Sector · Energy and utilities
Physical AI for energy and utilities
Grids and plants run on control systems, maintenance records and field crews that were never designed to share one picture of risk. These are the questions engineers ask first, answered from 8 published CodeNinja Atoms reference architectures.
- What does a physical AI system for energy and utilities look like?
- Which AI models can an energy and utilities operator run on its own hardware?
- How much compute and hardware does AI in energy and utilities need?
- Is it cheaper to own AI hardware or rent cloud GPUs in energy and utilities?
- What ontology or object model does an energy and utilities AI system need?
- Who approves the decisions an AI system makes in energy and utilities?
What does a physical AI system for energy and utilities look like?
A complete physical AI design for energy and utilities 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 8 such reference architectures for energy and utilities, each free to reuse under CC BY 4.0.
| Design | Country | What it does |
|---|---|---|
| Grid Context Watch | United States | one system of context that sees every device on a utility's control networks, flags abnormal behaviour and assembles the evidence a NERC CIP audit asks for, on the operator's own hardware. |
| Reliability Atlas | Saudi Arabia | a records-based reliability and availability assessment that joins each plant's work orders, trip history and design documents into one reliability object model the operator owns, so every availability figure, criticality rank and root cause finding traces to its source. |
| Feeder Firewatch | United States | a live ignition and outage risk score for every distribution feeder segment, built on the cooperative's own hardware, with every de-energisation and fast-trip change approved by a named operator. |
| Turbine Island Watch | Pakistan | An open reference architecture for system design in physical AI. A condition monitoring and diagnostics platform for a rehabilitated 280 MW hydrogen-cooled steam turbine generator at a combined cycle plant in Pakistan: it reads the DCS, the vibration monitoring rack and the generator protection systems through a one-way path, ranks anomalies into prescribed actions for the shift operator, and runs entirely on the operator's own hardware. |
| Sovereign AI Architecture for a Grid Control Room in Saudi Arabia | Saudi Arabia | An open reference architecture for system design in physical AI, built for a national grid operator in Saudi Arabia that runs the electricity transmission grid from regional and national control centers, where owning the inference node that runs it costs about 570,000 US dollars over three years. When one 380 kilovolt line trips at summer peak, the design joins the alarms, the network model, the outage plan, the protection record and the procedure library into one argument, so the control engineer sees which protection operated, what it affects, what work is already in progress nearby and which procedure governs the next step. The model runs entirely on the operator's own hardware under the national cybersecurity controls for operational technology, the grid code and export control on the GPU node. |
| Industrial Data Ontology for a Water Utility in the United States | United States | An open reference architecture for system design in physical AI, built for a municipal water utility in the United States that runs a treatment plant, wells, a distribution reservoir, pump and lift stations and a pipe network under PLCs and RTUs, in the middle of a control system migration, where owning the inference server that runs it costs about 59,000 US dollars over three years. Every alarm tag is tied to the pump it measures, the station and pressure zone the pump serves, the work orders on it and the sampling points and permit limits on its zone, so the operator on shift can ask, the night a duty pump trips, which asset tripped, what work is open on it and what the permit requires of the zone it feeds. The model runs entirely on the utility's own hardware, reads the segregated control network only through a hardware data diode, and answers to the drinking water and discharge permit regime and the obligations of a state and local critical infrastructure owner. |
| Sovereign AI Architecture for Distribution Loss and Theft Reduction in Pakistan | Pakistan | An open reference architecture for system design in physical AI, built for an electricity distribution company in Pakistan that serves several million customers through hundreds of grid stations and thousands of feeders, where owning the inference hardware that runs it costs about 641,000 US dollars over three years. When a transformer's monthly energy balance crosses its threshold, the design joins the smart meter interval load, the tamper events, the billing history, the network topology and the inspection record into one argument, so the subdivision engineer sees the ranked causes, the inspection officer opens a case file already populated with the load history and the billing record, and the revenue officer approves the correction. The models run entirely on the operator's own hardware in the country under the regulator's performance standards, the national data protection rules, the electricity theft laws and export control on the GPU node. |
| A Self-Improving AI System for OT Vulnerability Management at a Utility in the United States | United States | An open reference architecture for system design in physical AI, built for an electric utility in the United States whose substations, control centers and generating plants fall under the critical infrastructure protection standards, where owning the inference node that runs it costs about 477,000 US dollars over three years. When a vendor publishes a relay firmware advisory on day thirty of a cycle, the design joins the device inventory, the verified baseline, the advisory, the national vulnerability record and the decisions made the last time this vendor's advisories touched these devices into one argument, so the compliance engineer sees which devices it touches and why, the asset owner approves apply or mitigate with its outage window and change request, and the auditor finds the evidence package where audits already look. The model runs entirely on the utility's own hardware, on open weights of United States origin, with nothing about its assets leaving its network. |
Which AI models can an energy and utilities operator run on its own hardware?
