Sovereign road-edge compute

Every installed AiR unit is a tyre-safety appliance and a secure edge-compute node

Safety first, compute second. This page models how onboard inference, federated fine-tuning and scheduled tenant workloads share a bounded slice of each node, and what happens to that split as connectivity and power degrade.

Onboard inferenceFederated learningReticulum fallbackDry coolingZero-trust identity

Allocation, energy and round-time figures are produced by a deterministic planning model. Module TOPS and power envelopes are published vendor specifications, cited in the sources at the foot of this page.

Workload simulator

Edge AI workload and federated learning simulator

Choose a hardware tier, a connectivity state, a power state and a cohort size. The scheduler model reserves the tyre-safety control loop first, then allocates whatever the host has consented to share.

Controls

Hardware tier

Jetson Orin NX - Standard forecourt, 10 - 40 W.

Connectivity state

Power and thermal state

Cohort size1,000 nodes

Host-consented shared compute62%

Tyre-safety session duty cycle45%

Available per node

157TOPS

157 TOPS peak, derated by power state

Node utilisation

89.9%

15.8 TOPS idle headroom

Safety reservation

43.9TOPS

Control loop plus local intelligence

Shared compute

97.3TOPS

62% of the node

Scheduler allocation - Full cloud-connected, Mains, nominal thermals

Site ethernet or Wi-Fi plus IoT eSIM. Bulk model distribution and tenant inference payloads flow normally.

Tyre-safety control looprunning18.9 TOPS / 12%

Guaranteed reservation. Runs identically in every connectivity and power state.

On-unit local intelligencerunning25 TOPS / 15.9%

Local only. No dependency on upstream connectivity.

Federated fine-tuningrunning40.8 TOPS / 26%

Local passes over local data. Only signed deltas leave the node.

Scheduled tenant inferencerunning56.5 TOPS / 36%

Sandboxed tenant work units, quota-bounded and isolated from the tyre-safety control plane.

Federated learning round

Participating nodes947

Queued for later delivery53

Aggregate delta volume246.2 MB

Round completion6.1 min

Rounds per overnight window59

Raw driver and vehicle data never leaves the node. Only signed model deltas and secure aggregates move, and they only ever move over IP. Gradient traffic is never placed on the LoRa plane.

Cohort aggregate - 1,000 nodes

Peak installed throughput157,000 TOPS

Allocated throughput141,200 TOPS

Schedulable shared throughput97,300 TOPS

Compute power draw36 kW

Annual compute energy315.1 MWh

Of the 100 MW reference peak0.0455%

Compared on published peak INT8 throughput only. It does not imply equivalent training capability, memory capacity, model size, network performance or workload latency.

Scheduler notes for this state

A host share above 60 percent should be contracted with an explicit thermal, power and privacy schedule.

Safety first, compute second

Shared workloads are pre-empted automatically before any tyre-safety function is affected. That ordering is a hardware and scheduler guarantee, not a policy preference.

Strategic thesis

Four simultaneous roles in one installed asset

AiR already has a justified reason to exist at the point where journeys begin and end. That gives it a maintained, powered physical footprint distributed through forecourts, dealerships, homes, depots and mobility hubs.

Tyre-safety appliance

Local sensing, four-channel pressure control, ACT workflows, pre-journey checks and vehicle interaction.

Edge AI computer

Local inference, anomaly detection, vision where fitted, privacy-preserving feature extraction and model execution.

Communications node

Ethernet, Wi-Fi and cellular IoT for normal transport; LoRa and Reticulum for resilient low-bandwidth communications.

Distributed infrastructure node

Securely schedulable spare compute, local storage, signed telemetry and participation in a national sovereign edge fabric.

Exact architecture

Seven layers, from the valve to the sovereign control plane

Each layer has a fixed role, a fixed component set and a fixed data residency. Nothing above L1 can delay, block or degrade the safety control plane.

L0 Physical node

Node

The installed asset. Sensing, four-channel pneumatics, screen, secure element and radios in one sealed enclosure.

Pressure and temperature sensingValve and compressor controlTPM / secure elementNVMe encrypted storageUPS or supercapacitor

L1 Safety control plane

Node only

The deterministic tyre-safety engine. Runs to completion with no external dependency and pre-empts everything above it.

