Industrial AI · Reference architecture
Industrial AI reference architecture
An industrial AI reference architecture is a complete design for putting AI into one industrial operation: the sources it reads, the object model that joins them, the open-weight models it runs, the compute sized from those models, the three-year cost and the named person who approves every action. CodeNinja Atoms has published 23 of them, one per real operation, across 9 sectors in the United States (12), Pakistan (6), Saudi Arabia (4), Germany (1). Each is free to reuse under CC BY 4.0 with a DOI.
23published designs9sectors4countries317ontology objects57open-weight model choices
An industrial AI reference architecture for each sector
One design per real operation. Start from your own sector; the nearest design is the one to read first.
| Sector | Designs | Countries |
|---|---|---|
| agriculture and earth observation | Baseline, Fodder Watch, Field Ledger | Pakistan, Saudi Arabia, United States |
| aviation and aerospace manufacturing | Hangar Form Cell | United States |
| discrete manufacturing and automotive | Depot Operating Object Model, Pit Camera Watch | United States |
| energy and utilities | Turbine Island Watch, Reliability Atlas, Grid Context Watch, Feeder Firewatch | Pakistan, Saudi Arabia, United States |
| heavy industry and construction | Structure Phase Watch, Steel Count Ledger, Factory Fire Watch | Pakistan, Saudi Arabia |
| maritime and ports | Port Drone Watch, Port Twin, Terminal Pulse | Germany, United States |
| oil and gas | Pipeline Integrity Watch, Corridor Guard, Loop Integrity Watch, Sovereign HSE Watch | Pakistan, United States |
| rail and transportation | Wayside Trend Watch, Consist Watch | United States |
| warehousing and intralogistics | Warehouse Twin Command | United States |
Industrial AI reference architecture patterns
Three deployment patterns cover all 23 published designs, grouped by rule from each design's own compute figure. Pick the one that matches the hardware your operation can own.
| Pattern | What runs where | Designs |
|---|---|---|
| Frontier reasoning node (11) | eight 141 GB HBM-class GPUs holding an open-weight frontier model at FP8, owned or hosted in the country | Wayside Trend Watch, Warehouse Twin Command, Pipeline Integrity Watch, Port Drone Watch, Consist Watch, Grid Context Watch, Structure Phase Watch, Steel Count Ledger, Feeder Firewatch, Terminal Pulse, Sovereign HSE Watch |
| One inference card (6) | 48 GB PCIe inference cards, from one card to one eight-card server, at the site or in the country; no frontier model | Depot Operating Object Model, Turbine Island Watch, Hangar Form Cell, Corridor Guard, Pit Camera Watch, Fodder Watch |
| No new compute (6) | the design runs on servers or software the operator already owns; nothing to buy | Baseline, Field Ledger, Loop Integrity Watch, Reliability Atlas, Port Twin, Factory Fire Watch |
The layers of an industrial AI reference architecture
Every published design follows the same seven layers, bottom up. Across the 23 designs the object model layer holds 317 objects and 295 typed links, the inference layer names 57 open-weight model choices, and 23 of 23 designs name the person who approves each action. The sovereign boundary is what makes it sovereign: every layer runs on hardware the operator owns or controls, in its own country, with no third-party AI service in the serving path.
| Layer | What it holds |
|---|---|
| 1. Sources | the systems of record the operation already runs, read only; they stay the system of record |
| 2. Sensing | cameras, detectors and instruments, existing or added, sized from the streams they produce |
| 3. Adapters | one family per kind of source, owning its contract, retries and ordering; the only way data enters |
| 4. Object model | the operation as typed objects, properties and links (an ontology), held on infrastructure the operator controls |
| 5. Inference | open-weight models only, placed at the edge, the site or one in-country node by latency and link loss |
| 6. Services | what the design computes over the object model: forecasts, detections, joins, evidence |
| 7. Surfaces | where people see the result and a named person decides; actions are written one way into the owning system |
Every published industrial AI reference architecture
Newest first. The compute and cost columns are each design's own At a glance figures.