Each published energy and utilities 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 |
|---|---|
| Grid Context Watch | GLM 5.2 (MIT) for reasoning and Qwen3-Embedding-0.6B (Apache-2.0) for retrieval; detection stays in the procured sensors |
| Reliability Atlas | None: the register is empty by design because no inference workload is contracted |
| Feeder Firewatch | Self-hosted open-weight models: GLM 5.2 under MIT (reasoning) on site, RF-DETR (vision) at the edge |
| Turbine Island Watch | MOMENT-1 (MIT) for zero-shot anomaly detection over vibration, hydrogen, winding and process telemetry; IBM Granite 4.2 8B (Apache-2.0) for diagnostic notes and test-report drafts; no training run |
| Sovereign AI Architecture for a Grid Control Room in Saudi Arabia | GLM 5.3 (GLM-5.3 bespoke licence, 753B at FP8) for all four drafting agents (alarm triage, switching programs, handover and the sequence of events), the command center and ontology maintenance; BGE-M3 (MIT) for Arabic and English retrieval so every answer cites its procedure or record; no third party model service, nothing trained into weights |
| Industrial Data Ontology for a Water Utility in the United States | gpt-oss-120b (Apache-2.0, 117B, MXFP4) for the agentic work surface: the tag mapping agent, the alarm context agent, the monthly compliance drafting agent and cited answers; Nomic Embed Text v2 MoE (Apache-2.0) for retrieval so every answer cites its record; United States origin weights for a state and local critical infrastructure buyer; no third party model service, nothing trained into weights |
| Sovereign AI Architecture for Distribution Loss and Theft Reduction in Pakistan | GLM 5.3 (GLM-5.3 bespoke licence, 753B at FP8) for the loss, tamper and billing agents, the command center and ontology maintenance; BGE-M3 (MIT) for cited retrieval over tariff codes, meter numbers, PPC sections and inspection findings; Chronos-2 (Apache 2.0) forecasting each transformer's energy balance so the residual is the anomaly signal; no third party model service, nothing trained into weights |
| A Self-Improving AI System for OT Vulnerability Management at a Utility in the United States | gpt-oss-120b (Apache-2.0, about 120B parameters in MXFP4, United States origin) for the applicability, disposition and evidence agents, the command center's cited conversational surface and ontology maintenance; Nomic Embed Text v2 MoE (Apache-2.0, 475M parameters at FP16) embedding advisories, baselines and confirmed decisions for retrieval memory; no third party model service, nothing trained on the fly |
| Design | Choice | What was picked | Why |
|---|---|---|---|
| Grid Context Watch | Reasoning model | GLM 5.2, 753B parameters, frontier class, FP8, 1M context | Plain MIT license with no revenue trigger or security-review clause; one node of eight 141 GB HBM GPUs holds it; the work surface's role demands long-context agentic reasoning |
| Grid Context Watch | Embedding model | Qwen3-Embedding-0.6B, Apache-2.0, 32K window | Whole playbooks and CIP procedures retrieve as single passages; runs beside the frontier model for about 1.2 GB at BF16 |
| Grid Context Watch | Sizing rule | 1.2 planning factor over FP8 weights | 753 GB of weights times 1.2 is 904 GB against 1,128 GB of node memory, leaving governed headroom for KV cache and activations |
| Grid Context Watch | Frontier compute | One node of eight 141 GB HBM GPUs | The smallest class that holds the FP8 weights; sized from the model's filed parameters |
| Grid Context Watch | Edge compute | Fanless, extended-temperature, cabinet-mount class | Compute stays out of electronic security perimeters wherever topology allows |
| Grid Context Watch | Enclosure and power | NEMA 3R/4X class cabinets with uninterruptible power supplies under Network UPS Tools | Clean shutdown ordering protects capture integrity environment |
| Grid Context Watch | Time synchronization | OCP Time Card GNSS grandmaster with holdover, Linuxptp and chrony | Pcap, sensor and log clocks must agree for evidence to correlate |
| Grid Context Watch | One-way transfer | Hardware data diodes with the one-way transfer appliance at highest-impact perimeters | Demonstrable one-way flow where policy demands it; constrained DMZ conduit elsewhere |
| Grid Context Watch | Site networking | Hardened, fanless, dual-DC industrial switches with SPAN/TAP ports | Segmentation remains the operator's policy decision; the design reads a copy, never the control path |
| Grid Context Watch | Sensing | Passive network traffic capture via the procured OT monitoring sensors at SPAN/TAP sources | The scope watches OT traffic, not premises; no cameras or video models |
| Grid Context Watch | Primary pattern | System of Context over one Hyper-style object model of 14 objects | The lasting asset is the living model of the OT estate, not any single alert |
| Grid Context Watch | Ground | The operator's own on-premises OT hardware and data center | Ownership of weights, ontology, decision record and boundary stays with the operator; identity and update paths run inside the air gap |
| Reliability Atlas | Models | None: the register is empty by design | No inference workload is contracted; the deliverable is the study, and an empty register recorded plainly outranks a speculative entry the scope cannot justify |
| Reliability Atlas | Hardware classes and sizing rules | None procured; the study runs on the operator's own in-country servers, storage and data center capacity | No field hardware, no edge compute and no serving runtime is installed under this scope |
| Reliability Atlas | Sensing | No cameras, no sensors and no positioning; site verification is by engineer walkdowns | The world is read through records, so nothing is watched and nothing needs country legality stated |
| Reliability Atlas | The pattern it stands on | System of context as the primary pattern, with the reliability object model as a projection over the three systems of record | Unavailability ranking, sparing confirmation and RCA validation are cross-silo joins that no single source system can produce alone |
| Reliability Atlas | The ground it runs on | In-country hosting in Saudi Arabia, with identity through the operator's own directory and single sign-on, procured by class | The operator's requirement keeps data inside the country's boundary and under the operator's identity, which the operator already operates |
| Feeder Firewatch | Language model parameters filed, FP8, MIT license | GLM 5.2, mixture-of-experts, 753 billion weights the cooperative owns; one 8-GPU | Agentic reasoning over the ontology with FP8 node holds it |
