Edge AI video surveillance analyses camera feeds on the camera itself or on a local GPU server, so alerts fire in real time and raw footage never has to leave the site. In 2026 most enterprises use a hybrid design: lightweight detection on cameras, heavier analytics and video AI agents on an on-site edge server, and only events, clips and metadata sent to a central platform.
Updated October 2026
What changed in 2026
- Video AI agents reached the edge. NVIDIA's video search and summarization blueprint deploys "from the edge to the cloud", including on Jetson Thor, and Jetson Platform Services can run the generative AI app entirely on a Jetson Orin device (NVIDIA).
- Cloud video AI runs on site. Azure AI Video Indexer enabled by Arc runs video analysis and generative AI on edge Kubernetes clusters and supports live streams (Microsoft Learn).
- A closed edge appliance retired. AWS Panorama, an edge computer vision appliance service, reached end of support on 31 May 2026 (AWS docs).
- Surveillance rules tightened. The EU AI Act bans untargeted scraping of CCTV to build facial recognition databases and real-time remote biometric identification for law enforcement in public spaces (European Commission).
Who this is for: heads of security, facilities and IT infrastructure at campuses, plants, warehouses, airports, retail chains and city operators who are upgrading CCTV from recording to real-time detection.
This guide covers the architecture. For privacy, compliance and use cases, read our pillar guide to AI video surveillance.
The business problem: cameras record, nobody watches
Most organisations already own hundreds or thousands of cameras. Almost all of that footage is only reviewed after an incident. Operators cannot watch every screen, so the system works as evidence storage, not prevention. Streaming every camera to the cloud for AI is usually ruled out by bandwidth cost and data rules.
Edge AI flips the model. Analysis happens next to the cameras, operators only see events that matter, and footage stays on site unless it is needed.
Three places to run the AI
| Option | Strengths | Limits | Best for |
|---|---|---|---|
| AI on the camera | No extra server, lowest latency per camera | Small models, vendor-specific, hard to update across brands | Simple detections: line crossing, people, vehicles |
| On-site edge GPU server | Works with existing cameras, larger models and VLMs, data stays local | Hardware to buy and manage | Multi-camera analytics, safety, video agents |
| Cloud video AI | Elastic, large models, easy to start | Bandwidth, latency, data-residency limits | Forensic search on selected clips, small sites |
| Hybrid (recommended) | Real-time on site, central search and governance | Needs a control plane for many sites | Multi-site enterprises |
Reference architecture
existing CCTV, RTSP/ONVIF
decode, detect, track
verify event, summarise, decide
VMS alarm, access control, operator
Figure: Existing cameras stream to an on-site GPU server that detects and verifies events; only events and short clips go to a central plane that manages models, search and audit across sites.
This mirrors how NVIDIA describes Metropolis, a "vision AI application platform" for visual AI agents running "from the edge to the cloud" (NVIDIA). The video agent step is new in 2026: a vision language model checks an alert before it reaches an operator and writes a plain-language summary, which cuts false alarms.
Sizing questions to answer before you buy
- How many cameras per site, and at what resolution and frame rate do you need to analyse them? Many detections work on reduced frame rates.
- Which analytics run on every frame (intrusion, PPE) and which run only on events (summaries, search)?
- How long must footage and events be kept, and where?
- Which systems must receive alerts: your VMS, access control, ticketing, radio?
- What happens when the WAN link drops? The edge should keep detecting and queue events.
Ask vendors to benchmark on your own footage and hardware. Published stream counts vary widely with model, resolution and frame rate. For the GPU side, see GPU parallel processing for vision and real-time video streaming with edge AI.
Want to know what your existing cameras can do with edge AI? Book a free surveillance edge sizing review. Bring your camera count and use cases, and we will outline the hardware and architecture.
Privacy by design
- Detect behaviours and objects (intrusion, loitering, PPE, unattended bags) rather than identities where you can.
- Keep raw video on site; send events and short clips only.
- Blur or mask faces in exported clips unless there is a lawful need.
- Log every alert, model version and operator action for audit.
How to start in 30 days
- Week 1: pick one site, 10 to 20 cameras and two use cases with a clear cost, such as after-hours intrusion and blocked fire exits.
- Week 2: install an edge GPU server, connect cameras over RTSP or ONVIF and deploy detection models.
- Week 3: run in shadow mode, log alerts and have operators mark true and false alarms.
- Week 4: tune thresholds, add event verification, connect to the VMS and decide on rollout.
If you are migrating from AWS Panorama, read our Panorama migration guide.
Where it's heading: 2026 → 2030
| Period | Edge video surveillance |
|---|---|
| Around 2020 (history) | Motion detection and basic object classes; many false alarms; footage reviewed after incidents. |
| 2026 (now) | Edge GPUs run detection plus vision language models that verify and summarise events; natural-language video search. |
| 2028 (expected) | Our view: video agents coordinate responses across cameras, doors and staff, with humans approving actions. |
| 2030 (our view) | Sites run as "perceiving spaces" where video is one sensor among many, governed centrally. |
How NexaStack helps
NexaStack Vision AI analyses streams in real time with edge processing for "low latency and secure, on-site decision-making," and its use cases include flagging suspicious activity or unattended objects in public spaces. The platform deploys from edge devices to private cloud, with self-hosted and air-gapped options, and centrally manages models across sites. Read more about our approach to real-time video analytics.
FAQ
What is edge AI video surveillance?
It is CCTV where AI models analyse video on the camera or a local server, so alerts are immediate and raw footage stays on site.
Do I need to replace my cameras?
Usually not. An on-site edge server can analyse streams from most standard IP cameras.
Is it cheaper than cloud video analytics?
For many cameras running continuously, edge is often cheaper because you avoid streaming all video upstream. For a few cameras or occasional search, cloud can be simpler. Model your own bandwidth and hardware costs.
How do we stay compliant with privacy laws?
Detect behaviours rather than identities where possible, keep raw footage on site, mask faces in exports, log decisions and check each use case against local law and the EU AI Act where it applies.
What are the latest edge AI video surveillance trends in 2026?
Vision language models for event verification, natural-language video search, video agents at the edge, hybrid edge-cloud designs and stricter AI rules on biometric use.
Turn your CCTV into real-time detection without moving footage off site. Book a NexaStack edge video pilot discussion.
