Yobitel Edge AI

Overview#

Some workloads cannot wait for the cloud. A robot's grasp decision needs to happen in under 5 ms; a livestock-monitoring fleet runs in pastures with intermittent connectivity; a defence sensor must keep operating after the satellite uplink drops; a manufacturing-vision QC pipeline needs to inspect every part at line speed without exporting the image stream. Each of these is a physical AI or edge IoT workload, and each needs a runtime that is on-device, fleet-managed, and operated under the same identity, audit, and sovereignty surface as the customer's cloud-resident AI.

Yobitel Edge AI is the managed edge-inference capability inside the Yobitel Infrastructure pillar. It is the edge counterpart to cloud Yobibyte: customers register a fleet (a logical group of edge sites running a shared application), push a model image, and observe rollout, attestation, and telemetry through the same Yobibyte console used for cloud workloads. The runtime on each device is industry-standard (per platform) and operated by Yobitel; the customer's interaction is the fleet surface, not the per-device runtime.

Compared with AWS IoT Greengrass and Azure IoT Edge, the differentiator is two-fold: first, the platform surface above is the same Yobibyte workspace that operates the customer's cloud inference, so identity, audit, billing, and policy do not fragment between cloud and edge; second, the sovereignty discipline that runs in cloud regions extends to edge fleets — signed model images, attested rollouts, KMS-anchored update chains, and UK-resident audit destinations for UK-pinned workspaces. Compared with running an edge stack in-house on commodity orchestration, the difference is that Yobitel operates the fleet's runtime end-to-end; the customer configures the fleet, not the per-device update mechanics.

Yobitel Communications — a UK-headquartered AI infrastructure company and NVIDIA Inception partner — operates the edge capability. Fleets can be UK-only (NCSC OFFICIAL eligible), EU-only (EU Data Boundary), or US-only (FedRAMP-equivalent) depending on the host workspace. Pricing is in USD throughout per-device per-month with predictable rollover for adds and removals.

Quick start#

Edge deployment is a workspace activity. Sign in to the Yobibyte console with your corporate identity provider and open the Edge AI view; the surface is scoped to your workspace's sovereignty pin (a UK-bound workspace's fleets and audit destinations stay UK-resident).

Register a fleet. The dialog asks for fleet name, target hardware platform (Jetson Orin NX, AMD Versal AI Edge, etc.), expected fleet size envelope, and the OTA channel (`stable`, `staged`, or `canary`). Yobitel produces a fleet identity (a signing identity bound to the workspace's KMS chain) and an enrolment token that field-deployed devices use to join the fleet. The enrolment workflow is documented per hardware platform and is typically embedded in the customer's existing device-imaging pipeline.

Push a model image. From the marketplace or from a customer-private upload, pick the model and select the fleet as the deployment target. Yobitel produces a fleet-signed, attested model image for the fleet's hardware platform; the OTA channel determines whether the rollout is staged or pushed to all devices. The console surfaces per-device acceptance, attestation success, and the first inference's telemetry sample for each device as the rollout progresses.

Observe the rollout. The fleet dashboard surfaces rollout progress (percentage of devices on the new image), health (each device's last-seen, inference rate, telemetry sample), and any attestation failures. Customers running this for the first time typically watch a canary rollout to a single site, then promote to a staged channel, then to the full fleet.

Concepts#

Edge AI introduces a small set of concepts that sit alongside the workspace primitives. The mental model is: a fleet is a managed group of edge sites; an OTA channel is the rollout lever; a model image is the unit of deployment; attestation is the trust mechanism; telemetry is the observability surface.