| Feeder Firewatch | Detection model to 68 MB and 2XL at about 254 MB, BF16, Apache-2.0 | RF-DETR, Nano to Large checkpoints at 61 detection on the edge box class, fine-tuned on the cooperative's own fire-season frames | Smoke, pole, vegetation and equipment |
| Feeder Firewatch | Edge compute class and sealed, NEMA 3R/4 enclosures | Industrial edge accelerator modules, fanless streams, rated for dust, heat, hail and cold | Sized decode-first from the actual camera |
| Feeder Firewatch | Site inference server center, one node of 8 GPUs of the 141 GB HBM class | Ruggedised server class at the operations sized from filed parameters at FP8 | 1,128 GB against a 904 GB requirement, |
| Feeder Firewatch | Serving runtimes Runtime or OpenVINO class at the edge | vLLM on site; Triton Inference Server, ONNX a bench measurement of the real streams | The edge runtime is pinned at design against |
| Feeder Firewatch | Sizing rules thermal beats visible; reuse existing cameras or not; size GPUs from filed parameters | Size edge compute from streams; when measured duty rather than a vendor default | Every hardware choice derives from a |
| Feeder Firewatch | Sensing camera feeds, truck and drone cameras, mesonet wind and humidity, 310 recloser fault indicators, AMI last gasp from 61,000 meters, crew AVL on 20 trucks | 42 substation PTZ cameras, 12 wildfire thermal only where a coverage gap is proven | Reused through the reuse gates; new fixed |
| Feeder Firewatch | Patterns read-only systems of record; human-approved write-back; one-way diode for SCADA | System of context; adapters-only ingestion; record authoritative while the platform joins them | Keeps the grid protected and the systems of |
| Feeder Firewatch | Ground as primary; sovereign cloud for overflow and recovery only | Cooperative servers at the operations center during a storm, and the risk picture cannot live outside the boundary | The model must survive an internet outage |
| Turbine Island Watch | Model | MOMENT-1, a time series foundation model for anomaly detection, MIT | Runs zero-shot on the 48 GB-class site card over vibration, hydrogen, winding and process telemetry, with no training run, because the rehabilitated unit's history may be too sparse to baseline |
| Turbine Island Watch | Model | IBM Granite 4.2 8B, Apache-2.0 | Drafts diagnostic notes and test reports for engineer review; licence-clean with an enterprise vendor behind it and real KV headroom on one site card, chosen over Qwen3.8 at 27B |
| Turbine Island Watch | Hardware class | 48 GB PCIe inference GPU class on ruggedised edge servers inside the plant perimeter | Pakistan has no hyperscaler region, so every tier runs on the operator's own hardware |
| Turbine Island Watch | Tool: event streaming and messaging | Apache Kafka 4.3 (KRaft, three dedicated controllers, dynamic quorum), Apache-2.0 | Condition streams need a self-managed backbone inside the plant that no managed service can serve from Pakistan. |
| Turbine Island Watch | Tool: time series and telemetry storage | InfluxDB 3 Core, MIT | Thousands of condition channels need a self-hosted OT historian whose retention the five-year NEPRA log rule stands on. |
| Turbine Island Watch | Tool: edge model serving | ONNX Runtime, MIT | The anomaly model runs exported to ONNX at the site tier so serving survives without internet pulls. |
| Turbine Island Watch | Tool: on site language model serving | vLLM, Apache-2.0 | Granite 4.2 8B is served on the site inference card with the KV headroom vLLM's memory discipline gives. |
| Turbine Island Watch | Tool: edge orchestration of containers | Red Hat OpenShift AI 3.5 on OpenShift Container Platform 4.20, Red Hat subscription | The primary architecture's operator-managed, mirrored-install platform delivers serving, pipelines and the registry coherently to a plant with no internet route. |
| Turbine Island Watch | Tool: an air gapped artifact and model registry | MLflow model registry with the mirrored private registry as the artifact supply chain, Apache-2.0 | Every model version entering the plant crosses the mirror once and carries provenance into the NEPRA log trail. |
| Turbine Island Watch | Tool: labelling and retraining | Label Studio, Apache-2.0 | Engineer-confirmed vibration events become labelled training evidence for the anomaly model without leaving the perimeter. |
| Turbine Island Watch | Tool: time synchronisation across sensors and systems | linuxptp with PTP-aware switches, GPL-2.0 | The specification requires accurate stamping to a common time reference for sequence-of-events logs, which only a designed PTP spine delivers. |
| Turbine Island Watch | Tool: observability of the running system | Prometheus, Grafana and Loki, Apache-2.0 / AGPL-3.0 (Loki) | The 2003 blackout lesson demands a positive heartbeat and freshness indication for the monitoring layer itself. |
| Turbine Island Watch | Tool: identity and access on site | Keycloak, Apache-2.0 | NEPRA least-privilege and approved-remote-access duties need an on-site identity provider without credentials inside the control zone. |
| Turbine Island Watch | Tool: integration with the sector's own operational systems | open62541 OPC UA client, MPL-2.0 | The read contract into the DCS, vibration rack and detection systems is OPC UA subscriptions with bounded queues, run from the DMZ side. |
| Turbine Island Watch | Tool: time synchronisation hardware | OCP Time Card (PTP grandmaster with GNSS and holdover), MIT (repository; third-party FPGA cores LGPL-3.0) | One open grandmaster card disciplines the whole estate for sequence-of-events ordering, with a second unit for failover. |
| Turbine Island Watch | Tool: one way data transfer | lidi beside a hardware data diode, LGPL-3.0 | The platform reads the DCS through a physically one-way egress path so no inbound route into the control network exists. |
| Sovereign AI Architecture for a Grid Control Room in Saudi Arabia | Reasoning model | GLM 5.3, filed at 753B parameters, FP8 (753 GB of weights), GLM-5.3 bespoke licence, served on vLLM on one node in the operator's own data centre in the Kingdom | Licence allows commercial use, fine tuning and redistribution with an attribution notice and a managed access trigger above 10 billion US dollars of group revenue that internal control room use does not reach; the register's one frontier class model; carries all four drafting agents, the command centre and ontology maintenance |