| Design | Sector | Country | Compute | Three-year cost | DOI |
|---|---|---|---|---|---|
| Wayside Trend Watch | rail and transportation | United States | At the new four-camera yard-entry portal, an octa-core NPU edge box sized from its streams, so only detections, OCR strings and short evidence frames cross a cellular link that drops for hours; centrally, one 141 GB HBM GPU class card in one node in the railroad's own data center, holding the 65.2 GB checkpoint with a 1.2 planning factor and about 63 GB to spare | About 54,000 US dollars to own the central card for three years (a range of 52,000 to 57,000, card, support and power, host server by quotation), against about 625,000 dollars to rent the smallest AWS instance that holds one such card on the deepest three-year commitment in US East and about 1.66 million on demand; the edge box, portal cameras and pilot locomotive sensing are priced by quotation (Appendix A) | 10.5281/zenodo.23556421 |
| Warehouse Twin Command | warehousing and intralogistics | United States | At each site, one industrial edge server with one 48 GB PCIe inference card running vision in Triton and the site model under vLLM; centrally, one 141 GB HBM class node of eight cards, on the operator's own hardware | About 597,000 US dollars to own the twelve cards (510,000 to 689,000) before the four edge chassis, which are priced by quotation; renting the same cards in AWS US East runs about 1.86 million on demand and about 699,000 on the deepest three-year commitment (Appendix A) | 10.5281/zenodo.23556416 |
| Pipeline Integrity Watch | oil and gas | United States | At each station, volume production edge modules in 7W to 60W envelopes inside hazardous area enclosures, detecting through backhaul loss; centrally, one node of eight 141 GB HBM-class GPUs in the gas control center | About 465,000 US dollars to own the control center node for three years (432,000 to 485,000), about three quarters of the 625,000 the deepest AWS commitment in US East costs and well under the 1.66 million on demand; the station modules, enclosures, cameras, diodes and sensors are priced by quotation (Appendix A) | 10.5281/zenodo.23556374 |
| Depot Operating Object Model | discrete manufacturing and automotive | United States | Fanless DIN-rail edge boxes beside each machine group and at the test cell, their class fixed by a phase one measurement of the vibration load; centrally, two cards of the 48 GB PCIe class air cooled in the depot's own server room on its existing power, holding the 65.2 GB checkpoint with a 1.2 planning factor, about 78.2 GB against roughly 88 GB usable | About 67,000 US dollars to own the two-card server for three years (a range of 62,000 to 70,000: server, support and power), against about 104,000 dollars to rent the same card class on the deepest three-year commitment in US East and about 276,000 on demand; the cheapest rental that holds the model, one 96 GB card, is about 33,000 on that commitment, so ownership is chosen for the boundary, not the price, and renting is a price reference only because data and weights stay in the government's environment; edge boxes, sensors, cameras, time card and switches are priced by quotation (Appendix A) | 10.5281/zenodo.23570177 |
| Port Drone Watch | maritime and ports | Germany | Three tiers on the authority's own hardware: a GPU-class edge box at each dock, sized at the dock site survey; a site server at each port site with one to four 48 GB PCIe cards; and one node of eight 141 GB HBM cards (1,128 GB) in the port's own data centre, holding 753 GB of FP8 weights with a 1.2 planning factor, 904 GB. Detection events and evidence crops cross the boundary; pixels stay at the dock | About 735,000 US dollars to own the central node and the two site servers for three years (a range of 650,000 to 845,000: servers, support and power at the Eurostat non household tariff for Germany), against about 761,000 to rent the same card classes on the deepest three-year commitment in AWS Europe (Stockholm for the eight-card node, which Frankfurt does not sell, and Frankfurt for the site cards), so ownership costs about the same there, and about 2.02 million on demand, against which ownership costs about a third. Renting is a price reference only, because the imagery, the weights and the detection history stay on hardware the port owns; dock edge boxes are priced by quotation after the survey (Appendix A) | 10.5281/zenodo.23635729 |
| Turbine Island Watch | energy and utilities | Pakistan | One 48 GB PCIe inference GPU class card on ruggedised edge servers inside the plant perimeter, a PTP grandmaster time card and a one-way egress path; no cloud, because Pakistan has no hyperscaler region | About 22,500 US dollars to own the site envelope over three years (21,600 to 23,400), about the same as the deepest three-year AWS commitment for one L40S card in the UAE region, which could not keep the read one-way in any case (Appendix A) | 10.5281/zenodo.23365092 |
| Consist Watch | rail and transportation | United States | At the car, an EN 50155 and IEC 60571 module (8 by 8 by 8 inches, IP62, 24 to 48 VDC, CFR 238 materials); centrally, one 141 GB HBM GPU class node in a contractor-hosted US cloud | About 625,000 US dollars to run the eight-card node for three years on the deepest AWS commitment in US East, about three eighths of the 1.66 million on demand; the 225 modules are priced by quotation (Appendix A) | 10.5281/zenodo.23365118 |
