How to Choose the Top IoT Connectivity Platform for Vehicles
Choosing the top IoT connectivity platform for vehicles looks like a telecom decision and is really a data architecture decision. Where you run inference determines your payload, your payload determines your cost per vehicle-month, and most teams discover that order backwards. Here is how to evaluate properly.

Almost every evaluation of the top IoT connectivity platform for vehicles starts in the wrong place. A team assembles a feature matrix — global SIM coverage, eSIM provisioning, data plan tiers, device management console, dashboard quality — scores four vendors, and signs with whichever won. Eighteen months later the same team is rearchitecting, because the real constraint turned out to be something the matrix never asked about: how much data each vehicle actually needs to move, and that number is determined by where the machine learning runs.
Connectivity platform selection is downstream of inference placement. Decide first whether your models run in the vehicle, at a regional edge, or in the cloud, and the connectivity requirement falls out of that decision almost deterministically. Decide connectivity first and you will spend the next two years designing your data pipeline around a contract someone signed before the problem was understood.
This guide is written for the person who has to own that architecture — a CTO at a fleet operator, a VP Engineering at a telematics provider, a platform lead at an OEM or a mobility startup. It covers the layers a vehicle connectivity stack actually has, the cost model that governs it, how edge AI changes the arithmetic, and a concrete evaluation sequence.
What "connectivity platform" actually means — four distinct layers
Much of the confusion in this market comes from vendors describing four different products with the same phrase. Separate them before you compare anything.
- The connectivity layer — the SIM, eSIM or iSIM, the carrier relationships, roaming agreements, and the network technology, whether that is LTE Cat-1 and Cat-M, NB-IoT, 5G, or a satellite fallback. This is what MVNOs and connectivity management platforms sell.
- The device management layer — provisioning, over-the-air firmware updates, remote configuration, certificate lifecycle, and fleet-wide device health. This is where most of the operational pain lives and where most evaluations spend the least attention.
- The data ingestion and processing layer — how telemetry arrives, how it is parsed and normalised across vehicle makes and model years, where it is buffered, and how it reaches your systems.
- The application layer — dashboards, driver scoring, maintenance alerting, geofencing, compliance reporting. Highly visible, frequently the reason a platform gets chosen, and almost always the layer you will eventually replace with your own.
The strategic error is letting a strong application layer sell you a weak connectivity and data layer. Dashboards are the easiest part to rebuild and the easiest part to demo. Carrier economics and normalisation logic are the hard parts and the ones that will constrain you for years. Evaluate bottom-up, not top-down.
The cost model that actually governs the decision
Vehicle connectivity cost is not a per-megabyte question. Model it as cost per vehicle per month, decomposed into four components that behave very differently as you scale.
- Connectivity subscription — a per-SIM monthly fee, often 0.30 to 2 US dollars, which is largely fixed regardless of usage and which dominates at low data volumes.
- Data transfer — usually pooled across the fleet, with overage rates that are the single most common source of budget surprise. Pooling is what makes it survivable; unpooled per-SIM allowances are a trap for fleets with uneven duty cycles.
- Platform or device management fee — a per-device monthly charge for the management plane, frequently larger than the connectivity cost itself and frequently overlooked in early modelling.
- Roaming and regional premiums — the component that destroys international fleet budgets. Permanent roaming is restricted or prohibited in several jurisdictions including Brazil, Turkey, India and China, which means a single global SIM strategy is often not legally available and localisation is required.
Build the model at three payload levels before you talk to vendors: a minimal case of roughly 5 MB per vehicle-month for position and basic diagnostics, a moderate case of 50 to 200 MB for high-frequency CAN bus telemetry, and a heavy case of 2 to 20 GB for anything involving video or raw sensor streams. The vendor ranking frequently inverts between the minimal and heavy cases. A platform that is excellent at 5 MB can be unaffordable at 5 GB, and the vendor will not volunteer this.