| Sovereign AI Architecture for a Grid Control Room in Saudi Arabia | Embedding model | BGE-M3, MIT, FP16, about 1.1 GB, on the platform zone's dual socket x86 rack server class | MIT licence with no obligation; multilingual Arabic and English retrieval over procedures, switching instructions, records and past event analyses; recall on the operator's own Arabic procedures proved before go-live |
| Sovereign AI Architecture for a Grid Control Room in Saudi Arabia | Frontier hardware | 141 GB HBM GPU class, eight GPUs per node, one node (1,128 GB) | Holds GLM 5.3's 753 GB of FP8 weights on one node; held at one node so the export licence determination naming the end user is asked once, early, as a phase one checkpoint before the order |
| Sovereign AI Architecture for a Grid Control Room in Saudi Arabia | Frontier footprint sizing rule | Frontier model GPU footprint sizing rule, vLLM and SGLang | 753 GB of weights times the 1.2 planning factor is 904 GB against the node's 1,128 GB, leaving 375 GB beside the weights for KV cache and activations |
| Sovereign AI Architecture for a Grid Control Room in Saudi Arabia | Node count rule | GLM 5.3 node count by GPU class | One node of eight 141 GB GPUs; two nodes of the 80 GB class rejected because they add an RDMA fabric |
| Sovereign AI Architecture for a Grid Control Room in Saudi Arabia | One-way transfer | Data diode class with forward error correction | Committed fallback where the segmentation review demands demonstrable one-way flow; replication through a historian or DMZ feed is the default |
| Sovereign AI Architecture for a Grid Control Room in Saudi Arabia | Serving runtime | vLLM | Serves the FP8 weights within the KV cache ceiling the sizing reserves |
| Sovereign AI Architecture for a Grid Control Room in Saudi Arabia | Registry | Harbor, air gapped artifact and model registry | Updates are staged and signed inside the boundary; the platform keeps serving the last verified version if the update path fails |
| Sovereign AI Architecture for a Grid Control Room in Saudi Arabia | Sensing | Replicated SCADA and EMS streams, protection trip records, heartbeat and freshness signals, citation resolution checks | No cameras and no new field hardware; the platform watches its own data freshness and checks the citations on every answer |
| Sovereign AI Architecture for a Grid Control Room in Saudi Arabia | Pattern: memory | Memory loops over confirmed decisions, nothing trained into weights | The requirement asks that memory improve every shift without training a model, and a moving model would invalidate the frozen baselines |
| Sovereign AI Architecture for a Grid Control Room in Saudi Arabia | Pattern: validation | Replay of the operators' own recorded alarm floods | Trust is built on the control room's own history before summer go-live |
| Sovereign AI Architecture for a Grid Control Room in Saudi Arabia | Pattern: topology | CIM import from the network model seeds the ontology | Topology comes from the model of record rather than assembled from documents |
| Sovereign AI Architecture for a Grid Control Room in Saudi Arabia | Pattern: boundary | The operator's own hardware in its own data centre in Saudi Arabia; one way inbound, read only, no write path into SCADA or EMS | National cybersecurity segmentation controls and the operator's own rule that the operational network stays separated |
| Industrial Data Ontology for a Water Utility in the United States | Work surface model | gpt-oss-120b, 117B parameters filed, mixture of experts in MXFP4 at 4.25 bits with the rest at 16 bit (about 61 GB of weights), Apache-2.0, served by vLLM on the inference server in the utility's own server room | United States origin open weights for a state and local critical infrastructure buyer; the GLM 5.3 class is set aside because its maker sits on the US Entity List; Apache-2.0 carries no field of use restriction, so the utility holds the weights outright; carries the tag mapping agent, the alarm context agent, the monthly compliance drafting agent and the cited conversational interface |
| Industrial Data Ontology for a Water Utility in the United States | Embedding model | Nomic Embed Text v2 MoE, 475 million parameters, 768 dimensions, 8K window, Apache-2.0, about 0.95 GB at fp16, on the same node | United States publisher with published data provenance, chosen over BGE-M3 for origin; feeds the hybrid dense plus lexical index the citations depend on; Matryoshka truncation keeps the on-site index small enough for the utility's own hardware |
| Industrial Data Ontology for a Water Utility in the United States | Inference hardware | 48 GB PCIe GPU class, two cards in one air-cooled node, 96 GB total, about 88 GB usable | The 70.1 GB sized requirement fits with room for the key-value cache, and the class air-cools in a standard server room; frontier class nodes are set aside because 117B parameters sit below the 300B threshold |
| Industrial Data Ontology for a Water Utility in the United States | Footprint sizing rule | 117B parameters at INT4, 0.5 bytes per parameter, 58.4 GB, times the 1.2 planning factor for key-value cache and activations | 70.1 GB against the 96 GB the two cards hold; one node of two cards is the whole cluster |
| Industrial Data Ontology for a Water Utility in the United States | One way transfer | Lidi over a hardware-enforced data diode class, behind the historian mirror in the demilitarized zone | Makes the read a physical property that cannot be misconfigured into a write; the platform holds no credentials in the control network |
| Industrial Data Ontology for a Water Utility in the United States | Time synchronization | OCP Time Card grandmaster with linuxptp and chrony | One clock for sensors, alarms, lab result dates and work orders, so the night trip argument orders correctly |
| Industrial Data Ontology for a Water Utility in the United States | Sensing | Permanent vibration on critical pump and motor frames, motor current at starter panels, reservoir level and discharge pressure from existing instrumentation | Answers the night trip with condition data; discharge pressure carries the seal failure mode vibration cannot see; no cameras or video anywhere |
| Industrial Data Ontology for a Water Utility in the United States | Pattern: commitment | The model proposes, a deterministic layer commits, named staff decide | Keeps wrong mappings from becoming ground truth exactly when the cutover depends on them; nothing is trained into weights |
| Industrial Data Ontology for a Water Utility in the United States | Pattern: namespace | Dual tag namespace bridge holding legacy and new namespaces at once | The target historian is not yet selected, so connectors re-confirm against whatever is chosen before cutover |
