Read cofounder Muhammad Umar's memo, Compression and Reinflation
CodeNinja AtomsUnited StatesSaudi ArabiaPakistan

Sector · Maritime and ports

How ports and terminals use physical AI

How are ports and terminals using physical AI? Terminals and port authorities run gates, cranes, yards and finance on separate systems, so nobody sees the whole flow of a box or a truck. These are the questions engineers ask first, answered from 4 published CodeNinja Atoms reference architectures.

  1. What does a physical AI system for maritime and ports look like?
  2. Which AI models can a maritime and ports operator run on its own hardware?
  3. How much compute and hardware does AI in maritime and ports need?
  4. Is it cheaper to own AI hardware or rent cloud GPUs in maritime and ports?
  5. What ontology or object model does a maritime and ports AI system need?
  6. Who approves the decisions an AI system makes in maritime and ports?

What does a physical AI system for maritime and ports look like?

A complete physical AI design for maritime and ports names what to sense, which existing systems to join, the object model that joins them, the models and hardware, the three-year cost and the person who approves every action. CodeNinja Atoms has published 4 such reference architectures for maritime and ports, each free to reuse under CC BY 4.0.

DesignCountryWhat it does
Port TwinUnited Statesone authoritative digital twin that binds a port's operational systems, live sensor feeds and finance backbone into a thirteen-object ontology, self-hosted on the port's own virtual machines inside the continental United States.
Terminal PulseUnited Statestruck turn time predicted two hours out and yard congestion seen live, at a container terminal, on the operator's own hardware with no outbound connection.
Port Drone WatchGermanyAn open reference architecture for system design in physical AI, built for a maritime and ports authority in Germany that already flies a docked drone network across two port sites, where owning the inference hardware that runs it costs about 735,000 US dollars over three years. Every flight's imagery becomes detections, and every detection becomes a work item with a named responder role, anchored to a real berth, quay asset or restricted zone in a living map built from the port's own spatial data. Oil sheen, floating debris, persons in restricted zones, fender or bollard damage, smoke or vapour and works progress reach a person who decides, under the European specific category drone rules, the Seveso regime and GDPR.
Industrial Ontology ExamplesSaudi ArabiaOne of the industrial ontology examples in the Vertical-Driven Architectures series: an open reference architecture for system design in physical AI, built for a container terminal operator in Saudi Arabia whose expansion lengthens the quay, deepens the berths and adds ship to shore cranes and a yard block while the terminal keeps running, where owning the inference node that runs it costs about 570,000 US dollars over three years. When berth productivity drops on a working vessel call, the design joins the moves, the fault codes, the yard occupancy, the stowage plan, the granted berth window, the work orders and the commissioning records into one argument, so the operations manager sees which of the four candidate causes it was, the crane's fault, the yard block, the vessel's stowage or the berth window, within the shift. The models run entirely on the operator's own hardware in the country under the port authority's terms, the national data protection law, the port facility security code and export control on the GPU node.

Which AI models can a maritime and ports operator run on its own hardware?

Each published maritime and ports design names its models and why, and every model is open-weight or no model is used at all, so the operator can run it on hardware it owns.