Deterministic pressure engineChannel state machinesFault detectionSigned session record writer

L2 Local inference runtime

Node

Onboard AI. Anomaly detection, leak-pattern recognition, sensor fusion and privacy redaction before anything leaves the device.

Jetson accelerator runtimeSigned model packagesLeast-privilege inference containersLocal redaction pipeline

L3 Transport layer

Node and anchor

Dual plane. High-bandwidth IP for bulk traffic, Reticulum over LoRa as an independent out-of-band resilience path.

Dual EthernetWi-Fi 6/6EIndustrial eSIM, multi-networkRNode LoRa radioNominated Reticulum transport nodes

L4 Regional anchor layer

Region

170 backbone anchors provide regional persistence, job brokering, encrypted caching and Reticulum transport.

AGX Orin or multi-accelerator edge serverDual diverse backhaulHigh-gain engineered antennaGNSS timingRegional job broker and cache

L5 Sovereign control plane

UK sovereign service

UK-hosted identity, scheduling, model distribution, attestation and revocation. Owns orchestration, never owns raw driver data.

Device CA and key managementNational workload schedulerModel registry and signingAttestation and auditRevocation and blackholing

L6 Application and tenancy plane

UK analytics and tenant planes

Driver app, partner portals, research analytics and isolated compute tenancy, each with its own namespace and quota.

Driver and fleet applicationsPartner and OEM private namespacesResearch analytics planeTenant-isolated inference namespaces

Dual-plane transport

A primary data plane and an independent sovereign resilience plane

The two planes share no dependency. When commercial connectivity fails, the resilience plane keeps safety and operational messaging alive at low bandwidth.

Primary data plane

Ethernet

Permanent forecourt, dealership, fleet depot and commercial sites

Preferred where available: highest bandwidth, lowest cost per byte

Wi-Fi 6/6E

Home units, dealer premises and managed local networks

Primary path for the residential estate

Industrial eSIM / eUICC

Every node as either primary or failover

Multi-network UK roaming, private APN or static addressing where commercially available

5G fixed wireless or satellite

Selected backbone anchors only

Diverse second backhaul at strategic sites

Sovereign resilience plane

StackReticulum, a cryptography-based networking stack designed for heterogeneous interfaces under low-bandwidth and high-latency conditions, operating with or without infrastructure.

RadioRNode-compatible LoRa interfaces using raw LoRa modulation, with Reticulum providing the higher-level protocol stack and direct control of LoRa parameters.

CryptographyX25519 key exchange, Ed25519 signatures, encrypted single-destination traffic, ephemeral keys and forward secrecy.

AuthenticationInterface Access Codes for authenticated interfaces, plus network identities for discovery.

Transport nodesDeliberately nominated, not universal. Reticulum documentation warns that enabling forwarding everywhere degrades performance, so only anchors and selected dense-area nodes forward.

Honest wordingCryptographically secured and zero-trust by design. Never described as completely secure, because no credible network can promise that.

What LoRa carries, and what it must never carry

Message	Resilience plane	Transport
Node heartbeat and health	Carried	LoRa with a cellular duplicate
Alarm, tamper and outage alert	Carried	LoRa first when IP is unavailable
Signed location and site status	Carried	LoRa
Small command and job metadata	Carried	LoRa where the payload permits
Emergency text and service message	Carried	Reticulum
Model binaries	Never	Ethernet, Wi-Fi or cellular
Video and images	Never	Ethernet, Wi-Fi or cellular
Large telemetry batches	Never	Ethernet, Wi-Fi or cellular
Distributed inference tensors	Never	High-bandwidth IP only

Data flow

From valve to analytics plane, in eight stages

Classification happens on the node and is immutable afterwards. A message's class fixes its transport, its priority and where it is allowed to be stored.

Sense

Node, L0

Pressure, temperature, valve state and session context captured at the channel. Raw sensor data never leaves the enclosure in raw form.

Decide locally

Node, L1

The deterministic engine selects an approved profile, applies the thermal allowance and safety clamp, and drives the channels. No network dependency exists at this stage.

