Small Business Edge AI System: 4 Camera Budget Architecture

Verdict

For a 4-camera small business deployment, Jetson Orin Nano Super is the best starting point. It keeps cost low while providing enough headroom for basic object detection, people counting, and event alerts.

Try this in System Designer and adjust assumptions for your site.

Architecture Overview

This pattern is optimized for affordability. Keep inference local, store recent video and events on-device, and send only lightweight alerts or clips upstream.

Deployment Summary

Platform Choice Under a Budget Cap

The compute module is the one component you cannot cheaply swap later, so it deserves the most scrutiny. The requirement for this architecture is concrete: decode four 1080p streams and run a small detection model on each at 15 FPS. That is 4 × 15 = 60 inferred frames per second in aggregate.

Jetson Orin Nano Super in its 15W power mode runs YOLO11n at roughly 270 FPS (TensorRT, INT8, batch 1, 640×640 input) and YOLO11s at ~197 FPS. Against a 60 FPS aggregate requirement, that is 3–4× raw model throughput headroom — enough to absorb video decode, preprocessing, and tracking overhead, which is why practical stream capacity lands at 4–6 concurrent 1080p streams rather than the raw benchmark number. Step up to a mid-size model (YOLO11m, ~110 FPS at 15W) and the margin drops to under 2×; that is the point where this build stops being comfortable.

Why not the alternatives

The general rule from the kit selection guide applies: over-specifying hardware is the expensive mistake at scale, but with a single node the delta between Orin Nano Super and the cheaper boards is small enough that software maturity wins. One caveat on TOPS: ratings are not comparable across architectures — a 6 TOPS RK3588 NPU and a 67 TOPS Ampere GPU differ in memory bandwidth, supported operators, and runtime overhead, so validate with real benchmarks, not spec sheets.

Recommended Stack

Camera Layer

Prioritize coverage of business-critical areas: entry points, POS counter, inventory exits, and customer queue zones. Better placement beats extra camera count.

Alert Layer

Keep alerts local-first using app/web notifications and simple webhooks. Trigger workflows only for meaningful events to avoid alarm fatigue.

Compute Layer

Orin Nano Super is a cost-efficient baseline for 4 streams. Move to Orin NX if workloads add heavier models, higher FPS, or expansion toward 6-8 cameras.

Cost Breakdown

A realistic hardware budget for this build is $800–1,500, in line with the starter tier in the recommended builds guide. The honest part most vendor pages skip: the AI compute is not the biggest line item. Four cameras plus the network and power infrastructure around them typically cost more than the Jetson.

Operating cost is small but nonzero: at $0.10/kWh, 100W of continuous draw costs about $87/year. This system's 55–75W total load works out to roughly $48–65/year at that rate — scale by your local tariff.

Power and Performance

Size the PoE switch the same way the builds guide sizes larger systems: sum the camera nameplate draw, add ~10W switch overhead, then add 25% headroom. Four cameras at 10W worst case is 40W + 10W + 25% ≈ 63W — which is why a 60W switch is marginal and the 90–120W tier is the safe floor. PoE budget exhaustion does not fail loudly; it shows up as individual cameras rebooting under load.

Compute Power Modes

The Jetson's share of the budget depends on its nvpmodel power mode. Measured figures from the Orin Nano power consumption benchmarks (JetPack 6.2, YOLO11 TensorRT INT8 workloads):

Expected Performance

Network Math

Per-camera bitrate is the number everything downstream hangs on — network load and storage sizing both derive from it. Using EdgeAIStack's measured camera bitrate table (1080p, 30fps, medium scene complexity, CBR):

The conclusion is unambiguous: a gigabit switch is not a bottleneck at 4 cameras, and it will not become one at 8 or 16. Camera-side 100 Mbps PoE ports are also fine — no single 1080p stream approaches them. Do not buy 2.5GbE for this architecture; spend the difference on PoE budget or storage endurance.

Uplink and remote viewing

The LAN is trivial; the WAN uplink is where SMB deployments actually hurt. This architecture sends only event clips and alerts upstream, so steady-state upload is near zero. The exception is remote live viewing: each remotely watched stream adds its full bitrate to your upload path, and a 5 Mbps H.264 main stream can saturate the upload side of a typical asymmetric business connection on its own. Serve remote viewers the camera's low-resolution substream, or relay a re-encoded copy from the Jetson, and keep full-bitrate video on the LAN.

Storage Sizing

Storage is a straight chain from bitrate: Mbps ÷ 8 = MB/s → × 86,400 = GB/day → × retention days ÷ 0.8 = drive capacity (the ÷ 0.8 reserves 20% for OS, models, and filesystem overhead, per the Jetson NVMe guide). One continuous Mbps writes about 10.8 GB/day.

This is why the stack specifies H.265 and why event-based recording matters on a budget: the codec choice alone halves the drive you have to buy. Event-first recording cuts write volume further in low-traffic scenes, but size the drive for the continuous case so a busy week does not silently shorten your retention window.