DesignModels
Port TwinOne self-hosted open model, BGE-M3, for document retrieval; no generative model and no token stream
Terminal Pulse5 self-hosted open models: GLM 5.3 (reasoning), Chronos-2 (forecasting), Qwen3-Embedding-0.6B (retrieval), RF-DETR (vision), Roboflow trackers
Port Drone WatchRF-DETR (Apache-2.0) for per-frame detection at the dock; OWLv2 (Apache-2.0) for zero-shot passes on oil sheen, debris, smoke and damage; Ministral 3 8B (Apache-2.0) for evidence triage and work item drafting; Roboflow trackers (Apache-2.0) for motion and association; GLM 5.3 (bespoke licence, 753B at FP8) for the agentic work surface; nothing trained into weights
Industrial Ontology ExamplesGLM 5.3 (GLM-5.3 bespoke licence, 753B at FP8) for the three drafting agents (productivity, equipment, commissioning), the command centre and ontology maintenance; GLM-5.3-Flash (MIT, 320B total, 18B active, multimodal) reading the builders' commissioning drawings and pages on the terminal tier; BGE-M3 (MIT) for retrieval whose sparse leg keeps container numbers and fault codes exact; no third party model service, nothing trained into weights
DesignChoiceWhat was pickedWhy
Port TwinDocument retrieval modelBGE-M3 embedding model, 568M parameters, self-hosted at fp16The only model in the register; hybrid dense and sparse search over drawings, inspections and messages, with no token stream and no external provider in any interactive path
Port TwinHardware classesNone purchased; existing enterprise virtual machines onlyThe requirements buy no equipment; about 1.1 GB of fp16 weights fit standard virtual machine memory
Port TwinSizing ruleWeights occupy about 2.27 GB at float32, about 1.1 GB at fp16, about 0.6 GB at 8 bitThe fp16 figure is the sizing basis; no KV cache applies because an embedding model generates no token stream
Port TwinSensingNo new sensing; up to five existing environmental sensor feeds with threshold alerting and historical export, plus a sourced photorealistic 3D meshThe requirements integrate existing feeds; no camera, sensor or server is purchased
Port TwinGroundPort-managed servers under the port's own change management, inside the continental United States boundaryThe port's own security requirements set the boundary; every workload runs on infrastructure the port already operates
Port TwinPatternAdapter tier, one object model, read-only systems of record, Esri GeoEvent Server as the sole feed pathEvery source enters through an adapter and the twin never writes back into a system of record
Terminal PulseWork surface modelGLM 5.3, 753 billion parameters filed, FP8, on one 8-GPU node of the 141 GB HBM classBespoke license permits internal commercial use with attribution and no revenue trigger; one node holds the weights with KV headroom.
Terminal PulseEdge detectorRF-DETR, Apache-2.0, Nano to 2XL checkpoints from about 61 to 254 MB at 16 bitReal-time on edge GPUs, fine tunable on the operator's own footage, fixed auditable classes.
Terminal PulseSite forecasterChronos-2, Apache-2.0, about 0.48 GB at FP32Zero-shot multivariate forecasting with covariates for the two-hour turn time, queue and reefer predictions.
Terminal PulseSite embedding modelQwen3-Embedding-0.6B, Apache-2.0, about 1.2 GB at bf16The 32K window holds a whole shift's gate and appointment record in one passage.
Terminal PulseEdge trackerRoboflow trackers, Apache-2.0 clean-room SORT, ByteTrack and OC-SORTPer-object counts and conflict geometry at no model memory cost, on CPU beside the detector.
Terminal PulseEdge compute classFanless IP-rated enclosures with accelerator at the yard blocks, gate and quaySized from actual streams and models by bench measurement, never from a datasheet.
Terminal PulseSite inference serverOne GPU node, H100-class 80 GB or L40S-class 48 GBHolds forecaster and embedding model with room for 32K-window activation memory.
Terminal PulseFrontier nodeOne 8-GPU node of the 141 GB HBM class, about 10 kW904 GB required against 1,128 GB usable, inside the room's power envelope.