Infer and redact

Node, L2

Onboard inference runs anomaly, leak-pattern and integrity checks, then the redaction pipeline strips identifiers and derives features before anything is queued for transport.

Sign and classify

Node, L2

The session record is signed with the device key and each outbound message is tagged P0 to P4, which fixes its transport, priority and residency for the rest of its life.

Transport by class

P0 and P1 take whichever path is available, LoRa included. P2 to P4 wait for IP. Bulk traffic is deferred to off-peak windows and never touches the resilience plane.

Broker regionally

Anchor, L4

Anchors cache models, broker compute work units, forward Reticulum traffic and hold store-and-forward queues for nodes that are currently unreachable.

Reconcile centrally

Control plane, L5

Signed sequence numbers and idempotent processing mean queued events replay safely when IP returns. Duplicate delivery over two planes is expected and harmless.

Aggregate and serve

Personal records land in the UK sovereign service, anonymised trends in the analytics plane, tenant outputs in an isolated namespace. The three are never joined.

When the primary path fails

ImmediatelyPreserve local tyre operations. The safety function has no upstream dependency, so nothing about it changes.

Within secondsQueue bulk data locally on encrypted storage and stop attempting P2 to P4 delivery.

Within secondsShift only essential P0 and P1 messages to Reticulum over LoRa, where an available path exists.

While degradedContinue signing every record with the device key and monotonic sequence numbers so ordering survives the outage.

On IP returnReconcile queued events using signed sequence numbers and idempotent processing; duplicates from the dual planes are discarded.

After recoveryReport the outage window, path taken and reconciliation result into the security and audit plane.

Federated learning: raw data never moves

Local trainingFine-tuning runs on data that never leaves the node or its regional trust boundary.

Update extractionOnly model updates or anonymised aggregates are transmitted, never raw vehicle or user records.

Secure aggregationRegional anchors aggregate updates so the control plane cannot attribute an update to one node.

Differential privacyNoise is added where a research use case justifies it, at the cost of some model fidelity.

National reconstructionThe central model is reconstructed from aggregates and redistributed as a signed package.

Tenant isolationExternal compute tenants run in separate namespaces with no route to vehicle, driver or safety data.

Hardware-bound identity lifecycle

01Every unit receives a hardware-bound AiR identity during manufacture.
02Private keys are generated or injected inside a hardware security module or TPM-class secure element.
03The device certificate is tied to serial number, board identity, site and lifecycle state.
04A second, separate Reticulum identity is issued for the resilient communications plane.
05Compromised or retired devices are revoked centrally and blackholed on the sovereign plane.

Zero-trust controls at every node

Mutual authentication on every service-to-service connection
Signed firmware and signed model packages
Secure boot and measured boot
Least-privilege containers for inference workloads
Encrypted local storage
Remote attestation where hardware supports it
No acceptance of unsigned workloads
Per-tenant scheduling and resource quotas
Independent security logging and tamper events

Deployment assumptions

The 170-anchor backbone, stated as planning assumptions

These are assumptions and design targets, not measured coverage. They are published in this form so they can be tested rather than believed.

UK land area

0 km2

Backbone anchors

Average per anchor

0 km2

Planning-cell radius

0.0 km

Anchor site selection

Motorway and trunk-road service areas
Large dealer groups and mobility campuses
Roadside-assistance operational sites
Fuel retail estates and logistics hubs
Selected rooftops, towers and elevated commercial properties
Ports, airports and strategic freight corridors
Resilience sites near critical junctions and in regions with poor conventional connectivity

Required before any coverage claim

Propagation modelling against real terrain data
Antenna-height optimisation per site
Interference analysis in shared spectrum
Regulatory compliance for UK LoRa spectrum use
Drive testing to measure, not model, achieved coverage

The 21.4 km figure is an average planning-cell equivalent derived by dividing UK land area by 170, not a coverage claim. Urban clutter can reduce useful LoRa reach substantially and elevated line-of-sight paths can extend it well beyond it.

Anchors versus ordinary nodes

Anchors provide regional persistence and bridging. Ordinary forecourt and home nodes are Reticulum instances, and only a selected subset become local transport nodes where density and topology justify it. That is what turns 170 anchors into a practical national mesh rather than 170 isolated radios.

