A Hikvision AcuSense DS-2CD3786G2T camera ships 1.5 TOPS of on-chip compute and microSD storage up to 512 GB[3]. A Jetson Orin Nano Super delivers 67 TOPS at 7-15 W[2][5]. A query to a model in the cloud lives at 50-500+ ms round-trip latency[1][4].

The same question — “person or vehicle?” — gets answered in <3 ms on edge or hundreds of milliseconds on cloud[4].

The operational difference is brutal when you cross it with three real cases:

  • Store losing connectivity for 4 hours: edge keeps classifying, recording, triggering local alarms. Cloud is dead until the uplink returns.
  • Drive-thru camera that needs an alarm in <500 ms to activate an audio deterrent: edge handles it. Cloud at 200 ms network + 80 ms inference barely fits the window.
  • Point of sale with 8 cameras on a 50 Mbps link: pushing 8 streams of 4K to the cloud is 64 Mbps sustained — doesn’t fit. Processing local + sending only events changes the math entirely.

( IBM puts it this way: “Edge AI is low bandwidth because it processes data locally; Cloud AI is high bandwidth because it requires a network for data transmission to remote servers” — direct from their 2025 guide )

In every tv4agent deployment where the client says “I want analytics in the cloud”, the question that matters most is not “how accurate is the model?” It’s “what happens in your store when the internet drops?”. The answer defines whether you need edge, cloud, or a hybrid where inference runs local and dashboards query remote.

P.D. Does your current video analytics system run in the camera, in an NVR with a GPU, on a local server, or in a data center in another country? If you can’t answer that in one line, today’s a good day to open the device’s config page.

📰 Source: IBM Think — Edge AI vs Cloud AI; NVIDIA Jetson Orin Nano specs; Hikvision AcuSense DS-2CD3786G2T datasheet; NanoGPT 2026; Savant AI 2026.

Frequently asked questions

What’s better for CCTV: edge AI or cloud AI? Depends on 3 factors: (1) required latency, (2) available bandwidth, (3) data regulation. Edge AI = processing on the camera or on a local server. Cloud AI = processing on a remote server. Edge wins on latency (<100ms vs 1-5 seconds) and on regulation (no data transfer). Cloud wins on low initial cost and scalability. For 1-10 sites with good internet, cloud is cheaper. For 10+ sites or strict regulation, edge wins.

How much does edge AI cost? Initial CAPEX: 2-3× more than a non-AI IP camera. For example, standard IP camera $300 USD vs AI-edge camera (Hailo-8, Jetson Orin) $700-1,200 USD. But much lower OPEX: $0/month on cloud vs $50-200 USD/camera/month on cloud inference. For a 50-camera deployment over 5 years, edge TCO is 30-40% lower than cloud. Higher initial investment, lower TCO.

How much does cloud AI cost? Typical per-event model: $0.001-0.01 USD per inference (person, vehicle, etc. detection). If your camera generates 1,000 inferences/day = 365,000 inferences/year = $365-3,650 USD/camera/year. For 50 cameras, that’s $18K-180K USD/year. Above 100 cameras, cloud AI stops being competitive vs edge. Flat subscription models ($20-50 USD/camera/month) exist but are usually less flexible.

What about accuracy? Similar in normal conditions (90-95% accuracy on standard tasks). Cloud AI wins on complex tasks: facial recognition at >50m, plate OCR in extreme conditions, weapon detection. Edge AI wins on latency: a ‘person in restricted zone’ alert on edge reaches the operator in 100-500ms; on cloud, 1-3 seconds. For security alerts, the difference is operational. For forensic analytics (post-event), it doesn’t matter.

What about privacy? Edge AI wins. The image never leaves the site. For LATAM retail under LFPDPPP or for any deployment with biometric data, edge is the safest legal option. Cloud AI requires (1) contract with cloud vendor specifying jurisdiction, (2) end-to-end encryption, (3) audit of who accessed what footage and when. Edge eliminates all 3.

What happens if my internet goes down? Edge AI keeps working — cameras keep detecting and recording locally, no cloud needed. Cloud AI stops working — no internet, no inference. For 24/7 retail or critical deployment (airport, bank, datacenter), edge is the only reliable option. For an office with 9-5 hours and non-critical analytics, cloud is acceptable.

Sources

[1] IBM — Edge AI vs Cloud AI — https://www.ibm.com/think/topics/edge-vs-cloud-ai [2] NVIDIA Jetson Orin Nano Super Developer Kit — https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/ [3] Hikvision DS-2CD3786G2T-IZS AcuSense 8MP datasheet — https://www.hikvision.com/content/dam/hikvision/en/support/regional-materials/mea/DS-2CD3786G2T-IZSYH_Datasheet_MEA_2024Q1.pdf [4] NanoGPT — Edge vs Cloud AI latency and cost comparison — https://nano-gpt.com/blog/edge-ai-vs-cloud-ai-cost-latency-comparison [5] Savant AI — When NVIDIA Jetson Orin is the right fit — https://savant-ai.io/blog/jetson-orin-right-fit