Terminal PulseCamera estateReuse of the 220 existing fixed cameras, decided per camera by six gatesReuse existing CCTV or not is the first sizing rule; new buys carry no domestic US license restriction.
Terminal PulsePositioning stationRTKLIB with a site-owned GNSS base external positioning service.Grounds RTG and truck positions without an
Terminal PulseTime synchronization OCXO, linuxptp and chronyOCP Time Card grandmaster with holdover reads on one clock through power transfer.Keeps camera frames, PLC cycles and gate
Terminal PulseOne-way transferLidi over a hardware data diodeWeights and images move inward; nothing queries back across the boundary.
Port Drone WatchWork surface modelGLM 5.3, 753B parameters at FP8, bespoke licence, central node in the port's own data centreBespoke licence permits commercial, internal use and redistribution; the command centre role is committed by the requirement; runs on the ontology and memory, not an external service
Port Drone WatchPer-frame closed-set detector modelRF-DETR, Apache-2.0, dock edge tier, about 61 to 68 MB at 16 bit (Nano to Large)Small checkpoints keep pace with the camera stream; the authority holds the weights outright; detection survives an uplink loss
Port Drone WatchZero-shot open-vocabulary detector modelOWLv2, Apache-2.0, site tier, about 310 MB (base) and 875 MB (large) at 16 bitDetects oil sheen, floating debris, smoke or vapour and damage cues on selected frames without port-specific training data
Port Drone WatchEvidence triage and drafting modelMinistral 3 8B at FP8, Apache-2.0, site tier, about 12 GBFits the 48 GB PCIe class with room for concurrent requests; a describer and reader of evidence, never a localiser; European publisher suits an EU public buyer
Port Drone WatchObject tracking modelRoboflow trackers (clean-room SORT, ByteTrack, OC-SORT, BoT-SORT), Apache-2.0, dock edge tierMotion and association only, no re-identification branch, chosen on the port's own labelled clips
Port Drone WatchFrontier hardware141 GB HBM GPU class, eight GPUs per node, one node (1,128 GB)753 GB of FP8 weights times the 1.2 planning factor is 904 GB, inside 1,128 GB with about 375 GB beside the weights
Port Drone WatchSite inference hardware48 GB PCIe inference server class, one to four cards, air-cooled, one server at each of the two port sitesSized by memory class, never by product name; one card holds the site models, more cards buy concurrency
Port Drone WatchEdge AI computeGPU-class edge module at each dock, sized by the dock site surveyRated against salt air, wind and temperature; silicon fixed only at the survey
Port Drone WatchTime synchronisationOCP Time Card grandmaster with linuxptp and chronyFrames, telemetry and detections share one clock across tiers
Port Drone WatchSensingCameras as payloads on the retained drone fleet (patrol nadir, berth survey, incident response; thermal as an annual trial), dock weather station, drone GNSS and telemetry, pipeline health sensorsDetections are geo-located from telemetry against the standardised berth plan
Port Drone WatchPattern: governingSystem of Context, with a work surface where the authority's own people build their agents on memoryPeople decide every response; the loop learns through memory, not weight training
Port Drone WatchPattern: boundaryEvents and evidence crops cross the boundary, not pixelsHours of daily imagery cannot cross the dock uplink, and detection must survive its loss
Industrial Ontology ExamplesReasoning modelGLM 5.3, filed at 753 billion parameters, FP8 (753 GB of weights), GLM-5.3 bespoke licence, served with vLLM on one node of eight 141 GB GPUs in the operator's own machine room in Saudi ArabiaNamed by the operator; the licence permits commercial use and redistribution with attribution, and its model as a service trigger does not touch a terminal operator's internal use; carries the three drafting agents, the command centre and ontology maintenance
