Fleet Autonomy Edge Compute

edge · telematics · upload · OTA

Autonomous vehicles produce far more data than cellular links can carry. The depot stop is the data window, and its length sets the network.

Link rate needed to move one terabyte off a single vehicle: 8.9 gigabits per second in a 15-minute stop, 4.4 in a 30-minute charge, 2.2 in a one-hour dwell and 0.28 over an eight-hour night. A dashed line marks 10 gigabit Ethernet. Figures are sustained throughput; protocol overhead and shared links push real needs higher. LINK RATE TO MOVE 1 TB OFF ONE VEHICLE 15-minute stop 8.9 Gb/s 30-minute charge 4.4 Gb/s One-hour dwell 2.2 Gb/s Eight-hour night 0.28 Gb/s 10 GbE line rate Sustained throughput; protocol overhead and shared links push real needs higher.
1 TB in 30 min
4.4 Gb/s
Upload window
Depot dwell
Highest-risk step
OTA
Hub
Autonomy

Finding

A robotaxi's logs leave the vehicle in the depot, not over the air, so dwell time and depot networking set how much the fleet can learn each day

Moving one terabyte in a 30-minute charging stop takes a sustained 4.4 Gb/s per vehicle - more than most Wi-Fi links deliver in a busy yard and close to half of a 10-gigabit Ethernet port. Cellular is reserved for telematics, remote assistance video and urgent events. Fleets therefore filter and prioritize on the vehicle, then upload bulk logs over high-capacity depot links during charging or shift change.

The edge stack at each depot is usually a managed appliance from the AV developer: servers and storage for log ingest, local networking (wired, Wi-Fi 6E or 7, sometimes private 5G), gateways to chargers and telematics, and containerized analytics. It filters and compresses data, flags safety events for fast review, stages software and map updates, and keeps the depot running when the wide-area link drops. Tesla's training clusters in Austin, Waymo's Google TPU capacity and Chinese clusters at Baidu and XPeng sit at the other end of the pipe.

Over-the-air updates are the highest-leverage and highest-risk step. A bad release can ground a fleet, so updates go out in stages - a few vehicles, then a depot, then a city - with rollback and with blackout periods written into operator contracts. Under UNECE R156, vehicles sold in Europe need a certified software-update management system; China requires OTA filing with MIIT. See OTA Architecture and Telemetry.

Data flows

Data flows through the depot

Depot data flows
DirectionWhat movesWhere processed
Vehicle to depotLogs, sensor clips, fault codes, remote-assistance annotationsDepot edge ingest
At the depotFiltering, compression, sampling, event detectionEdge servers
Depot to cloudPrioritized subsets for training and safety reviewTraining and simulation clusters
Cloud to vehicleSoftware, maps, configuration, geofencesStaged through the depot edge

Ownership

Who controls what

Most autonomy logic, data schemas and OTA mechanisms belong to the AV developer; operators shape the environment and the contract.

Edge pipeline control
PartyControls
AV developer or OEMAutonomy stack, data schemas, training pipeline, OTA delivery
Fleet operatorDepot networking, backup links, retention of operator data, uptime SLAs, physical security, update windows
SharedEdge node placement, separation of personal data, safety review workflows, dashboards

Resilience

Failure modes and mitigations

Edge pipeline failure modes
FailureEffectMitigation
Wide-area link downUploads and OTA campaigns stallSecond fiber path or cellular backup
Edge storage fullLogs lost or vehicles heldSize storage for several days of backlog
Bad updateVehicles held or rolled backStaged rollout, canary vehicles, tested rollback
Clock driftEvents cannot be reconstructedPrecision time sync across vehicles and edge
Edge power lossDepot operations degradeUPS and a protected microgrid segment

Sources

Sources and scope

Scope
Fleet Autonomy Edge Compute
Sources
UNECE R156; MIIT OTA rules; ElectronsX arithmetic (1 TB = 8,000 gigabits)
Updated
October 2026