Sector · Mining
Physical AI for mining
These are the questions engineers ask first, answered from 1 published CodeNinja Atoms reference architecture.
- What does a physical AI system for mining look like?
- Which AI models can a mining operator run on its own hardware?
- How much compute and hardware does AI in mining need?
- Is it cheaper to own AI hardware or rent cloud GPUs in mining?
- What ontology or object model does a mining AI system need?
- Who approves the decisions an AI system makes in mining?
What does a physical AI system for mining look like?
A complete physical AI design for mining 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 1 such reference architecture for mining, each free to reuse under CC BY 4.0.
| Design | Country | What it does |
|---|---|---|
| Industrial AI for an Open Pit Mine Haul Fleet in Saudi Arabia | Saudi Arabia | An open reference architecture for system design in physical AI, built for an open pit mine operator in Saudi Arabia whose pit is worked by a contractor with a haul fleet of about 35 trucks of the 100 tonne class and one shovel of the 600 tonne class, where owning the reasoning node that runs it costs about 556,000 US dollars over three years. Every alarm, oil result, tyre reading, payload record and rebuild is joined on one governed fleet ontology, so the maintenance planner can ask which component on which truck fails next and catch a final drive in the workshop rather than on the ramp. The models run entirely on the operator's own hardware on site, under the Mining Investment Law and its safety regulations, the national cybersecurity controls for operational technology, the Personal Data Protection Law and export control on the GPU node. |
Which AI models can a mining operator run on its own hardware?
Each published mining 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.
| Design | Models |
|---|---|
| Industrial AI for an Open Pit Mine Haul Fleet in Saudi Arabia | GLM 5.3 (bespoke licence, 753B at FP8) for reasoning, the conversational interface and the agents' drafts; Grounding DINO Swin-T (Apache-2.0, about 345 MB at 16 bit) for shovel bucket tooth detection at the edge; Chronos-2 (Apache-2.0) for zero shot telemetry and oil trend forecasting with covariates; BGE-M3 (MIT) for evidence retrieval so every answer cites its record; nothing trained into weights |
| Design | Choice | What was picked | Why |
|---|---|---|---|
| Industrial AI for an Open Pit Mine Haul Fleet in Saudi Arabia | Reasoning model | GLM 5.3, 753B parameters filed, FP8, GLM-5.3 bespoke license, served on vLLM on one node of eight 141 GB GPUs in the site server room | Bespoke license permits commercial use, fine tuning and redistribution, so the operator holds the weights; carries reasoning, the conversational interface and the drafts of the health, tooth loss and dispatch agents |
| Industrial AI for an Open Pit Mine Haul Fleet in Saudi Arabia | Edge vision detector model | Grounding DINO, Swin-T open weights checkpoint, about 345 MB at 16 bit, Apache-2.0, served under ONNX Runtime on K3s on the automotive grade SoC at the bucket camera | Apache-2.0 lets the operator hold and serve the checkpoint fanless at the bucket; detects the bucket, its teeth and the ground engaging tools per frame; Swin-B at about 467 MB at 16 bit is the fallback if evaluation on representative footage demands it |
| Industrial AI for an Open Pit Mine Haul Fleet in Saudi Arabia | Forecasting model | Chronos-2, zero-shot with covariates, about 0.48 GB at 32 bit as published and about 0.24 GB at 16 bit, Apache-2.0, on the site servers | Forecasts telemetry and oil trends and thresholds the residual for anomaly with no tuning; chosen over TimesFM 3.0 and Moirai 2.0, whose non-commercial terms cannot ship inside the boundary, and over Chronos-Bolt, because covariates and co-evolving channels matter |
| Industrial AI for an Open Pit Mine Haul Fleet in Saudi Arabia | Embedding model | BGE-M3, MIT, 8,192 token window, about 1.1 GB at fp16, on the site servers | Dense, sparse and multi-vector retrieval in one pass keeps part numbers and alarm codes retrievable across Arabic and English records, so every answer cites its evidence |
| Industrial AI for an Open Pit Mine Haul Fleet in Saudi Arabia | Edge compute | Automotive grade octa core SoC variant, heat rated for cab and shovel mounts, fanless enclosure | Carries the bucket watch model where a GPU module would need a heat path a sealed enclosure cannot give; the conflicting temperature datasheets are resolved as a commissioning check |