Home / Orin Nano25 W peak - 219 kWh/yr at 100% - Wall-mounted, passive-first, Wi-Fi primary

Forecourt / Orin NX40 W peak - 350 kWh/yr at 100% - Ethernet primary, filtered low-speed fans when required

Backbone / AGX Orin60 W peak - 526 kWh/yr at 100% - Anchor duty: transport, cache and job brokering

Why the cost case is a systems argument

Shared site acquisitionNo separate land acquisition per unit of edge capacity; the site is already justified by tyre safety.

Shared enclosure and serviceEnclosure, connectivity, power and field-service cost are carried by the core business.

No water at the nodeDry cooling by design, so no evaporative or chilled-water loop at the edge.

No single grid connectionTens of watts across many sites instead of one 100 MW grid connection and one thermal zone.

Reduced backhaulInference happens close to the data, so only features and aggregates travel.

Incremental capacityCapacity scales with each commercial deployment rather than a separate construction programme.

Honest limitCentral hyperscale achieves exceptional utilisation and network density. The edge advantage holds where local inference, low-latency control, distributed redundancy, sovereignty or otherwise-idle installed compute matter.

IoT connectivity contract requirements

Industrial eSIM / eUICC profiles supporting more than one UK mobile network
Static or private addressing options
VPN or private APN capability
Roaming control and pooled data allowances
Remote SIM lifecycle management
Full API access for provisioning and diagnostics
The SIM is transport, not sovereignty: sovereignty comes from AiR-controlled identity, keys, software, data routing and the independent fallback path

Hardware tiers

Three node classes, tens of watts each

Peak AI performance figures are the published NVIDIA Jetson specifications for each module. Annual energy assumes the full compute envelope held continuously, which no real duty cycle reaches.

Tier	Deployment	Module	Peak INT8	Power	Storage	Annual kWh at 100%
H1	Home wall-mounted	Jetson Orin Nano Super	67 TOPS	7 - 25 W	512 GB - 1 TB NVMe	219 kWh
F1	Standard forecourt	Jetson Orin NX	157 TOPS	10 - 40 W	1 - 2 TB NVMe	350 kWh
F2	High-value site / backbone anchor	Jetson AGX Orin	275 TOPS	15 - 60 W	2 TB NVMe + regional cache	526 kWh

Suited to the edge fabric

Embarrassingly parallel inference batches
Regional risk scoring
Anonymised tyre-health feature extraction
Model A/B evaluation
Synthetic-data generation at idle nodes
Federated fine-tuning where data stays local
Map-reduce analytics across geographically partitioned datasets

Not suited, and not claimed

Tightly coupled large-model training that depends on NVLink or InfiniBand fabrics
Single jobs needing hundreds of gigabytes of unified accelerator memory
Latency-locked synchronous all-reduce at national scale
Anything requiring raw driver or vehicle data to leave the node

Large training runs depend on tightly coupled NVLink and InfiniBand fabrics. The honest position is a complementary national inference fabric, not a replacement for hyperscale.

What the national scheduler knows

Hardware tier and available accelerator memory
Live thermal headroom and fan health
Local queue depth and safety-session duty cycle
Connectivity class and measured uplink quality
Site policy and host consent tier
Power state, UPS status and tariff window
Attestation state and firmware signature validity

Jobs are decomposed into small signed work units and dispatched only to eligible nodes. Unsigned workloads are never accepted, and tenant namespaces cannot reach vehicle or user data.

Dry-cooling thermal hierarchy

01Passive chassis conduction as the first thermal path
02Large external fin area integrated into the cabinet structure
03Low-speed filtered fans only when required
04Temperature-aware inference scheduling
05Automatic compute throttling before tyre-safety functions are affected
06No evaporative or chilled-water loop at the edge node

A central AI campus concentrates tens or hundreds of megawatts into one thermal zone. AiR distributes tens of watts across thousands of sites, so zero on-site cooling water at the node is a design target rather than a marketing line.

Equivalence, stated honestly

What a 100 MW comparison actually means

A 100 MW rating is an electrical capacity figure, not a compute figure. This model fills 100 MW with DGX H200 systems at their maximum system input and ignores facility overhead, storage and switching so the arithmetic stays checkable.

