This is Part 3 of The Uptime Equation. Part 1 covered vehicle context, and Part 2 covered the hardware-or-software trade-off. This part covers where the work should run.
A long-haul truck crosses a stretch of highway where the cellular signal drops out. Somewhere along that stretch, an intermittent electrical fault appears for 90 seconds and then clears. If detection depends on the cloud, one of two things happens. The event is missed, or a delayed snapshot arrives after the relevant signals have already been overwritten. The next time the issue arises, it might be in a shop with no record of the conditions that previously caused it.
This is the field version of the re-drive problem described in The Agentic Loop. In commercial vehicles it happens more often, because of where and how these vehicles operate.
Connectivity is not a given
Coverage maps are optimistic. In a 2019 FCC staff investigation, field agents ran 24,649 drive tests over nearly 10,000 miles in 12 states. Only 62.3% of the tests reached the download speeds that carriers’ maps predicted. Networks have improved since 2019, but the lesson holds: a map showing coverage doesn’t mean a truck will have a signal at the moment a fault happens. Off-highway equipment on mine sites, farms, construction sites, and ports operates in places coverage planners rarely prioritize. Fleets already plan around this. Drivers wait until they reach a yard to send logs, technicians pull data by hand when a truck comes in, and a truck that goes quiet on the road is treated as normal rather than a warning. Each workaround costs uptime, and none of them catches a 90-second fault on an empty stretch of highway.
That uncertainty makes offline operation a design requirement, not just a fallback. In McKinsey’s survey on edge AI in automotive, 39% of respondents cited offline availability as a key requirement for all AI use cases, ahead of lower latency (35%) and data privacy and security (20%). Only 6% cited network data cost. The case for the edge is less about the cellular bill and more about what the cloud can’t see.
Volume: the issue is signal-to-noise, not the data bill
Commercial vehicle networks produce data continuously. A single SAE J1939/14 bus can run at up to 500 kbit/s, many heavy trucks run more than one, and newer platforms add Automotive Ethernet, much of it carrying camera feeds for ADAS and safety systems. Over a full driving day, that is far more data than is worth sending, and most of it only confirms that nothing went wrong.
As the Agentic Loop outlines, always-on logging is “expensive to transmit over cellular networks, expensive to store, and expensive to work through.” The alternative is to decide — on the vehicle — what data matters:
- Event-triggered capture. Record the signals before and after an event, rather than everything all the time. With Fastlane™ Edge Smart Triggers, for example, a lightweight anomaly-detection model on the vehicle starts a Fastlane™ Collector policy only when it detects abnormal behavior, such as an engine running outside its normal range, so the vehicle captures the signals around the abnormal event instead of a continuous stream.
- On-vehicle compression. Fastlane™ Collector cites up to 28× reduction in data payload size through time-series compression.
- On-vehicle history. A time-series database on the vehicle keeps weeks or months of recent signal history. If a fault happens in a dead zone, the context is still there when the truck reconnects, and engineers can pull the exact window they need instead of waiting for the fault to come back.
- Filtering before upload. Send summarized, relevant results instead of raw streams. In one partner case study on in-vehicle intrusion detection, processing locally and uploading only relevant alerts cut cloud costs by 60%.
Latency: some decisions can’t wait
Some actions — protecting a component, adjusting a system’s operation, alerting a driver, or flagging a possible cyber intrusion — need to happen within seconds. That makes the vehicle the natural place to handle them, though latency alone doesn’t determine where intelligence should run. McKinsey notes that the edge doesn’t always offer a decisive latency advantage. For many diagnostic use cases, availability and context matter more than milliseconds.
None of this makes the cloud less important. It does the work a single vehicle can’t:
- Fleet-wide pattern detection. A failure mode can look like noise on one truck and be an obvious issue across thousands.
- Root-cause investigation. It combines vehicle data with engineering documents, service history, and past cases. Fastlane™ Insight does this in the cloud, using AI to correlate these sources and recommend likely root causes.
- Model training and validation. This uses the context-rich events captured at the edge.
- Distribution. Updated models and collection policies are pushed back to vehicles over the air.
The edge detects and captures, the cloud reasons and learns, and the edge then acts on what the cloud learned. Fastlane™ Edge is built around that loop: it runs anomaly detection on the vehicle and receives updated models over the air as the cloud learns. Running intelligence at the edge reduces dependence on connectivity, while learning across the fleet keeps improving vehicles already on the road.
Drawing the line
A useful rule: put the work where the data and the deadline are.
| On the vehicle | In the cloud |
| Continuous monitoring and anomaly detection | Pattern detection across the fleet |
| Event-triggered capture and on-vehicle history | Root-cause investigation with service and engineering data |
| Time-critical actions | Model training and validation |
| Filtering and summarizing before upload | Model and policy distribution |
If a decision depends on one vehicle’s data in the moment, it belongs on the vehicle. If it depends on many vehicles’ data over time, it belongs in the cloud. Capturing context on the vehicle also helps mixed fleets. When it’s structured at the source using shared signal definitions such as COVESA VSS, it can reach a fleet’s telematics provider in the same shape from every brand. And market forecasts suggest OEMs are investing in this model. GM Insights estimates that commercial vehicles accounted for $3.59 billion of the automotive edge computing market in 2025, growing about 22% a year.
Once the edge captures the right context, that data becomes the raw material for AI that can predict failures instead of just reporting them. Part 4 picks that up.
The bottom line: Commercial vehicles need intelligence at both ends of the connection. The vehicle handles detection, context capture, and time-critical action. The cloud handles learning across the whole fleet. The recommendation: default those jobs to the vehicle, and keep the cloud for anything that needs more than one truck’s data. Design the two as one loop from the start, so models trained in the cloud flow back to the trucks that captured the data. A fault the truck doesn’t capture is a fault that comes back, and every comeback costs uptime.