Data Autonomy

local inference · open weights · data rules · audit

A robot or vehicle that waits on a remote model is only as autonomous as the network and the vendor. Critical decisions run on hardware and weights the operator controls.

Bar chart of how far a car at 70 miles per hour travels while waiting on added network delay: on-board compute adds nothing; a metro edge site at about 10 milliseconds adds about 1 foot; a cloud region at about 80 milliseconds about 8 feet; congested cellular at about 500 milliseconds about 51 feet; and in a network outage no decision arrives at all. Latencies are illustrative. DISTANCE A CAR AT 70 MPH COVERS WAITING ON THE NETWORK On-board compute 0 ft added Metro edge, ~10 ms ~1 ft Cloud region, ~80 ms ~8 ft Congested cellular, ~500 ms ~51 ft Network outage no decision Added round-trip delay only; illustrative latencies. 70 mph = 103 ft/s.
Driving decisions
On-board
China car data
Stored in China
Framework layer
Fifth
Hub
Autonomy

Finding

Every production autonomy system already runs its driving decisions on board; data autonomy is now being decided by governments, which require car data to stay inside their borders

Waymo, Tesla FSD, Baidu Apollo Go, Pony.ai and every humanoid program run perception and planning on the vehicle or robot, because a car at 70 mph covers 8 feet in a typical cloud round trip and stops getting answers in an outage. The cloud handles training, fleet analytics and remote assistance, not the control loop.

The harder question is who controls the data and models. China has required vehicle data collected in China to be stored in China since its 2021 automotive data rules, which is why Tesla built a Shanghai data center. The US Commerce Department's January 2025 connected-vehicle rule bars Chinese and Russian vehicle software from model year 2027 and hardware from model year 2030. The EU AI Act phases in obligations for high-risk systems over 2026-27. Each rule pushes fleets toward in-region training, local storage and auditable models.

Open-weight models make local control practical for plants, hospitals and depots: Meta Llama, Alibaba Qwen, DeepSeek, Mistral and OpenAI's gpt-oss can run on owned hardware. Data sits fifth in the Six Foundation Domains, after materials, silicon, energy and thermal, because local inference needs local compute, power and cooling, and just before Operational Autonomy, because a site cannot run without people if its decisions belong to someone else.

Domains

What data autonomy covers

Data autonomy domains
DomainWhat it includesNamed examples
Inference controlOn-device, edge and on-premise inference for critical functionsTesla AI4, Nvidia DRIVE Thor, Horizon Journey 6, Nvidia Jetson Thor in robots
Model sovereigntyOwned, open-weight or auditable models with known lineageMeta Llama, Alibaba Qwen, DeepSeek, Mistral, OpenAI gpt-oss
Data-flow controlWhat leaves the site, under what retention and training termsIn-country storage (Tesla Shanghai); private cloud regions
AuditabilityLogging, decision traceability, validation recordsEvent data recorders; safety cases under UL 4600 and ISO 21448
Training-loop governanceApproved retraining, dataset provenance, staged OTA updatesFleet OTA release gates; shadow-mode validation

Rules

Data rules shaping autonomy

Data sovereignty rules affecting autonomous systems
JurisdictionRuleEffect on fleets
ChinaAutomotive data security provisions (2021); data export security reviewsVehicle data stored and trained in China; foreign OEMs build local data centers
United StatesCommerce connected-vehicle rule (January 2025)Chinese and Russian vehicle software barred from MY2027, hardware from MY2030
European UnionAI Act (in force August 2024); GDPR; Data ActHigh-risk AI obligations phase in over 2026-27; data portability for connected products
United StatesExport controls on advanced AI chipsLimits which compute China-based training can use

Readiness

Data readiness bands

ElectronsX rating scale.

Data autonomy readiness bands
BandPrimary inferenceCharacteristics
DA-0 DependentCloud onlyProprietary remote models; operations stop without the vendor
DA-1 AwareCloud, some edgeLatency-critical functions local; risk mapped
DA-2 HybridCritical functions localAuditable models for sensitive work; data flows documented
DA-3 AutonomousLocal or federatedOwned weights, controlled training data; cloud only for auxiliary tasks

Sources

Sources and scope

Scope
Data Autonomy
Sources
CAC automotive data provisions (2021); US Commerce BIS connected-vehicle rule (January 2025); EU AI Act; company disclosures
Updated
October 2026