Reference systems

DGX H200 at 10.2 kW system input

Reference GPUs

Eight H200 SXM per system

Peak INT8 per GPU

0 TOPS

Reference peak

0M TOPS

Theoretical peak INT8 only

Jetson Orin Nano Super at 67 TOPS~4.63 million nodes

Jetson Orin NX at 157 TOPS~1.98 million nodes

Jetson AGX Orin at 275 TOPS~1.13 million nodes

If the network were only 170 backbone sites, each anchor would need roughly 1.83 million TOPS, which is mini data-centre territory. The attractive model is the inverse: use the 170 anchors to coordinate a very large population of low-power nodes already justified by the core tyre-safety business.

Cohort against the reference

This cohort, allocated141k TOPS

This cohort, peak157k TOPS

100 MW reference peak310429k TOPS

Move the cohort slider in the simulator to see where the installed estate crosses one percent, ten percent and parity on this narrow peak-throughput metric.

Dual-plane network

A high-bandwidth commercial data plane for normal operation, plus a cryptographically secured sovereign resilience plane that keeps exchanging essential messages independently of the public Internet.

Traffic classes and priority

P0 SafetyHighest - LoRa first if IP is unavailable

Device alarms, safety fault, critical control

P1 OperationsHigh - LoRa with cellular duplicate

Telemetry, job status, software health

P2 DataMedium - IP only

Compressed analytics and research uploads

P3 ComputePolicy controlled - high-bandwidth IP only

Distributed inference payloads

P4 BulkOff-peak - IP only

Model downloads, logs, archives

Graceful degradation in five layers

Full cloud-connected

Mode 1

Site IP plus cellular. All five traffic classes flow, tenant inference and model distribution included.

Cellular-only

Mode 2

Private APN carries P0 to P2 normally; P3 and P4 move into policy-controlled off-peak windows.

Local-site

Mode 3

No upstream IP. Local inference and site aggregation continue, bulk data queues on encrypted local storage.

Reticulum / LoRa fallback

Mode 4

Essential P0 and P1 messages shift to the sovereign plane via nominated transport nodes and store-and-forward.

Isolated safe-local

Mode 5

Every external link is gone and the core tyre-safety function still runs, signing records for later reconciliation.

Data sovereignty

Where each class of data is allowed to live

Sovereignty is not the SIM card. It comes from AiR-controlled device identity, cryptographic trust, software, key management and an independent fallback path.

Class	Examples	Default location
Safety-critical local	Pressure control state, valve status	Node
Personal / vehicle linked	Registration-linked history, account data	UK sovereign service
Research aggregate	Anonymised pressure and usage trends	UK analytics plane
Compute tenant	Tenant model inputs and outputs	Tenant-isolated namespace
Security	Attestation, audit, tamper events	UK security plane

Sovereign requirements

UK-hosted control plane and identity infrastructure
UK-controlled keys and certificate authorities
Documented data residency and processor chain
Offline-capable core safety function
Software bill of materials and signed supply-chain artefacts
Independent penetration testing and recurring security assurance
Clear separation of public-interest data from commercial compute tenancy

The correct wording is cryptographically secured and zero-trust by design. No credible network promises absolute security.

Reference architecture

What actually goes inside each node class

Forecourt node

Jetson Orin NX-class accelerator as baseline, AGX Orin at high-value sites
1 - 2 TB NVMe local storage
TPM or secure element, dual ethernet where practical, Wi-Fi 6/6E
Industrial eSIM with multi-network management
RNode-compatible LoRa radio and external antenna, optional GNSS
Thermal and fan-health sensing, UPS or supercapacitor for graceful shutdown

Home node

Orin Nano-class accelerator
Wi-Fi or ethernet primary, IoT eSIM as a premium option
LoRa and Reticulum radio
512 GB - 1 TB local storage
Passive-first wall-mounted thermal design

Backbone anchor

AGX Orin or a small multi-accelerator edge server
Dual diverse IP backhaul
High-gain, professionally engineered LoRa antenna system
Larger UPS and monitored power
Regional Reticulum transport-node role
Encrypted regional cache and job broker

Deployment phases

Phase 0 - Lab

10 - 25 nodes

Validate hardware, Reticulum, LoRa, the scheduler and secure identity end to end.