| Industrial Data Ontology for a Water Utility in the United States | Pattern: boundary | The utility's own server room, on premises in the United States, with Harbor offline registry, vLLM serving and RKE2 orchestration; no data leaves the boundary | The intake requires self hosting on the utility's own servers, no third party model service and a segregated control network |
| Sovereign AI Architecture for Distribution Loss and Theft Reduction in Pakistan | Reasoning model | GLM 5.3, filed at 753B parameters, FP8 (753 GB of weights), GLM-5.3 bespoke licence, served on vLLM on one node of the 141 GB HBM class in an in-country sovereign data centre or colocation on the operator's own hardware | Licence permits commercial use, fine tuning and redistribution, so the operator holds the weights; carries the agentic work surface (the loss, tamper and billing agents' drafts and the command centre's conversational interface) and ontology maintenance |
| Sovereign AI Architecture for Distribution Loss and Theft Reduction in Pakistan | Embedding model | BGE-M3, MIT, FP16, about 1.1 GB (2.27 GB at float32, 0.6 GB at 8 bit), on the 48 GB PCIe inference card of the dual socket on site inference server | Dense, sparse and multi-vector retrieval in one pass over an 8,192 token window, so tariff codes, meter numbers, PPC sections and inspection findings survive retrieval exactly and every answer cites its record; chosen over multilingual E5 (512 token window) and Nomic v2 |
| Sovereign AI Architecture for Distribution Loss and Theft Reduction in Pakistan | Forecast model | Chronos-2, Apache 2.0, FP32, about 0.48 GB (0.24 GB at 16 bit), beside the embedder on the same 48 GB card | Zero shot multivariate forecasting with covariates; the residual between each transformer's forecast and its measured energy balance is the anomaly signal, with no training run anywhere in the loop; chosen over TimesFM 3.0 and Moirai 2.0, whose non commercial terms cannot ship in an operator owned system |
| Sovereign AI Architecture for Distribution Loss and Theft Reduction in Pakistan | Frontier hardware | 141 GB HBM GPU class, eight GPUs per node, one node (1,128 GB), the frontier inference cluster class for 300B to 1T parameter open weight models | 753 GB of FP8 weights times the 1.2 planning factor is 904 GB against 1,128 GB; BF16 on two nodes rejected because it doubles export licence exposure for an accuracy gain the ranking and drafting tasks have not been shown to need |
| Sovereign AI Architecture for Distribution Loss and Theft Reduction in Pakistan | On site inference server | A dual socket server with one 48 GB PCIe inference GPU, the on site inference server class | Hosts the embedder (1.1 GB) and the forecaster (0.48 GB) together with two orders of magnitude of headroom, so the design gives them one server rather than two |
| Sovereign AI Architecture for Distribution Loss and Theft Reduction in Pakistan | Frontier footprint sizing rule | Frontier model GPU footprint sizing rule, vLLM and SGLang | Weights times the 1.2 planning factor for KV cache and activations, checked per GPU class before any card is committed, with a non US accelerator class held open |
| Sovereign AI Architecture for Distribution Loss and Theft Reduction in Pakistan | Node count rule | GLM 5.3 node count by GPU class | One node of eight 141 GB GPUs holds the weights with 375 GB left beside them; the export licence determination precedes any card commitment for Pakistan and may change the accelerator class but never the country the node sits in |
| Sovereign AI Architecture for Distribution Loss and Theft Reduction in Pakistan | Serving runtime | vLLM | Serves the FP8 weights inside the KV cache ceiling the sizing reserves; on site language model serving on the operator's own hardware |
| Sovereign AI Architecture for Distribution Loss and Theft Reduction in Pakistan | Event backbone | Apache Kafka | Interval load and tamper events on topics partitioned by meter number, so ordering holds per meter; at least once delivery with idempotent adapter writes; buffering absorbs head end bursts after an outage and replication inside the data centre |
| Sovereign AI Architecture for Distribution Loss and Theft Reduction in Pakistan | Time series store | TimescaleDB on PostgreSQL, committed at design against the measured interval rate | Interval load for several hundred thousand meters is the bulk of the design's storage and its rate is not yet stated; the shelf's InfluxDB 3, TimescaleDB and QuestDB comparison covers exactly this case |
| Sovereign AI Architecture for Distribution Loss and Theft Reduction in Pakistan | Registry | Harbor, air gapped artifact and model registry | Weights and container images enter only as versioned, signed artifacts through the registry, never over an open link; a failed update leaves the previous version serving |
| Sovereign AI Architecture for Distribution Loss and Theft Reduction in Pakistan | Identity and observability | Keycloak inside the boundary; Chrony; Prometheus | Identity is held inside the boundary with no external links; Chrony aligns clocks so an event joins to the right billing period; Prometheus watches consumer lag so a feed falling behind is seen before it goes stale |
| Sovereign AI Architecture for Distribution Loss and Theft Reduction in Pakistan | Sensing | Smart meter interval load and tamper events from the meters and feeder head meters through the meter data management system; the records of seven systems of record | No cameras, no field sensors and no edge compute; the meters and their head end already exist and the platform reads them from the data centre |
| Sovereign AI Architecture for Distribution Loss and Theft Reduction in Pakistan | Pattern: asset to alert to work order | A threshold crossing or tamper alarm becomes a tamper event or an agent draft linked to the exact meter or transformer, resolved by an inspection or an approval | Supplied by the pinned energy and utilities brief; an anomaly is never closed by editing a status field |
| Sovereign AI Architecture for Distribution Loss and Theft Reduction in Pakistan | Pattern: memory | The kept record of confirmed causes and closed cases, with customer identifiers minimised, that the next ranking reads; nothing trained into weights | The requirement asks that the next ranking be better without training a model on the fly, and historical labels are biased by where inspections happened to be sent |