Industrial Ontology ExamplesDocument reading modelGLM-5.3-Flash, 320 billion total parameters with 18 billion active, natively multimodal, MIT, on the terminal tier (a two socket air cooled server class in the terminal machine room)MIT terms let the operator hold the weights outright; it reads the builders' scanned commissioning drawings and pages at the rate the expansion produces them and keeps load off the frontier node
Industrial Ontology ExamplesEmbedding modelBGE-M3, MIT, FP16, about 1.1 GB, dense, sparse and multi vector retrieval over 8,192 tokens, on the central nodeThe sparse leg keeps container numbers and fault codes exactly retrievable; embeddings for the conversational interface and ontology retrieval
Industrial Ontology ExamplesFrontier hardware141 GB HBM GPU class, eight GPUs per node, one node (1,128 GB)Holds the 753 GB of FP8 weights with the 1.2 planning factor (904 GB) and 375 GB beside the weights; the class needs an export licence determination before order, run at phase one start with the importer of record named
Industrial Ontology ExamplesFrontier footprint sizing ruleFrontier Model GPU Footprint Sizing Rule for vLLM and SGLangOne byte per FP8 parameter times the 1.2 factor for KV cache and activations fixes the node count at one
Industrial Ontology ExamplesServing runtimevLLMServes the FP8 weights on the frontier node within the KV cache ceiling the sizing reserves
Industrial Ontology ExamplesOne-way transferLidi over a hardware data diodeControl zone egress ends in hardware; fault logs and machine state leave the control zone and nothing can return
Industrial Ontology ExamplesThermal camera legalityAny US sourced thermal unit held at 9 Hz or lessStays under the export control line (ECCN 6A003.b.4.b) that would otherwise require a licence; belongs to the offered crane tier only
Industrial Ontology ExamplesEdge detector servingOpenVINO on fanless, sealed, sunshielded edge compute rated for quayside heatOffered continuous crane sensing tier only, not committed; a development board measurement is the gate before any unit is fixed
Industrial Ontology ExamplesSensing (offered tier)Cameras on hoist machinery, the quay crane rail and a yard block transfer zone, detecting spreader, container, twistlock, crane boom, straddle carrier and terminal tractorBelongs to the offered crane tier only; existing telematics is the committed signal
Industrial Ontology ExamplesTelemetry signalsShip to shore crane telematics (hoist position, drive load), crane fault logs, terminal operating system move events and the port authority berth window feedThe drift and productivity arguments run on records the operator already holds
Industrial Ontology ExamplesEvent backboneApache Kafka 4.3.x in KRaft modeOrdering per key (moves by vessel call, faults by crane), at least once delivery made safe by idempotent upserts, buffer and replay when a link fails
Industrial Ontology ExamplesTime synchronisationChrony with linuxptp under an OCP Time Card grandmasterOperating system events, control system fault logs and telematics join on one clock
Industrial Ontology ExamplesRegistryHarbor, air gapped artifact and model registryModel artifacts are staged, verified and applied on the operator's schedule, never pulled live from outside
Industrial Ontology ExamplesPattern: ontologySystem of context, with Hyper Ontology v0.9 Semantics and KineticsThe ontology is a projection over the systems of record, never a shadow copy of them
Industrial Ontology ExamplesPattern: memoryA work surface where the operator's people build their own agents, with a kept record of decisions and outcomes that the next decision readsAgents draft inside the boundary and memory of confirmed outcomes improves the next trace without training
Industrial Ontology ExamplesPattern: boundaryThe operator's own hardware, in country, on the terminal's own network; the port authority feed the one inbound external linkCargo, vessel and equipment data never leave the country and no third party model service is called

