and air-gapped options keep images on your premises.
FAQ
What is edge AI in computer vision?
It is running vision models on hardware near the camera instead of in the cloud, so results are instant and images stay on site.
What replaced Azure Percept and AWS Panorama?
Neither has a single like-for-like successor. The common path is off-the-shelf edge GPU hardware, containerised models and central management through tools such as Azure Arc, AWS IoT Greengrass or a platform like NexaStack.
Will edge vision work with our existing cameras and MES?
Usually yes. Most edge servers read standard IP camera streams, and results can be sent to MES, SCADA or quality systems through standard interfaces. Confirm image quality on a sample first.
How much data do we need to train an edge vision model?
Less than before. Vision language models can work zero-shot for some tasks, and fine-tuning from a pre-trained model often needs a few hundred good examples to start. Accuracy targets decide the final amount.
What are the latest edge computer vision trends in 2026?
VLMs at the edge, retirement of vendor kits, Kubernetes-based edge runtimes, automatic hard-example capture and agents that act on what cameras see.
Scale edge vision from one line to every site. Book a NexaStack edge vision demo.