Why inference placement decides everything
Here is the dependency most feature matrices miss entirely. The gap between the minimal and heavy payload cases above is not a product decision — it is a direct consequence of where your models execute.
Run inference in the cloud and the vehicle becomes a sensor that ships raw data. A dashcam feeding a cloud vision model for driver monitoring moves gigabytes per vehicle-month. At a thousand vehicles that is a telecom bill with a seven-figure annual run rate, plus latency that makes real-time intervention impossible, plus a hard dependency on coverage in places where coverage is unreliable.
Run inference at the edge — on the telematics unit, a vehicle gateway, or an in-cab device — and the vehicle ships events rather than data. The same driver monitoring use case becomes a few kilobytes per event: a classification, a confidence score, a timestamp, perhaps a single compressed keyframe for audit. The payload drops by three to four orders of magnitude, the latency becomes low enough to actually warn a driver, and the system degrades gracefully when connectivity drops because detection no longer depends on the network.
The practical consequence for platform selection is that these two architectures want different vendors. Cloud inference needs aggressive pooled data pricing, high-throughput ingestion, and strong 5G coverage. Edge inference needs excellent over-the-air update tooling, because you will be shipping model updates to vehicles for the rest of the product's life, and it needs a device management plane that can handle staged rollouts and rollback of a model that misbehaves in the field.
That second requirement is the one teams consistently underestimate. Shipping a model to ten thousand vehicles is not a firmware update — it is a continuous delivery problem with physical consequences and no easy rollback if a vehicle is mid-route. If your roadmap includes on-vehicle models, the quality of the platform's OTA and staged-rollout capability matters more than any line item in its data pricing. We cover the broader pattern in our work on IoT app development for business and on computer vision for business.
The hybrid pattern that most mature fleets converge on
In practice the durable architecture is neither pure edge nor pure cloud. It is a three-tier split, and it is worth designing for deliberately rather than arriving at it accidentally.
- On-vehicle — safety-critical and latency-sensitive inference that must work without a network: collision warning, driver distraction detection, harsh-event classification, local anomaly detection on engine parameters.
- Regional edge — aggregation across vehicles that does not need to reach a central cloud: route-level traffic inference, depot-scoped optimisation, and the first pass of data reduction before anything crosses an expensive link.
- Cloud — model training and retraining, fleet-wide analytics, long-horizon maintenance prediction, and anything requiring a view across the whole estate.
Designing for this split from the start changes your platform requirements concretely: you need a connectivity layer that supports differentiated quality of service, a device management plane capable of pushing different model versions to different vehicle cohorts, and an ingestion layer that accepts both high-frequency raw telemetry and low-frequency event summaries without forcing both through the same expensive path.
Normalisation is the hidden cost centre
Vehicle data is not a standard. A mixed fleet means multiple CAN bus dialects, manufacturer-specific parameter identifiers, model-year variation within the same make, and in heavy vehicles a mixture of J1939, OBD-II and proprietary extensions. Electric vehicles add battery management data with almost no cross-manufacturer consistency.
Some platforms do this normalisation for you and the quality varies enormously. Some hand you raw frames and leave it as your problem. The honest question to put to a vendor is not whether they support your vehicles but which specific parameters they normalise for each make and model year you operate, and what happens when you add a make they have not seen. Ask for the list. A vendor that cannot produce it does not have one.
Price the alternative before you decide it is cheaper. Building and maintaining your own normalisation layer for a mixed fleet is a standing engineering commitment measured in person-quarters per year, not a one-time project, because every new model year changes something. For most operators the right answer is to buy normalisation and own the analytics on top of it. For telematics providers reselling to others, owning the normalisation layer is frequently the actual product.
Coverage: the questions that matter and the ones that do not
Coverage maps are marketing artefacts. Three questions produce more signal than any map.
