When not driving, electric vehicles sit idle for most of the day. According to CATL chairman Robin Zeng, those idle hours could become the new frontier of distributed computing. In an interview reported by Bloomberg, the executive claimed that parked EVs could act as “AI token factories,” using their onboard computers to run inference for large language models.

Hidden compute in parking lots

The idea builds on a simple fact: modern EVs pack increasingly powerful processors, required for assisted driving and infotainment. While the car is parked or charging, those chips go unused. By linking them into an orchestrated compute network, you could create a massive distributed resource capable of generating tokens — the basic unit of processing in language models — on behalf of central servers.

This isn't science fiction. Platforms like BOINC or Folding@Home already show that distributed computing on consumer devices works at a global scale. Applying the same principle to EVs would require an orchestration layer able to dispatch batches of inference requests to available nodes, balancing latency, energy consumption, and intermittent connectivity.

From batteries to bytes: the technical challenge

Turning a car into a token factory raises multiple technical questions. Embedded vehicle systems are not designed for sustained LLM inference: thermal dissipation, component wear, and battery degradation must be carefully managed. Moreover, network bandwidth — typically cellular or Wi-Fi — could become a bottleneck for large models, unless aggressive quantization or heavily optimized on-device models are used.

CATL's proposal, however, leans on a critical point: the marginal cost of that compute is close to zero, because the hardware is already paid for and the energy comes from the charging grid. For a fleet operator, the ROI could be immediate, transforming a fixed cost into a productive asset.

Data sovereignty on four wheels

For those evaluating on-premise deployment, the vehicle-as-compute concept adds a novel dimension. A company with an EV fleet could theoretically process sensitive data directly on its own vehicles, without sending anything to the public cloud. This fleet-edge approach would give full control over processing locations, reducing compliance risks and satisfying regulations like GDPR. Paired with quantized models optimized for mobile hardware, it could yield a local, geo-distributed inference infrastructure with low operational cost.

Of course, security becomes paramount: every vehicle is an exposed endpoint, and data and model integrity must be secured with end-to-end encryption and hardware attestation. But the sovereignty and TCO gains are significant, especially for use cases like predictive maintenance, intelligent logistics, and mobile assistance.

Beyond the cloud: a distributed vision

Zeng's statement signals a broader paradigm shift. Not just centralized server farms or static edge devices, but entire mobile fleets participating in an organization's total compute capacity. For AI-RADAR readers, this is a natural progression of the on-premise conversation: infrastructure is no longer just the rack in the server room, but anything with a processor, a connection, and a power supply. The line between data center and vehicle fleet blurs, and data sovereignty literally moves on wheels.