| Hangar Form Cell | aviation and aerospace manufacturing | United States | A fanless industrial inference computer in the cell panel and an eight-card 48 GB PCIe inference server in the facility, joined one way across a hardware data diode; an OCP Time Card grandmaster orders every stream | About 202,000 to 220,000 US dollars to own the priced classes over three years, the scanner and the server most of it; the two compute tiers alone cost about two fifths of the deepest three-year AWS commitment for eight L40S cards, which the offline requirement rules out anyway (Appendix A) | 10.5281/zenodo.23308095 |
| Corridor Guard | oil and gas | Pakistan | One server-class host of the 48 GB PCIe inference GPU class inside the operator's estate; no field compute, no cloud, nothing crosses the boundary | About 19,500 US dollars to own the host over three years (18,700 to 20,200), about four fifths of the deepest three-year AWS commitment for one L40S card, which could not keep the data in Pakistan in any case (Appendix A) | 10.5281/zenodo.23308093 |
| Pit Camera Watch | discrete manufacturing and automotive | United States | One 48 GB L40S-class card per recorder host, in the conditioned server room; nothing at the poles | About 474,000 to 871,000 US dollars to own over three years, mostly camera heads; the two recorder hosts alone cost about four fifths of the deepest three-year AWS commitment, and there is no per camera license at any pole count (Appendix A) | 10.5281/zenodo.23296077 |
| Baseline | agriculture and earth observation | United States | The operator's ArcGIS Enterprise; nothing runs in a cloud the operator does not control | No compute to price; the acquisition is priced by the survey firm against the operator's own task table (Appendix A) | 10.5281/zenodo.23186673 |
| Fodder Watch | agriculture and earth observation | Saudi Arabia | One 48 GB L40S-class inference node inside the Kingdom; the card class needs a US export licence | About 15,200 dollars to own the node, about two thirds the deepest three-year AWS commitment, and renting would move the data outside the Kingdom (Appendix A) | 10.5281/zenodo.23186675 |
| Field Ledger | agriculture and earth observation | Pakistan | Open-source PostgreSQL with PostGIS on the operator's own servers; source code and full intellectual property handed over | No hardware line to price: the cost is the integration and software work (Appendix A) | 10.5281/zenodo.23186671 |
| Loop Integrity Watch | oil and gas | Pakistan | Two containers on the operator's own site application server, inside its OT boundary | No compute to price; the field equipment is priced by OEM quotation against the survey (Appendix A) | 10.5281/zenodo.23157967 |
| Reliability Atlas | energy and utilities | Saudi Arabia | No hardware bought; the study runs on the operator's own in-country servers | No hardware line to price: the cost is the assessment work itself (Appendix A) | 10.5281/zenodo.23157965 |
| Grid Context Watch | energy and utilities | United States | One node of eight 141 GB HBM-class GPUs: 753 GB of FP8 weights, 904 GB with the 1.2 planning factor, against 1,128 GB | About 510,000 dollars to own the reasoning node, about four fifths the deepest three-year AWS commitment (Appendix A) | 10.5281/zenodo.23157957 |
| Structure Phase Watch | heavy industry and construction | Saudi Arabia | One node of eight 141 GB HBM-class GPUs holds GLM 5.3 at FP8 (753 GB of weights, 904 GB with headroom) | About 642,000 US dollars with support and power at the Saudi industrial tariff | 10.5281/zenodo.23126448 |
| Steel Count Ledger | heavy industry and construction | Pakistan | One node of eight 141 GB HBM-class GPUs holds GLM 5.3 at FP8 (753 GB of weights, 904 GB with headroom), in a dedicated data center behind an export licence checkpoint | About 4,445,000 US dollars for 300 installation points, of which the counting kits are 2,832,000 and cannot be rented; owning the frontier node costs about 686,000 against 751,000 on AWS's deepest three-year plan | 10.5281/zenodo.23126563 |
| Port Twin | maritime and ports | United States | No equipment bought: about 1.1 GB of fp16 weights on the port's existing enterprise virtual machines | No hardware line to price: the design adds software and integration work to servers the port already runs | 10.5281/zenodo.23126431 |
| Factory Fire Watch | heavy industry and construction | Saudi Arabia | None bought: the requirement asks for no server or GPU, and the field gateways carry no model runtime | About 328,000 to 439,000 US dollars with support and power; the eight-factory pilot is about 21,000 to 26,000 of equipment | 10.5281/zenodo.23126565 |
| Feeder Firewatch | energy and utilities | United States | One node of eight 141 GB HBM-class GPUs holds GLM 5.2 at FP8 (753 GB of weights, 904 GB with headroom) | About 841,000 US dollars with support and power at the Texas industrial power price | 10.5281/zenodo.23159328 |
| Terminal Pulse | maritime and ports | United States | One node of eight 141 GB HBM-class GPUs holds GLM 5.3 at FP8 (753 GB of weights, 904 GB with headroom), at about 10 kW | About 722,000 US dollars with support and power at the US industrial power price | 10.5281/zenodo.23159331 |
| Sovereign HSE Watch | oil and gas | Pakistan | One node of eight 141 GB HBM-class GPUs holds GLM 5.3 at FP8 (753 GB of weights, 904 GB with headroom) | About 670,000 US dollars with support and power at Pakistan's industrial tariff | 10.5281/zenodo.23119714 |
Built from the published designs only: each paper's At a glance table and the dataset at Hugging Face. The same designs as sovereign AI reference architectures; their object models at industrial ontology examples; also read as physical AI system design and as the output of a self-improving AI system. Questions by sector at sectors. Designed on Praxis; object models load into Hyper Ontology. Updated 2026-10-26.
Want one of these designs for your own operation? Talk to us.