First, what happens at the edges of coverage rather than outside it. Degraded connectivity — high latency, high packet loss, repeated reattachment — is far more common than total absence and far more damaging, because naive clients retry aggressively and burn both data allowance and device battery. Ask how the platform's agent behaves under 20 percent packet loss. A good answer involves local buffering with backpressure and exponential backoff. A vague answer means you will discover the behaviour in production.
Second, which carriers specifically, in which countries, under what roaming agreement, and with what permanent-roaming restrictions. Multi-carrier eSIM with local profile switching is the correct architecture for anything crossing borders, but verify that profile switching is automatic and tested rather than a manual operation on a per-vehicle basis.
Third, how the device stores and forwards. Every vehicle will spend time without usable connectivity — underground parking, rural routes, tunnels, ferry crossings. The platform must buffer locally with a defined retention window and forward in order with deduplication on reconnect. Ask for the buffer size in both megabytes and hours at your telemetry rate, and ask what is dropped first when it fills.
Network technology selection without the marketing
The 5G narrative has distorted a decision that is usually straightforward. Match the technology to the payload and the device lifetime, not to the roadmap slide.
- LTE Cat-1 and Cat-1 bis — the sensible default for most telematics. Adequate throughput for event and telemetry workloads, broad global coverage, mature and inexpensive modules, and long carrier support horizons.
- LTE-M and NB-IoT — for low-power, low-payload, long-life devices such as trailer and asset trackers. Not suitable for anything with meaningful throughput requirements, and NB-IoT in particular handles mobility poorly, which matters in a moving vehicle.
- 5G — genuinely necessary only for high-bandwidth continuous workloads such as multi-camera streaming or remote operation. Carries higher module cost and power draw, and coverage is still patchy along the rural corridors where commercial fleets actually operate.
- Satellite — increasingly viable as a fallback rather than a primary link, and the right answer for maritime, mining, agricultural and remote-corridor operations. Price it as insurance on a subset of the fleet, not as a fleet-wide plan.
A note on device lifetime that catches people out: a telematics unit installed today may be in service for seven to ten years. 2G and 3G sunsets have already stranded large deployments, and further LTE band retirements are scheduled through the 2030s. Check the module's certified bands against the retirement schedules in every country you operate in, and treat the modem as a replaceable component rather than a permanent one.
Lock-in and exit: what to establish before signing
Vehicle connectivity contracts are unusually sticky because switching involves physically touching vehicles. Establish four things in writing before you commit.
- SIM portability and unlocking. Are the SIMs locked to this provider, and what is the process and cost to move the fleet to another? A provider that will not answer plainly has given you the answer.
- Raw data access. Can you export complete historical telemetry in a documented format, at any time, without a fee that scales with volume? Normalised-only export is a lock-in mechanism.
- Device firmware ownership. If the provider supplies the hardware, who can sign firmware? If only they can, you cannot change providers without changing devices, which means the real switching cost is a fleet-wide hardware replacement.
- Contract term against hardware amortisation. A three-year connectivity commitment on devices with a seven-year life is a reasonable trade. A seven-year commitment is not, because the technology and the pricing will both move.
The strongest structural defence is to keep your own data store as the system of record. Stream telemetry into infrastructure you control and treat the platform as a transport and management layer rather than a database. This costs a little more to build and converts a provider change from a migration project into a configuration change. Our custom software development practice treats this as the default pattern for connected-product work.
Security requirements that are genuinely non-negotiable
A connected vehicle is a networked computer attached to a moving physical object, and the regulatory environment has caught up. UNECE R155 obliges a certified cybersecurity management system for vehicle type approval in the European Union, Japan, South Korea and other markets, and R156 does the same for software update management. If you are an OEM or a tier-one supplier these are compliance gates, not best practices.
- Mutual TLS with per-device certificates, and a documented, tested rotation and revocation process. Shared credentials across a fleet are disqualifying.
- Secure boot and signed firmware, with the signing key held by you or under a clearly defined joint arrangement.
- Private connectivity to your cloud — a private APN or equivalent, with no device reachable from the public internet.