| Sovereign AI Architecture for Distribution Loss and Theft Reduction in Pakistan | Pattern: validation | One circle of feeders for one quarter against a measured baseline with unchanged control feeders; known answer probes in the normal approval queue | Only untouched control feeders make the numbers defensible before the billing audit; probes catch rubber stamping and step an agent class back to shadow mode when the upheld rate falls |
| Sovereign AI Architecture for Distribution Loss and Theft Reduction in Pakistan | Pattern: boundary | In-country sovereign data centre or colocation on the operator's own hardware; reads only across the network boundary; the approved correction as the single write path into a system of record | Nothing about customers leaves the country or reaches a foreign model service; the national data protection rules, the regulator's performance standards and the privatisation conditions bind |
| A Self-Improving AI System for OT Vulnerability Management at a Utility in the United States | Reasoning model | gpt-oss-120b, filed at 117B parameters (about 120B), sparse mixture of experts, MXFP4 with the dense portions at 16 bit, Apache-2.0, served on vLLM 0.32 on one node of eight 80 GB GPUs inside the utility's own network | The requirement names open weights of United States origin on owned hardware; Apache-2.0 is a permissive licence with no usage based trigger and no field of use restriction, so the utility holds the weights with no third party model service; carries the three drafting agents, the command centre's cited conversational surface and ontology maintenance |
| A Self-Improving AI System for OT Vulnerability Management at a Utility in the United States | Embedding model | Nomic Embed Text v2 MoE, 475M parameters, sparse mixture of experts, FP16, about 0.95 GB resident (about 0.5 GB at 8 bit; the Hub weights are float32, about 1.9 GB to download), Apache-2.0, on the same node | Embeds advisories, baselines and confirmed decisions for retrieval memory, which is how the system improves cycle to cycle; all 475 million parameters stay resident because every expert is in play at inference |
| A Self-Improving AI System for OT Vulnerability Management at a Utility in the United States | Serving runtime | vLLM 0.32 | Serves the reasoning model from the node the weights sit on, with headroom for concurrent agent sessions; KV cache about 4.8 GB per session |
| A Self-Improving AI System for OT Vulnerability Management at a Utility in the United States | GPU class | 80 GB HBM GPU class, one node of eight GPUs (640 GB), about 10 kW | Holds the 70.1 GB the sizing rule demands with generous headroom; owned hardware inside the utility's own network; domestic supply, so no export control complication for a United States end user |
| A Self-Improving AI System for OT Vulnerability Management at a Utility in the United States | Sizing rule | Frontier model GPU footprint rule: filed parameters at INT4, 0.5 bytes each, times 1.2 for KV cache and activations | 117 billion parameters at 0.5 bytes is 58.4 GB; times 1.2 is 70.1 GB against the node's 640 GB; the checkpoint itself is 60.8 GiB (65.2 GB of root safetensors), and expanded to 16 bit it would need about 234 GB, which rules out a smaller host |
| A Self-Improving AI System for OT Vulnerability Management at a Utility in the United States | One-way transfer | Hardware enforced data diode class, with a DMZ conduit as the lighter alternative | OT side data (the network monitoring tool's device list) leaves the electronic security perimeter with no return path and no credentials inside the perimeter where the utility's policy demands it; the choice between the two is an open item |
| A Self-Improving AI System for OT Vulnerability Management at a Utility in the United States | Sensing | None: the network monitoring tool's device list read only across the one way path, with Chrony 4.9 holding UTC ordering | The design watches advisories, baselines and decisions; no cameras, no telemetry feed, no historian in scope |
| A Self-Improving AI System for OT Vulnerability Management at a Utility in the United States | Event backbone | Apache Kafka 4.4 | Ordering keyed on advisory and record identifiers, delivery at least once with idempotent upserts keyed on record identity and content hash, buffering for scheduled exports, replication inside the utility's own cluster |
| A Self-Improving AI System for OT Vulnerability Management at a Utility in the United States | Registry | Harbor v2.15.2, air gapped artifact and model registry | Model and artifact updates travel as signed artifacts; a fine tuned model is promoted only through the instruction gate with the prior version retained for rollback, never inside an active cycle week |
| A Self-Improving AI System for OT Vulnerability Management at a Utility in the United States | Labelling and retraining | Label Studio 1.23.2 | Supports the governed offline fine tune lane; nothing changes the served weights inside a cycle |
| A Self-Improving AI System for OT Vulnerability Management at a Utility in the United States | Identity | Keycloak 26.8, federated to the utility's own directory | No person acts anonymously; every confirmation carries the person who made it |
| A Self-Improving AI System for OT Vulnerability Management at a Utility in the United States | Observability | Prometheus 3.13 LTS | Watches the running system; record freshness and the acceptance rate are instrumented and visible in the command centre |
| A Self-Improving AI System for OT Vulnerability Management at a Utility in the United States | Pattern: memory | Retrieval memory of confirmed decisions plus a governed offline fine tune lane, nothing trained on the fly | The requirement asks for improvement from confirmed decisions without training a model on the fly; served weights must not change inside a cycle and no third party may read the record |
| A Self-Improving AI System for OT Vulnerability Management at a Utility in the United States | Pattern: projection | A projection ontology over the seven systems of record, never a migration | The governed systems stay authoritative and evidence lands where audits already look; the write surface stays two paths wide |
| A Self-Improving AI System for OT Vulnerability Management at a Utility in the United States | Pattern: boundary | The utility's own data centre rack, power, cooling and virtualisation in the United States; inbound feeds verified for source identity and integrity; nothing about the assets leaves the network | The critical infrastructure protection standards and the utility's own segregation commitment; every weight on the node is one it holds |
How much compute and hardware does AI in energy and utilities need?
