Edge AI for Aquaculture & Fish Farming 2026:
Underwater Vision, Biomass & Industrial PC Hardware

Aquaculture is the most hostile siting problem in food production. A salmon pen sits in open water, hundreds of metres offshore, reachable only by boat, powered from a feed barge or a solar-battery island, and wrapped in a chloride atmosphere that destroys ordinary electronics within a season. Land-based recirculating systems trade salt for the opposite problem: a permanently wet, warm, condensation-soaked hall where a fan failure is a stock-loss event. Both now depend on machine vision — Norway's Food Safety Authority has accepted automated lice counters in place of manual counting, and over 30% of Norwegian salmon sites run smart camera systems. Mandatory vision, no datacentre, no maintenance window: that is the case for edge computing.

Where the Compute Actually Sits

Aquaculture inference splits into three tiers, and picking one accelerator for the whole farm is the most common BOM mistake.

In-pen and in-tank sensing nodes. A subsea camera plus a compact module on the pen collar or tank wall. Its job is narrow and local: run one or two detection models on one or two streams, then transmit a decision — lice count per fish, size distribution, feeding intensity, a flagged thumbnail. Raw video never leaves the pen. Fanless Jetson Orin Nano and Orin NX hardware is the right register, and PoE the right power-and-data interface.

The feed barge or farm control room. The farm's decision layer: aggregating every pen, running the feeding controller, holding the day's clip archive, presenting one dashboard. With mains-equivalent power and several times the stream count, an AGX Orin or an x86 industrial PC with a low-profile GPU is the natural choice.

The land-based RAS hall. A recirculating facility is a 24/7 process plant — pumps, biofilters, oxygen cones, dozens of tanks — with the compute racked in a wet room doing process control and multi-camera vision at once. Treat it as a small industrial datacentre: redundant power, RAID, and a real thermal design.

What the Models Have to Do

The workload mix determines the silicon far more than the TOPS rating does, because these are mostly small, high-frame-rate detectors rather than large foundation models.

Two commercial anchors explain why farms pay for this. Camera-based biomass estimation reaches 95%+ accuracy against the 65–75% typical of manual sampling, and feed conversion ratio — the key economic number in the business — runs near 1.0 early in the cycle and climbs toward 1.5 as fish approach market size. A tenth of a point of FCR across a pen holding 100,000-plus fish is worth more than the entire compute BOM. Note the last table row: water quality and actuation belong on a deterministic industrial gateway, not the vision GPU — separating those paths is what keeps a lice model from being blamed for an oxygen-dosing fault.

Compute Tiers Compared

The rule: the camera determines the module, the stream count determines the tier. One camera per pen is an Orin NX problem, not an AGX Orin problem — reach for 64 GB only when a single box aggregates several pens or fuses vision with acoustic and sensor data.

The Environmental Spec That Decides the BOM

Aquaculture diverges from every other edge AI vertical here. Farm hardware sees salt fog, pressure washing, condensation and constant flex, on a service cycle measured in boat trips.

Two items are routinely underestimated. First, the fan: a dust-and-salt filter on an outdoor cabinet is a maintenance item with a boat trip attached, which is why fanless is the default rather than a preference. Second, camera interface: subsea machine-vision cameras are GigE Vision with PoE and wide input voltage — a 3 MP subsea unit at 55 fps runs 10–36 VDC or 17–72 VDC over PoE — so the hub needs real PoE budgeting at the pen, not a generic 8-port switch.

Storage and Retention Math

Underwater video is expensive to move and expensive to keep, and retention policy is where farm projects quietly overspend. A continuous 1080p underwater camera runs 10–20 GB/day and a 2–4 MP stereo pair 30–60 GB/day — while the lice-count and biomass telemetry that actually matters is measured in kilobytes. Archive decisions and events, not raw video: a pen node with 1 TB of wide-temperature, power-loss-protected NVMe holds 30–90 days of event retention for 2–4 cameras and transmits kilobytes rather than gigabytes over a metered satellite or 5G link. Power-loss protection is not optional on a node that loses power when the barge does — a yanked write can corrupt the evidence you need for a welfare or regulatory record.

Specification Checklist

1. Fanless, sealed chassis with published wide-temperature range and IP66/IP67 (IP69K in washdown zones).
2. Conformal-coated boards, 316 stainless or marine aluminium hardware, with salt-mist test evidence.
3. 9–36 VDC wide input with surge protection and a controlled-shutdown UPS.
4. PoE ports budgeted for the real camera load — including inrush and cable-length voltage drop to the pen.
5. Wide-temperature, power-loss-protected NVMe with a published TBW figure, mirrored where the archive is a compliance record.
6. Separated control and vision paths: industrial gateway for water quality and actuators, GPU node for vision.
7. 7–10 year availability statement — a farm is a 15-year asset, and a component EOL forces requalification.
8. Secure boot, TPM 2.0 and signed OTA updates — the node is unattended and reachable only by boat.

QSCompute supplies Jetson Orin Nano Super, Orin NX and AGX Orin modules for in-pen nodes, AGX Orin and x86 industrial edge systems for feed barges and RAS halls, fanless IP67 and IP69K stainless industrial PCs for wet environments, and the wide-temperature NVMe, PoE switches and wide-input PSUs these sites require — validated against your site's worst-case conditions, with DDP shipping to 85+ countries.