- Command authorisation and audit for anything that actuates — remote immobilisation, door unlock, charge control. Every such command needs an authenticated actor and an immutable log entry.
- A tested revocation path for a compromised device, exercisable within hours rather than at the next service interval.
Ask to see the results of a third-party penetration test of the device agent and the management plane. A platform selling into automotive that cannot produce one is not ready for a safety-adjacent deployment.
A concrete evaluation sequence
Run these five steps in order. The order is the point — each step constrains the next, and reversing them is how teams end up rearchitecting.
- Fix your inference placement first. Decide what runs on the vehicle, at the regional edge and in the cloud, for the use cases on your eighteen-month roadmap rather than only the ones live today.
- Derive the payload envelope. Convert that placement into bytes per vehicle-month at the minimal, moderate and heavy levels, and state the peak as well as the average, because overage is priced on peaks.
- Model total cost per vehicle-month across all four cost components and all three payload levels, for every candidate, including roaming for every country you operate in. Rankings commonly invert between levels.
- Pilot on real vehicles in bad conditions. Twenty to fifty vehicles for six to eight weeks, deliberately including the worst routes you operate — tunnels, borders, rural corridors, underground depots. Measure data loss, buffer behaviour, OTA success rate and time-to-detect for a deliberately bricked device.
- Verify exit before signing. Get SIM portability, raw data export, firmware signing and term alignment in the contract, not in an email from a sales engineer.
The pilot step is the one teams skip under time pressure and the one that reliably surfaces the problems a procurement process cannot. An OTA success rate of 94 percent looks acceptable in a slide and means six hundred vehicles out of ten thousand need a physical visit, which is a different business entirely.
Where AI is genuinely changing fleet economics
Beyond inference placement, three applications are producing measurable returns rather than pilot-stage promise, and each has a connectivity implication worth planning for.
Predictive maintenance has moved from threshold alarms on individual parameters to models trained on multivariate sequences — vibration, temperature, current draw and duty cycle together. The connectivity implication is that training needs periodic high-resolution data from a sample of vehicles, while inference needs only continuous low-resolution data from all of them. Plan for a deliberately asymmetric data policy: a rotating subset of the fleet uploading rich data for retraining, the remainder uploading summaries.
Driver behaviour analysis has moved on-device for a straightforward reason — cloud round trips are too slow to intervene and too expensive to run continuously. The implication is that your platform must support model versioning per vehicle cohort, because you will be running an A and B comparison of behaviour models across the fleet whether or not you planned to.
Agentic operational workflows are the newest and most interesting. Rather than alerting a human that a vehicle will likely fail a component within 1,500 kilometres, systems now check parts availability, find a service slot compatible with the vehicle's scheduled routes, and propose a rescheduling of affected loads. That requires the telemetry platform to be addressable by other systems — a real API with webhooks, not a dashboard with CSV export. Weight API quality accordingly; we build this class of system as agentic workflow development, and the integration layer is usually where it succeeds or fails.
Build, buy, or assemble
Three honest positions, with the conditions under which each is correct.
Buy a full stack from a single telematics provider when you operate fewer than roughly five hundred vehicles, your use cases are standard compliance and basic efficiency, and connected data is not part of your product. The premium you pay is less than the engineering you avoid.
Assemble — connectivity from a management platform, devices from a hardware vendor, your own ingestion and application layer — when you operate at scale, when telemetry informs a product your customers touch, or when your differentiation depends on models you own. This is the right position for most mobility startups and most operators above a few thousand vehicles.
Build the connectivity layer itself only if you are becoming a telematics provider. Carrier relationships and regulatory work across jurisdictions are a business, not a feature, and the firms that do it well do nothing else.
The common mistake is assembling without the engineering capacity to own the assembly, which produces a stack nobody fully understands and a normalisation layer that silently rots. If you are choosing the assemble path, staff for it deliberately — or bring in a partner who has done it before. If you want a second opinion on an architecture or a shortlist you are already holding, talk to us and we will tell you where we think it breaks.