| Industrial AI for an Open Pit Mine Haul Fleet in Saudi Arabia | Site inference servers | Rack servers in the site server room, sized for the desert environment, fed from the internal Harbor registry with no egress | Join the evidence and run the forecasting and embedding tiers beside the ontology |
| Industrial AI for an Open Pit Mine Haul Fleet in Saudi Arabia | GPU node | One node of eight 141 GB GPUs (1,128 GB), roughly ten kilowatts, in the site server room | Holds GLM 5.3 at FP8 for reasoning and the work surface |
| Industrial AI for an Open Pit Mine Haul Fleet in Saudi Arabia | Frontier footprint sizing rule | 753 GB of FP8 weights times the 1.2 planning factor | 904 GB against the node's 1,128 GB; 375 GB beside the weights against the 151 GB the planning factor reserves |
| Industrial AI for an Open Pit Mine Haul Fleet in Saudi Arabia | Sensing | Bucket, ramp and rebuild bay cameras; vibration and motor current sensing on critical rotating components; oil findings; tyre readings; payload records; telematics alarms | The kinds of evidence that never meet today, joined in one ontology with every fact tied to its source |
| Industrial AI for an Open Pit Mine Haul Fleet in Saudi Arabia | Pattern: governing | Safety loop on the machine and optimization loop in the center; memory of work order outcomes, never weights; write back only through the owning systems; local buffering with fail safe on link loss | Each pattern is a mining safety or sovereignty rule, not a preference; nothing is trained into weights |
| Industrial AI for an Open Pit Mine Haul Fleet in Saudi Arabia | Pattern: boundary | The operator's own hardware on site in Saudi Arabia, an air gapped Harbor registry, updates inbound only through a hardware enforced data diode running Lidi | The pit keeps working when the link to the city fails, and nothing leaves |
How much compute and hardware does AI in mining need?
The compute follows from the models: the published mining designs size it as follows, from no new hardware to a full GPU node.
| Design | Part | The design |
|---|---|---|
| Industrial AI for an Open Pit Mine Haul Fleet in Saudi Arabia | Compute | One node of eight 141 GB cards (1,128 GB) in the site server room, holding 753 GB of FP8 weights with a 1.2 planning factor, 904 GB, at roughly ten kilowatts; rack servers beside it for forecasting and embedding; heat rated automotive grade SoC edge units, fanless, on the trucks and at the shovel bucket camera. The node is a 4A090.a class item for a D:3 and D:4 destination, so a BIS licence determination belongs in phase one (Appendix A.3). Nothing leaves; updates arrive one way through a hardware enforced data diode |
Is it cheaper to own AI hardware or rent cloud GPUs in mining?
Each mining 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.
| Design | Line | Three years |
|---|---|---|
| Industrial AI for an Open Pit Mine Haul Fleet in Saudi Arabia | Three-year cost | About 556,000 US dollars to own the reasoning node for three years (a range of 528,000 to 580,000: server, support and power at the Saudi industrial electricity tariff), against about 751,000 to rent the same eight cards on the cheapest three year commitment in AWS Middle East (UAE), an EC2 Instance Savings Plan, all upfront, so ownership costs about three quarters of it. Renting is a price reference only, because AWS lists no Saudi region on the day read and every rented option would carry fleet records across the border (Appendix A) |
What ontology or object model does a mining 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 mining 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.
| Design | Size | Objects | Download |
|---|---|---|---|
| Industrial AI for an Open Pit Mine Haul Fleet in Saudi Arabia | 16 objects, 13 links | Haul truck, Shovel, Component, Haul road segment, Haul cycle, Payload record, Telematics alarm, Oil sample, Tyre reading, Work order, Rebuild, Ground engaging tool, Shift, Maintenance planner, Dispatcher, Mine superintendent | objects.json |
Who approves the decisions an AI system makes in mining?
In every published mining design a named person makes the decision that changes the physical world; the system prepares it.
| Design | Human control |
|---|---|
| Industrial AI for an Open Pit Mine Haul Fleet in Saudi Arabia | Agents draft and named people decide: the maintenance planner approves each health ranking, the dispatcher approves each haul redraft, the shovel operator confirms each tooth loss, and the mine superintendent owns the command center and holds escalation authority; no agent stops a machine or changes a work order without a named approval, except the tooth loss stop, which follows the site safety procedure |
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.