Endurance check

Capacity is only half the sizing. Continuous H.265 recording at 108 GB/day is ~39 TB/year of writes; with a 1.2–1.3× write amplification factor that is roughly 50 TB/year, or ~250 TB over five years. A ≥360 TBW consumer TLC drive covers that with margin; the H.264 profile (~99 TB/year with WAF) needs the ≥600 TBW class. Formula from the NVMe guide: daily writes × WAF 1.2–1.3 × years + 30% margin.

Why NVMe, not SD card

• Endurance: UHS-I SD cards are rated around 10–30 TBW lifetime vs 600+ TBW for a decent NVMe drive. At ~50 TB/year of ring-buffer writes, an SD card is a consumable measured in months.
• Sustained writes: SD sustains 50–100 MB/s vs 200–350 MB/s NVMe. Four cameras only need ~1–3 MB/s, but SD write-latency spikes under mixed I/O cause dropped segments.
• Boot and resilience: JetPack 6.x boots from NVMe in 15–25s vs 45–60s from SD, and NVMe filesystem journaling survives unclean shutdowns far better. NVMe boot is the production standard for unattended nodes.

Prefer a TLC drive with DRAM cache: DRAM-less (HMB) drives borrow 64–128 MB of the Orin's shared 8 GB memory, which comes straight out of the inference pool.

Install Realities for Small Businesses

An SMB has no server room. The Jetson, switch, NVMe, and UPS end up in a back office, above a drop ceiling, or in a stock room — and that changes the engineering.

• Placement: Put the node where the camera cable runs converge, not where it is convenient to look at. Every meter of Cat6 through finished walls costs more in labor than the cable. Wired Ethernet throughout — Wi-Fi backhaul for cameras adds latency variance and a second failure domain.
• Thermals: At the recommended 15W mode the module runs 8–16W, which passive cooling handles only with genuine airflow. A sealed cabinet or a shelf above a fridge compressor is where fanless plans die: the SoC throttles at 80°C and clock oscillation shows up as inconsistent alert latency, not a clean error. Fit the small axial fan; it is the cheapest reliability component in the BOM.
• The NVMe has thermals too: In a fanless or semi-sealed enclosure, bridge the drive to the chassis with a 1.0–1.5mm thermal pad rather than a desktop-style finned heatsink that assumes airflow. Validate before commissioning: 15 minutes of sustained write load should hold the drive ≤65°C.
• Dust and noise: Retail and cafe back rooms are dustier than offices. Filtered vents beat open mesh, and cleaning the filter goes on the same quarterly checklist as checking SMART data. Noise matters in customer-adjacent spaces — one quiet 40mm fan at 15W mode is inaudible; a switch with a screaming server fan is not, so check the switch's cooling design too.
• Power: The site's wall power is shared with espresso machines and HVAC. The 600–1000 VA UPS is not optional equipment — it is what stands between a brownout and a corrupted ring buffer, and its USB port should trigger a clean shutdown, not just delay the crash.

Bottlenecks and Failure Modes

Primary risk: treating this like an 8-camera system. The budget build works best when model size, FPS, and retention expectations stay aligned with the hardware tier.

Scaling Decisions

• 1-2 cameras: Orin Nano Super is typically more than enough.
• 4 cameras: Orin Nano Super is the default recommendation.
• 6 cameras: validate carefully and tune FPS/model complexity.
• 8 cameras: move to Orin NX 16GB.
• Multi-model analytics: prefer Orin NX class hardware.

Upgrade Path: 4 to 8 Cameras

If there is any realistic chance the site grows to 8 cameras, the cheapest time to prepare is now — but only for the components where the delta is small. The expensive mistake is buying twice; the other expensive mistake is buying the 8-camera compute tier for a 4-camera site.

Buy differently today

• Switch: Skip the 8-port unit and buy the 16-port PoE+ tier (180–250W). The 8-camera math from the builds guide — 8 × 13W + 10W overhead + 25% headroom ≈ 142W — already exceeds an 8-port/120W switch, so it would be replaced, not extended.
• Cabling: Pull spare Cat6 drops to likely future camera positions while the walls are already open. Cable is the cheapest component and the most expensive retrofit.
• Storage endurance: Eight 1080p30 H.264 streams write ~64 Mbps ≈ 691 GB/day — a different endurance class entirely. If growth is planned, buy 2 TB at ≥600 TBW now instead of 1 TB at 360 TBW.
• UPS: Step to the 1000–1500 VA pure sine tier so the doubled PoE load still gets useful runtime.

Swap later

The compute is the one component to defer. At 8 cameras, move from Orin Nano Super to Orin NX 16GB (~$599, 157 TOPS, ~12 practical streams) — same module form factor, so a compatible carrier or dev-kit-style deployment carries over, and the TensorRT/DeepStream software stack is unchanged. Between 4 and 8, the Nano Super can often stretch to 6 streams by dropping inference FPS or staying on YOLO11n, or by moving to 25W Super mode with active cooling; validate in the System Designer before spending.

For the full 8-camera pattern, see the Retail 8-Camera reference architecture.

Validate This Architecture With EdgeAIStack

• System Designer — recommendation, headroom, risks, and alternatives.
• Network Bandwidth — stream load estimation for 4 cameras.
• Storage Endurance — SSD/NVMe retention and endurance sizing.
• Power Budget — PoE and compute power checks.

FAQ