
Most vehicle connectivity evaluations compare feature matrices and then discover the cost structure afterwards. Invert it. Build the cost model first, at three payload levels, and the vendor shortlist largely selects itself.
Here is the model and the parameter that dominates it.
The model
Cost per vehicle-month, C, decomposes into four terms that behave differently under scale:
C = S + D(p) + M + R
-
S— connectivity subscription. Per-SIM monthly, roughly $0.30–$2. Effectively fixed. Dominates whenpis small. -
D(p)— data transfer as a function of payloadp. The only usage-dependent term, and the one that must be modelled as a function rather than a constant. -
M— platform / device management fee. Per-device monthly. Frequently exceedsS, and is the term most often missing from first-pass spreadsheets. -
R— roaming and regional premiums. Zero domestically, and capable of exceedingS + D + Mcombined for international fleets.
Two properties of this model matter more than the individual numbers.
D(p) is usually piecewise. Pooled allowance up to a threshold, then overage at a materially worse rate. So you need p_peak, not just p_mean — overage is assessed on peaks, and a fleet with uneven duty cycles has a peak-to-mean ratio well above 1.
Insist on pooling. With per-SIM allowances you pay overage on highway vehicles while leaving allowance stranded on depot-bound ones. Pooling across the fleet is what makes D(p) tractable.
The three payload levels
Evaluate at all three. Do not pick the one matching today's use cases.
| Level | | Workload |
|---|---|---|
| Minimal | ~5 MB | GPS position, basic diagnostics, ignition events |
| Moderate | 50–200 MB | High-frequency CAN bus telemetry |
| Heavy | 2–20 GB | Video, raw sensor streams, uncompressed IMU |
The reason all three matter: vendor rankings invert between minimal and heavy with uncomfortable regularity. A provider with an excellent low-tier bundle and punitive overage wins decisively at 5 MB and is unaffordable at 5 GB. One with flat per-GB pricing and a higher S loses at minimal and wins at heavy.
No vendor volunteers this during a demo, because the demo is built around whichever tier they win.
The parameter that dominates p
Here is the thing that makes the whole exercise tractable, and the reason cost modelling is really an architecture exercise.
p is not a product requirement. It is a consequence of where inference runs, and the two options differ by three to four orders of magnitude.
Consider camera-based driver monitoring:
-
Cloud inference. The vehicle ships frames.
p≈ 2–20 GB/vehicle-month. Plus latency too high to intervene in real time, plus a hard dependency on coverage where coverage is worst. -
On-device inference. The vehicle ships events:
{class, confidence, timestamp, optional keyframe}.p≈ 1–50 KB/vehicle-month for the same function.
Identical business outcome. D(p) differs by ~1000×.
So the correct sequence is:
-
Fix inference placement for the use cases on your 18-month roadmap, not only those live today.
-
Derive
p_meanandp_peakfrom that placement. -
Compute
Cfor every candidate at all three levels, withRper country of operation. -
Only then compare features.
Teams that run steps 1–3 after signing are the ones rearchitecting eighteen months later.
What edge inference costs you instead
Pushing inference to the vehicle does not make cost disappear — it relocates it, and the relocation is worth pricing explicitly.
You now own a continuous model delivery pipeline to physical devices. Concretely:
-
OTA infrastructure with staged rollout by cohort, atomic swap and rollback that works on a vehicle mid-route.
-
Model versioning on every event, or analytics silently blend generations during any rollout and a metric step change becomes unattributable.
-
Asymmetric upload policy. Retraining needs periodic high-resolution raw data from a rotating subset of vehicles; inference needs events from all of them. Plans priced for a uniform per-vehicle profile force you to either overprovision the fleet or starve training.
And the metric that decides whether this is viable: OTA success rate measured under real conditions. 94% reads fine and means ~600 physical service visits on a 10,000-vehicle fleet. Put that number in the model as a cost, because it is one.
The term that is not in the model
C omits the cost that most reliably consumes fleet programmes: normalisation.
Vehicle data is not a standard. Mixed fleets mean multiple CAN dialects, manufacturer-specific PIDs, model-year variation within a single make, and in heavy vehicles a mixture of J1939, OBD-II and proprietary extensions. EVs add battery management data with almost no cross-manufacturer consistency.
The evaluation question is not "do you support our vehicles" — every vendor says yes. It is: which specific parameters do you normalise, per make and per model year we operate, and what happens when we add an unfamiliar make? Ask for the list. A vendor who cannot produce one does not have one.
If you plan to build it, model it as a standing commitment in person-quarters per year rather than a project, because every model year changes something. For most operators, buying normalisation and owning the analytics above it is correct. For telematics providers reselling to others, the normalisation layer frequently is the product.
Sanity-checking the model in production
Once running, one metric tells you whether a workload is misplaced: connectivity cost should scale roughly linearly with vehicle count. When it scales superlinearly, a high-compression-ratio or cross-vehicle workload is sitting in the cloud, and your invoice history usually identifies the quarter it moved there.
Full treatment — the three-tier architecture, roaming legality, security obligations under UNECE R155/R156, and the lock-in terms to secure before signing: How to Choose the Top IoT Connectivity Platform for Vehicles. We build this layer as IoT app development.
Frequently Asked Questions
What are the four components of vehicle connectivity cost?
Per-SIM connectivity subscription (roughly $0.30–$2, effectively fixed), data transfer as a function of payload, a per-device platform or management fee that frequently exceeds the subscription, and roaming or regional premiums. The third is the one most often missing from first-pass models.
Why model three payload levels instead of one?
Because vendor rankings invert between them. A provider with a cheap low tier and punitive overage wins at 5 MB per vehicle-month and is unaffordable at 5 GB, while flat per-GB pricing loses at minimal volume and wins at heavy. Evaluating only today's payload selects for a vendor who may be wrong in a year.
Should I model average or peak payload?
Both, and overage specifically against peak. Data pricing is typically piecewise — a pooled allowance then a materially worse overage rate — and fleets with uneven duty cycles have peak-to-mean ratios well above one. Modelling the mean alone systematically understates cost.
Why does inference placement dominate the cost model?
Because payload is a consequence of it rather than an independent requirement. Camera-based driver monitoring run in the cloud ships 2–20 GB per vehicle-month; the same function run on-device ships kilobyte-scale events. That is roughly a thousandfold difference in the only usage-dependent term in the model.
What costs does edge inference add back?
A continuous model delivery pipeline: OTA infrastructure with cohort staged rollout and mid-route rollback, model version tagging on every event, and an asymmetric upload policy so a rotating subset can send training data. Include the measured OTA failure rate as a cost — 6% failure means hundreds of physical service visits at fleet scale.
How do I tell from production data that a workload is in the wrong tier?
Connectivity cost should grow roughly linearly with fleet size. Superlinear growth indicates a high-compression-ratio or cross-vehicle workload running in the cloud that belongs on the vehicle or at a regional edge, and the invoice history usually pinpoints when it was introduced.