The compute follows from the models: the published energy and utilities designs size it as follows, from no new hardware to a full GPU node.
| Design | Part | The design |
|---|---|---|
| Grid Context Watch | Compute | 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 |
| Grid Context Watch | Boundary | Read only and one way out of the control networks; nothing writes into an electronic security perimeter, and no data leaves the operator |
| Reliability Atlas | Compute | No hardware bought; the study runs on the operator's own in-country servers |
| Feeder Firewatch | Frontier compute | One node of eight 141 GB HBM-class GPUs holds GLM 5.2 at FP8 (753 GB of weights, 904 GB with headroom) |
| Feeder Firewatch | Edge | Sealed industrial boxes at substations and on patrol trucks detect smoke and damaged equipment when cellular coverage drops |
| Turbine Island Watch | Compute | 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 |
| Sovereign AI Architecture for a Grid Control Room in Saudi Arabia | Compute | One node of eight 141 GB HBM cards (1,128 GB) in the operator's own data center in the Kingdom, holding 753 GB of FP8 weights with a 1.2 planning factor, 904 GB; BGE-M3 on the platform zone's rack server class; no edge tier. The node is a 4A090 class item for a D:3 and D:4 destination, so an export licence determination naming the end user is a phase one checkpoint before the order |
| Industrial Data Ontology for a Water Utility in the United States | Compute | One air-cooled node of two 48 GB PCIe cards (96 GB, about 88 GB usable) in the utility's own server room, holding 58.4 GB of INT4 weights with a 1.2 planning factor, 70.1 GB; no inference at the stations or wells, and no data leaves the boundary. The room's power, cooling and rack space are confirmed at design and serving acceptance is gated on a measured load run |
| Sovereign AI Architecture for Distribution Loss and Theft Reduction in Pakistan | Compute | One node of eight 141 GB HBM GPUs (1,128 GB) holding 753 GB of FP8 weights with the 1.2 planning factor, 904 GB, plus one dual socket server with a 48 GB PCIe card for the embedder and the forecaster, both in an in-country sovereign data center or colocation on the operator's own hardware; no edge tier. The node is a 4A090 class item for a D:3 and D:4 destination, so the export licence determination precedes any card commitment, with a non US accelerator class held open |
| A Self-Improving AI System for OT Vulnerability Management at a Utility in the United States | Compute | One node of eight 80 GB HBM cards (640 GB, about 10 kW) in the utility's own data center, holding 70.1 GB by the sizing rule (117 billion parameters at INT4 times 1.2) with generous headroom for concurrent agent sessions; the embedding model on the same node; no edge tier; domestic supply, so no export control complication for a United States end user |
Is it cheaper to own AI hardware or rent cloud GPUs in energy and utilities?
Each energy and utilities 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 |
|---|---|---|
| Grid Context Watch | Three-year cost | About 510,000 dollars to own the reasoning node, about four fifths the deepest three-year AWS commitment (Appendix A) |
| Reliability Atlas | Three-year cost | No hardware line to price: the cost is the assessment work itself (Appendix A) |
| Feeder Firewatch | Three-year cost, owned | About 841,000 US dollars with support and power at the Texas industrial power price |
| Feeder Firewatch | Three-year cost, rented | 0.88 million to 1.92 million US dollars for the same GPUs around the clock; ownership costs about the same as the deepest three-year commitment (version 2) |
| Feeder Firewatch | Closed model break-even | The cheapest closed model matches the owned stack at about 41 users; above that, ownership is cheaper and the gap grows with every user |
| Turbine Island Watch | Three-year cost | 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) |
| Sovereign AI Architecture for a Grid Control Room in Saudi Arabia | Three-year cost | About 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 lists no Saudi region on the day read and every rented option would carry grid data across the border (Appendix A) |
| Industrial Data Ontology for a Water Utility in the United States | Three-year cost | About 59,000 US dollars to own the server for three years (a range of 57,000 to 61,000: server, two 48 GB cards, support and power at the United States average commercial tariff), against about 104,000 to rent the same card class on the deepest three-year commitment in AWS US East, so ownership costs a little under three fifths of it, and about 276,000 on demand. The lowest priced rented instance that holds the model, one 96 GB card of a different class, is about 33,000, against which ownership costs about 1.8 times as much. Renting is a price reference only, because every rented option would carry the utility's records outside its boundary (Appendix A) |
| Sovereign AI Architecture for Distribution Loss and Theft Reduction in Pakistan | Three-year cost | About 641,000 US dollars to own the node and the inference server for three years (a range of 615,000 to 667,000: servers, support and power at the regulator's uniform commercial tariff), against about 773,000 to rent the same cards on the deepest three year commitment in AWS Asia Pacific (Mumbai), so ownership costs about five sixths of it, and about 2.05 million on demand, against which ownership costs a little under a third. Renting is a price reference only, because AWS lists no Pakistan region and every rented option would carry customer and meter data across the border (Appendix A) |
| A Self-Improving AI System for OT Vulnerability Management at a Utility in the United States | Three-year cost | About 477,000 US dollars to own the node for three years (a range of 437,000 to 496,000: server, support and power at the national average commercial price), against about 544,000 to rent the same eight H100 cards on the deepest three year commitment in AWS US East, so ownership costs about seven eighths of it, and about 1.45 million on demand, against which ownership costs about a third; the region's other 80 GB class instance, eight A100 cards, rents for about 268,000 on the same plan, against which owning the H100 node costs about 1.8 times as much. Renting is a price reference only, because every rented option would carry the asset and vulnerability record outside the utility's network (Appendix A) |