Frequently Asked Questions
What is the best IoT connectivity platform for vehicles?
There is no single best platform, because the ranking inverts depending on your data payload. Platforms optimised for low-payload telematics of roughly 5 to 50 MB per vehicle-month are frequently unaffordable for video or raw sensor workloads in the gigabyte range, and vice versa. Decide where your inference runs first, derive your payload envelope from that, then compare total cost per vehicle-month at your actual numbers rather than against a feature matrix.
How much does vehicle IoT connectivity cost per vehicle per month?
For low-payload telematics, expect roughly 1 to 5 US dollars per vehicle-month all in, combining the SIM subscription, pooled data and the device management fee. Video-enabled or high-frequency sensor workloads commonly run 15 to 60 dollars or more. International operation with roaming premiums and required local profiles can double either figure, so model it country by country rather than as a blended global rate.
Should vehicle AI inference run at the edge or in the cloud?
Safety-critical and latency-sensitive inference belongs on the vehicle, because a cloud round trip is too slow to intervene and fails when connectivity does. Model training, fleet-wide analytics and long-horizon prediction belong in the cloud. Most mature deployments run a three-tier split across vehicle, regional edge and cloud. The decision is not primarily about accuracy — it determines your data payload and therefore your connectivity cost, by three to four orders of magnitude.
Is 5G necessary for connected vehicles?
Rarely. LTE Cat-1 or Cat-1 bis handles the overwhelming majority of telematics and event-based workloads at lower module cost, lower power draw and much broader coverage. 5G is genuinely required for continuous high-bandwidth streaming such as multi-camera video or remote vehicle operation. Given that a telematics unit may stay in service for seven to ten years, verify band support against published network retirement schedules in every market you operate in.
What is permanent roaming and why does it matter for fleets?
Permanent roaming is a device operating indefinitely on a foreign carrier's network rather than transiently. Several countries including Brazil, Turkey, India and China restrict or prohibit it, which means a single global SIM strategy is often not legally available. The practical answer is multi-carrier eSIM with automatic local profile switching, and the question to verify is whether switching is automated and tested at fleet scale or a manual per-vehicle operation.
How do I avoid vendor lock-in with a telematics platform?
Secure four things in the contract: SIM portability with a stated unlocking process and cost, raw historical data export in a documented format without volume-scaled fees, firmware signing rights you hold or jointly control, and a contract term shorter than your hardware amortisation period. Structurally, stream telemetry into a data store you own and treat the platform as transport and management rather than as your system of record.
What does UNECE R155 require for connected vehicles?
R155 requires a certified cybersecurity management system covering the vehicle lifecycle as a condition of type approval in the European Union, Japan, South Korea and other adopting markets, and the companion R156 requires a software update management system. For OEMs and tier-one suppliers these are compliance gates rather than guidance. Fleet operators are affected indirectly, through what their suppliers must demonstrate, so ask any platform vendor what they can evidence.
How long should a connectivity platform pilot run before committing?
Six to eight weeks across twenty to fifty vehicles, deliberately including the worst conditions you operate in — tunnels, underground depots, rural corridors and border crossings. Measure data loss, local buffer behaviour under sustained packet loss, over-the-air update success rate, and how long it takes to detect and recover a deliberately bricked device. A 94 percent OTA success rate reads fine on a slide and means hundreds of physical service visits at fleet scale.
Can I use one platform for a mixed fleet of different makes and electric vehicles?
Technically yes, but normalisation quality is where platforms genuinely differ. Mixed fleets involve multiple CAN dialects, manufacturer-specific parameter identifiers and model-year variation, and electric vehicles add battery management data with very little cross-manufacturer consistency. Ask the vendor for the specific list of parameters they normalise per make and model year you operate, and what happens when you add an unfamiliar make. If they cannot produce the list, they do not have one.