Phase 1 - Regional pilot

100 - 250 nodes

Prove mixed cellular and LoRa operation plus real distributed inference against live sites.

Phase 2 - 170-anchor backbone

170 anchors + thousands of edge nodes

Establish the national resilience spine with elevated, redundantly backhauled transport nodes.

Phase 3 - Commercial density

10,000 - 100,000 nodes

Create a distributed compute fabric with enough aggregate throughput to sell capacity.

Phase 4 - National infrastructure

100,000s to millions

Sovereign road-edge AI network at hyperscale-class aggregate peak throughput.

Gate criteria before any national claim

Measured RF coverage, not theoretical coverage
Measured task throughput and utilisation
Measured power and thermal envelope
Independent cyber-security assessment
Failure-mode testing with cellular and Internet removed
Demonstrated store-and-forward delivery across Reticulum transport nodes
Documented UK spectrum compliance

Revenue layers on one installed asset

Core AiR tyre-safety service
Premium AiR Home hardware and subscriptions
Fleet connectivity and management
Connectivity margin and managed IoT service
Sovereign edge-inference capacity
Research and aggregated road-intelligence services
Resilience-network service contracts
OEM and dealer private edge namespaces

Engineering truths

Claims we will not make, and the position we hold instead

Each row is a misconception this architecture is routinely credited with, paired with the position AiR states publicly.

'170 LoRa sites cover the whole UK'

Treat 170 as backbone anchors. Coverage is proven with RF engineering and edge-node density, not map circles.

'LoRa can carry distributed AI traffic'

No. LoRa carries low-rate control, status and resilience traffic. Bulk compute traffic uses IP.

'100 MW equals a fixed amount of compute'

No. A reference architecture and numeric precision must be stated before any comparison.

'Edge compute replaces hyperscale'

Only for suitable distributed inference workloads, never tightly coupled training.

'Completely secure'

Zero-trust, hardware identity and encryption. No credible network promises absolute security.

'No water'

Zero on-site cooling water is claimable only where the production thermal design genuinely uses dry cooling.

'Cellular is sovereign'

Cellular is a transport dependency. Sovereignty comes from AiR-controlled cryptography, data, software and fallback paths.

Technical basis

Sources

[1]Reticulum Network Stack - 'What is Reticulum?' and networking documentation. Cryptographic local and wide-area networking under low-bandwidth, high-latency conditions.
[2]Reticulum Network Stack - RNode LoRa interface documentation. Reticulum operates over LoRa with direct control of LoRa parameters.
[3]NVIDIA Jetson Orin Nano Super - up to 67 INT8 TOPS, published low-power operating envelope.
[4]NVIDIA Jetson Orin NX - up to 157 TOPS, configurable low-power edge module.
[5]NVIDIA Jetson AGX Orin - up to 275 TOPS at 15 - 60 W configurable power.
[6]UK House of Commons Library - 'Data centres: planning policy, sustainability, and resilience', 27 May 2026.
[7]Reticulum - 'Building Networks'. Every node can be a transport node, but enabling transport everywhere degrades performance.
[8]Reticulum communications hardware documentation - RNodes use raw LoRa modulation with Reticulum providing higher-level protocol functions.
[9]Reticulum - 'Understanding Reticulum' and API reference. X25519, Ed25519 and AES-based encryption with ephemeral keys.
[10]Reticulum - 'Building Networks'. Trustless networking, interface access codes and encrypted transport behaviour.
[11]NVIDIA DGX H100/H200 user documentation - eight GPUs per system, 10.2 kW maximum system input.
[12]NVIDIA H200 specifications - up to 3,958 INT8 Tensor Core TOPS per H200 SXM GPU, up to 700 W TDP.
[13]LoRa Alliance - LoRaWAN overview. Low-power wide-area networking for massive IoT and small, low-rate data flows.

Methodology note: the 100 MW equivalence assumes 100 MW of DGX H200 system input and deliberately ignores facility PUE, storage, network switching and redundancy so the arithmetic remains transparent. The result is a theoretical peak-throughput reference, not a data-centre design or a benchmark.

See the comms architecture and the 170-anchor backbone