From thousands of devices to hundreds of thousands.

Mass OTA campaigns

Phased rollouts (1% → 10% → 50% → 100%) with TPM 2.0 attestation gating between phases. Blue-green for atomic version swaps. Mesh self-healing through Thread / LoRa / private-5G when partial outages happen. Rollouts simulate cleanly before they touch a single live device — RolloutSim is the dispatch guard.

Cross-vendor fleet management

Mixed-vendor fleets are the norm — Jetsons, Pis, ESP32s, Corals, Hailos, custom OEM boards. Furcate runs as a KubeEdge component above the silicon mix and presents one operations surface. Akri makes leaf devices first-class K8s objects via ONVIF / udev / OPC UA discovery.

Federated learning rounds

NVIDIA FLARE's hierarchical FL architecture scales to thousands of edge devices. Schedule a round, gate participation by device class + region + battery state, aggregate deltas centrally without raw data exfiltration. Eli Lilly TuneLab, Taiwan MOHW, Tri-Labs (Sandia/LANL/LLNL) are the reference deployments.

Fleet-wide observability

Inference latency, uplink reliability, battery life, model drift, CPU utilisation — all addressable per-device, per-class, per-region. TimesFM forecasts each KPI 24h-30d ahead. Anomaly thresholds auto-generated from fleet baselines. Operator dashboards include the agent's recommended action attached to each anomaly.

Offline-capable operation

OpenYurt keeps edge nodes operational without continuous cloud connectivity. The control plane re-syncs on reconnect with bounded staleness. The right pattern for mining, maritime, remote energy, defence, and any environment where the WAN isn't reliable.

Sovereign control plane

Run the control plane at the customer's perimeter — Microsoft Azure Local (thousands-of-nodes scale, April 2026 announcement), Crusoe Edge Zones, HPE Private Cloud AI, or the customer's own bare-metal cluster. Customer data never leaves the trust boundary.