• Edge Site — a single physical device or grouped set of devices at one location (a robot, a camera, a vehicle, a kiosk, a sensor cluster). Sites enrol against a fleet identity and inherit the fleet's policy.
• Fleet — a logical group of edge sites running a shared application or model. Fleets are scoped to a workspace, inherit its sovereignty pin and audit destination, and are the unit at which OTA, rollback, and policy apply.
• OTA Update — the over-the-air update path for fleet model images. Updates are signed by the fleet identity, attested on-device, and applied via the customer's chosen channel (`stable`, `staged`, `canary`).
• Local Inference Runtime — the per-platform managed runtime that loads and serves models on the device. Customers do not select or operate the runtime; Yobitel ships the platform-appropriate runtime as part of the fleet's pinned profile.
• Model Sync — the mechanism by which model images reach devices: pull from the workspace's UK/EU/US-resident object store, with delta updates and resumable transfers for low-bandwidth or intermittent links.
• Telemetry Stream — the per-device telemetry the fleet emits back to the workspace: inference rate, model output sample (subject to policy), device health (CPU, memory, thermals), and observability signals. Subject to the customer's redaction policy and sovereignty rules.
• Attestation Chain — the signature and verification chain that proves a model image came from the workspace's KMS-anchored identity. Devices refuse images that fail attestation.
• Fleet Policy — the customer-configurable rules that govern what a fleet can run: model size ceiling, OTA channel, telemetry sampling rate, redaction rules, audit destination, and the kill-switch.

Reference — supported edge hardware platforms#

The table below is the supported hardware index. Per-platform runtime profiles (quantisation defaults, supported model families, memory and compute envelopes) are published in each platform's edge-hardware listing.

Workload patterns#

Three deployment shapes cover most of what customers do with Edge AI — a livestock-monitoring fleet running with intermittent connectivity, an industrial-vision QC pipeline on a manufacturing line, and a sovereign defence edge with attested updates only.

• Livestock fleet monitor — an agricultural customer deploys a fleet of 800 Jetson Orin Nano units across rural sites. The fleet's OTA channel is `staged` with a 7-day window; telemetry sampling is 1% to conserve bandwidth; the fleet's audit destination is the customer's EU-resident SIEM. Model images are signed by the fleet identity; each site verifies the signature against the workspace's KMS chain before applying. Sites operate offline for up to 72 hours; telemetry buffers locally and syncs on reconnect.
• Industrial vision QC — a manufacturer deploys a fleet of 60 Hailo-10 inference accelerators across a production-line vision system. The fleet's OTA channel is `canary` to one line for new model versions; promotion to `staged` and full rollout requires the platform team's approval. Telemetry includes per-part inference results streamed to the manufacturer's MES; the policy redacts image content from telemetry by default to keep the stream small.
• Sovereign defence edge — a defence-adjacent supplier deploys an AMD Versal AI Edge fleet with attestation-only updates and a UK-resident audit destination. The fleet's policy requires attestation success before any device accepts an image; failed attestations alert immediately. The fleet is in a workspace bound to UK NCSC OFFICIAL with the OFFICIAL-SENSITIVE handling option enabled.

# PREVIEW - the EdgeFleet declarative shape is in active development;
# this is the planned shape and is not runnable today. Edge fleets
# are registered today through the Yobibyte console.
#
# apiVersion: yobibyte.yobitel.com/v1
# kind: EdgeFleet
# metadata:
#   name: livestock-uk
#   workspace: agri-customer-uk
# spec:
#   hardwarePlatform: jetson-orin-nano
#   expectedSize: 800
#   sovereignty: uk-ncsc-official
#   otaChannel: staged
#   model:
#     marketplaceEntry: livestock-vision-2026.06
#     signedBy: fleet-identity-livestock-uk
#   policy:
#     telemetrySamplingRate: 0.01
#     redactImageContent: true
#     offlineBufferHours: 72
#     auditDestination: eu-siem-prod
#   spendCap:
#     amount: 12000
#     currency: USD
#     window: monthly

Fleet size and edge SKU pairings#

The table below pairs typical fleet sizes with the edge SKU most often chosen for the workload. Fleet size is an envelope, not a hard limit; specific deployments may run much larger fleets on the same SKU.

Limits and quotas#

Default per-fleet and per-workspace limits exist to protect the OTA fabric and the customer's bandwidth. Almost every limit is raisable on request.

Observability#

Every fleet emits a stable `yobitel_edge_*` metric set: fleet health (online site count, last-seen age, model version distribution), rollout health (rollout percentage, acceptance rate, attestation success rate), and inference health (per-site inference rate, telemetry sample, on-device thermals and power). The hosted Grafana dashboard surfaces these by default; customers ship the same stream to their own Prometheus or OpenTelemetry collector.