How much compute and hardware does AI in maritime and ports need?

The compute follows from the models: the published maritime and ports designs size it as follows, from no new hardware to a full GPU node.

DesignPartThe design
Port TwinComputeNo equipment bought: about 1.1 GB of fp16 weights on the port's existing enterprise virtual machines
Port TwinBoundaryEverything runs inside the continental United States on infrastructure the port already operates; no hosted document AI service in the path
Terminal PulseFrontier computeOne 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
Terminal PulseEdgeFanless IP-rated enclosures at the yard blocks, gate and quay; no safety reflex crosses a network hop
Port Drone WatchComputeThree 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
Industrial Ontology ExamplesComputeOne node of eight 141 GB HBM cards (1,128 GB) in the operator's own machine room in the Kingdom, holding 753 GB of FP8 weights with a 1.2 planning factor, 904 GB, 375 GB beside the weights; a two socket air cooled server class for the terminal tier; no edge tier committed (continuous crane sensing is offered and priced separately). The node is a 4A090 class item for a D:3 and D:4 destination, so an export licence determination naming the importer of record is a phase one start item

Is it cheaper to own AI hardware or rent cloud GPUs in maritime and ports?

Each maritime and ports design prices three years of ownership in its Appendix A, with every price cited, against renting the same capacity from a cloud region at its deepest three-year commitment where hardware is bought.

DesignLineThree years
Port TwinThree-year costNo hardware line to price: the design adds software and integration work to servers the port already runs
Terminal PulseThree-year cost, ownedAbout 722,000 US dollars with support and power at the US industrial power price
Terminal PulseThree-year cost, rented0.99 million to 2.52 million US dollars for the same GPUs around the clock; ownership is about four fifths the deepest three-year commitment (version 2)
Terminal PulseClosed model break-evenThe cheapest closed model matches the owned stack at about 35 users; above that, ownership is cheaper and the gap grows with every user
Port Drone WatchThree-year costAbout 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)
Industrial Ontology ExamplesThree-year costAbout 570,000 US dollars to own the node for three years (a range of 542,000 to 594,000: server, support and power at the national commercial electricity tariff), against about 751,000 to rent the same eight cards on the deepest three year commitment in AWS Middle East (UAE), so ownership costs about three quarters of it, and about 2.00 million on demand, against which ownership costs a little over a quarter. Renting is a price reference only, because AWS listed no Saudi region on the day read and every rented option would carry cargo, vessel and equipment data across the border (Appendix A)

What ontology or object model does a maritime and ports AI system need?

The object model is the part that makes the system an ontology rather than a pipeline: typed objects for the things in the maritime and ports operation, with properties, status values and typed links. Every published design ships its object model as hyper-ontology/1 JSON that loads into Hyper Ontology.

DesignSizeObjectsDownload
Port Twin13 objects, 13 linksBerth, Navigation channel, Subsurface utility layer, Environmental sensor feed, Parcel / facility, Bathymetric survey surface, Utility gap record, Capital project, Inspection record, Vessel movement record, Truck movement record, PCS message, Engineering documentobjects.json
Terminal Pulse12 objects, 11 linksVessel Call, Container, Yard Block, RTG, Ship to Shore Crane, Truck Visit, Gate Lane, Rail Cut, Reefer Plug, Yard Person, Transfer Zone, Safety Eventobjects.json
Port Drone Watch15 objects, 14 linksMission, Construction Work, Detection, Work Item, Flight Authorisation, Coverage Statement, Berth, Dock Station, Restricted Zone, Vessel, Quay Asset, Drone, Responder Role, Drone Operator, Weather Observationobjects.json
Industrial Ontology Examples16 objects, 13 linksBerth, Quay section, Yard block, Ship to shore crane, Yard crane, Straddle carrier, Equipment component, Vessel call, Berth window, Fault, Move, Incident, Work order, Commissioning record, Box, Productivity measureobjects.json

Who approves the decisions an AI system makes in maritime and ports?

In every published maritime and ports design a named person makes the decision that changes the physical world; the system prepares it.

DesignHuman control
Port TwinThe twin shows; pilots, berth planners, engineers and finance staff decide in their own systems, and nothing writes back into a system of record
Terminal PulseSurfaces warn and propose; the named planner acts in the terminal operating system and the named safety supervisor acknowledges every safety event
Port Drone WatchThe system proposes and records and people decide every response: a detection stays proposed until a responder confirms, rejects or merges it, every work item is acknowledged with reason codes by its named responder role, flight approvers grant every mission, and automatic acts are limited to confirmation re-flights and notification escalation
Industrial Ontology ExamplesAgents draft and people decide, by name: the operations manager confirms the productivity agent's ranked causes, the engineering manager approves the equipment agent's work order and berth plan effect, and the expansion project engineer approves the commissioning agent's objects; approval latency is instrumented and known answer probes sit in the draft queue from day one, and every confirmation is kept as the memory the next trace reads

Every answer on this page is drawn from the papers linked in it: each paper's At a glance table, model register and cost appendix. Full text for agents: llms-full.txt. Designed on Praxis; object models load into Hyper Ontology.

Want one of these designs for your own operation? Talk to us.