Private 5G for AI Video Analytics: Uplink Sizing Guide

Private 5G for AI video analytics works when you design for the uplink first. Cameras, drones and smart helmets send video up to an edge AI server, but most 5G frames are built for downloads. Size the uplink per cell, place inference on site, and pick a TDD pattern and camera profile that fit, and you can run computer vision for safety, security and quality at industrial scale.

The timing matters. On 6 November 2026, e& UAE announced a contract for what it called the world's first commercial mobile network on Upper 6GHz spectrum, and Huawei speakers at the same summit pointed to traffic shifting toward uplink-heavy flows. In parallel, EMSTEEL and e& UAE piloted a private 5G network with AI safety cameras that detect missing protective equipment in real time. AI video over private 5G has moved from demo to procurement. If you are still at the licensing stage, read our guide to the TDRA private IMT framework first.

Consumer 5G was built to deliver video to phones. Industrial AI does the opposite: hundreds of cameras push video up to an analytics engine that never sleeps. NGMN notes that all currently supported TDD frame structures are more or less downlink heavy, with 8 downlink slots and 2 uplink slots in the common pattern. A network that looks fast on a speed test can still choke when 150 cameras stream at once.

Ookla's 2026 AI readiness research makes the same point for public networks: operators need to prioritise uplink, latency under load and cloud reach for AI traffic. The UAE scores well on upload, but a factory, port or campus needs guaranteed capacity, not a national median. That is what a private network, or a dedicated slice, is for.

There is no universal number. It depends on spectrum, TDD pattern, antenna placement and radio conditions at each camera. The reliable method is to measure, then divide.

A worked example using the bitrate guide: 120 fixed 1080p H.265 cameras at 2 Mbps each need about 240 Mbps of sustained uplink. Add 20 body or helmet cameras at 4K H.265 and 5 Mbps each, another 100 Mbps, for roughly 340 Mbps in total. With 30 percent headroom you need close to 490 Mbps of measured uplink across the cells that serve those cameras. Machine vision for quality inspection is a different class: NGMN lists uplink needs from 80 Mbps to 12 Gbps for that use case, so treat it as its own design.

Four choices drive cost and performance more than anything else.

Edge inference also shrinks the problem upstream. When the AI model runs on site, only alerts, metadata and short clips leave the premises. For the trade offs in buying versus renting the GPUs behind that edge, see our GPU as a Service vs owned GPUs guide.

Many tenders still specify radio and cameras in separate lots, then discover in the pilot that the analytics cannot keep up. Write one integrated requirement set.

• ✓A stream inventory with resolution, frame rate, codec and analytics per camera group.
• ✓Measured sustained uplink per cell from a site survey, with headroom stated in the design.
• ✓The TDD pattern proposed and how coexistence with neighbouring networks is handled.
• ✓Edge inference location, GPU capacity and model accuracy targets per use case.
• ✓End to end latency target from camera to alert, tested under full load.
• ✓Data residency: where raw video, clips and metadata are stored, and for how long.
• ✓Licensing route under TDRA rules and who holds the spectrum or slice agreement.
• ✓Pilot acceptance criteria and a named owner for each layer: radio, edge, analytics, integration.

No single vendor covers radio, edge compute, analytics and site integration well. The projects that land on time use a clear prime contractor with specialist partners, which is the same structure we described for government AI tenders.