Three signals matter most for edge fleets: fleet online ratio (are sites reachable?), model rollout success rate (is the new image actually landing?), and attestation failure rate (is the trust chain intact?). The PromQL block below is the alert most fleet operators add first.

groups:
- name: yobitel-edge
  interval: 30s
  rules:
  - alert: EdgeFleetOnlineRatioDegraded
    expr: |
      avg by (workspace, fleet) (
        yobitel_edge_fleet_online_ratio
      ) < 0.85
    for: 30m
    labels: { severity: page }
    annotations:
      summary: "{{ $labels.fleet }} online ratio below 85%"

  - alert: EdgeOtaRolloutStuck
    expr: |
      yobitel_edge_ota_rollout_acceptance_ratio < 0.9
      and on (workspace, fleet)
      yobitel_edge_ota_rollout_active == 1
    for: 1h
    labels: { severity: warn }

  - alert: EdgeAttestationFailures
    expr: increase(yobitel_edge_attestation_failed_total[15m]) > 0
    for: 1m
    labels: { severity: page }

Cost and FinOps#

Edge AI is priced per-device per-month in USD, with a tier ladder that covers the on-device runtime, OTA bandwidth envelope, telemetry ingestion, and the workspace's audit pipeline. Hardware itself is customer-supplied; the Yobitel pricing covers the operating envelope around it. Spend caps can be set per fleet and per workspace.

Security and compliance#

Signed model images anchor the trust model. Every model image is signed by the fleet's identity, which is itself bound to the workspace's KMS chain; devices refuse images that fail attestation against the workspace's published public-key chain. Attestation failures alert immediately and are retained in the audit stream.

Sovereignty extends to the edge. A fleet in a workspace bound to UK NCSC OFFICIAL ships telemetry to a UK-resident audit destination, refuses to enrol sites whose declared location is outside the UK, and uses UK-resident object storage for model image distribution. The same discipline applies to EU Data Boundary and US FedRAMP-equivalent fleets in their respective workspaces.

• Signed model images — every image signed by the fleet identity, anchored in the workspace's KMS chain.
• On-device attestation — devices verify image signatures before applying; failures are alertable and audited.
• Hardware root of trust — supported on Jetson and AMD Versal AI Edge; required for OFFICIAL-SENSITIVE fleets.
• NCSC Cloud Security Principles — applies to the workspace; edge fleet configurations inherit the workspace's mapping.
• G-Cloud framework — sovereign edge fleets procurable via the framework alongside the rest of the Yobitel stack.
• ISO 27001 / SOC 2 Type II — current certificates available under NDA.
• Cyber Essentials Plus — annual third-party assessment maintained.
• GDPR / UK DPA 2018 — DPA, sub-processor list, EU SCCs available; edge telemetry residency enforced.
• EU AI Act — edge-deployed models declared per fleet; risk classification inherited from the model entry.
• Sovereign edge attestation — UK NCSC OFFICIAL fleets require attestation success on every update.

Alternatives and customer-owned baseline#

Without Yobitel Edge AI, customers either stitch together open edge stacks (Greengrass, IoT Edge, k3s + manual signing) or run a single-vendor edge SaaS that does not extend their cloud workspace surface. The comparison below positions the offering.

Troubleshooting#

The errors below cover the failure modes seen most often during fleet onboarding and the first rollout cycles. The full runbook library is at docs.yobitel.com/runbooks.

Where Yobitel Edge AI fits in the Yobitel stack#

Edge AI is the edge counterpart to cloud Yobibyte. The same workspace, the same marketplace, the same identity and audit surface, the same FOCUS billing pipeline — all extend to the edge fleet. Customers who already use Yobibyte for cloud inference adopt Edge AI as an additional surface, not a new platform; customers who are edge-first can still benefit from the workspace's cloud features when the workload pattern needs them (training, fleet-level analytics, knowledge-base maintenance).

The AI Applications Suite is a frequent consumer: applications like Livestock Monitor publish to fleet targets out of the box, and the workspace's policy bindings extend to the edge fleets they deploy onto. Yobitel GPU Cloud is the cloud counterpart for any training or evaluation the customer needs to support the edge workload — the cloud workspace trains the model, the edge fleet runs it.

Omniscient Compute indexes a near-edge tier (regional inference points of presence) for workloads that fit between on-device and central-cloud. Edge AI is the on-device surface; near-edge is Omniscient territory. Customers picking between them choose by latency, bandwidth, and sovereignty; the Yobibyte workspace is the same on both sides of the line.