What ontology or object model does an energy and utilities 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 energy and utilities 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 |
|---|---|---|---|
| Grid Context Watch | 14 objects, 14 links | Substation, OT Asset, Network Sensor, Anomaly Alert, Investigation Case, Vulnerability Finding, Detection Rule, Threat Intelligence Item, Pcap Evidence Record, CIP Evidence Artefact, OT Security Analyst, SIEM, Energy management system, Substation data platform | objects.json |
| Reliability Atlas | 13 objects, 14 links | Plant, Production train, Equipment, Failure mode, Work order, Downtime event, Availability prediction, Availability target, Criticality index, RCA report, Major event, Spare part, Reliability program | objects.json |
| Feeder Firewatch | 14 objects, 12 links | Substation, Feeder, Feeder segment, Pole, Recloser, Meter, Pole inspection record, Outage event, Feeder segment ignition risk score, Red flag warning, Wildfire camera station, Field crew, Work order, PSPS decision record | objects.json |
| Turbine Island Watch | 10 objects, 10 links | Steam Turbine STG-1, New Steam Turbine Generator (280 MW, 20 kV hydrogen-cooled), 3500 Vibration Monitoring Rack, Condition Anomaly Alert, Condition Work Order, Instrument Calibration Record, Commissioning and Performance Test Record, Shift Operator (Employer's operational personnel), Warranty Engineer (Contractor, DLP), STG-1 Availability (RTR-derived) | objects.json |
| Sovereign AI Architecture for a Grid Control Room in Saudi Arabia | 16 objects, 13 links | Substation, Transmission line, Busbar, Transformer, Breaker, Protection relay, Alarm, Grid event, Switching program, Planned outage, Shift log, Operating procedure, Weather cell, Connected customer, Control engineer, Shift supervisor | objects.json |
| Industrial Data Ontology for a Water Utility in the United States | 16 objects, 14 links | Treatment plant, Well, Distribution reservoir, Pump station, Lift station, Pressure zone, Pump, Pipe segment, PLC, RTU, Tag, Alarm, Work order, Sampling point, Lab result, Permit limit | objects.json |
| Sovereign AI Architecture for Distribution Loss and Theft Reduction in Pakistan | 15 objects, 14 links | Grid station, Feeder, Transformer, Service connection, Meter, Interval load, Reading, Billing account, Tariff, Inspection, Tamper event, Theft case, Recovery, Agent draft, Approval | objects.json |
| A Self-Improving AI System for OT Vulnerability Management at a Utility in the United States | 16 objects, 13 links | Substation, Protective relay, Engineering workstation, Firmware and software version, Configuration baseline, Vendor, Advisory, Vulnerability, Applicability decision, Disposition, Change request, Outage window, Work order, Evidence item, Auditor question, Cycle agent | objects.json |
Who approves the decisions an AI system makes in energy and utilities?
In every published energy and utilities design a named person makes the decision that changes the physical world; the system prepares it.
| Design | Human control |
|---|---|
| Grid Context Watch | Every triage, case and risk acceptance carries a named OT security analyst; agents draft, the analyst decides |
| Reliability Atlas | Predictions are accepted only by a named reviewer and RCA reports validated only by a named engineer |
| Feeder Firewatch | Every public safety power shutoff (PSPS) recommendation becomes a decision record a named operator approves or declines; the design never opens or closes a recloser |
| Turbine Island Watch | The platform advises and the protection layer trips: every alert carries a prescribed action a named shift operator acknowledges, and nothing writes to the DCS |
| Sovereign AI Architecture for a Grid Control Room in Saudi Arabia | Agents draft and people decide, by name: the control engineer confirms the alarm triage ranking, the shift supervisor approves every switching program draft, the incoming supervisor approves the handover, and the event analysis is prepared for the protection engineer; disagreement is a first class path faster than agreement, and every confirmation is kept as memory the next shift reads |
| Industrial Data Ontology for a Water Utility in the United States | Agents draft and named staff decide: the engineer confirms each tag mapping, the operator on shift confirms each alarm context pack, and the compliance manager approves the monthly permit report draft before any figure leaves the room; a monthly sampled audit of confirmed mappings logs the error rate, and a 15 minute staleness alarm on the historian mirror puts the command center into an explicit degraded mode |
| Sovereign AI Architecture for Distribution Loss and Theft Reduction in Pakistan | Agents draft and people decide, by name: the subdivision engineer approves the ranked causes and the inspection visits, the inspection officer approves the case file, the revenue officer approves the correction, and the Grade 17 officer signs the complaint; every approval carries the approver role, the officer grade, the timestamp, the artifact hash and a written reason, known answer probes police the approvers, and every confirmed cause and closed case is kept as the record the next ranking reads |
| A Self-Improving AI System for OT Vulnerability Management at a Utility in the United States | Agents draft and people decide, by name: the compliance engineer confirms each applicability decision, the asset owner approves each disposition and its change request, the compliance manager confirms each evidence package, and no agent applies a patch, changes a baseline or closes a change; the acceptance rate is instrumented both ways, unflagged drafts are sampled independently each cycle, and a confirmed call becomes a precedent the next cycle retrieves |
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.