A vehicle camera system can keep its footage on the vehicle, send it to a server, or split the work. The split is the settled answer in this market, written into the same transport standards the rest of this series walks page by page. The trade-offs still need stating, because card sizes, data plans and cloud fees are all specification lines a fleet signs. The question is what each class of footage is for. The answer places each class where it serves that purpose at the lowest workable cost.
The two pure designs sit at the ends. Pure local recording keeps every frame on the vehicle’s card. Pure cloud recording sends every frame to a server as it is captured, the design home and shop security cameras use on fixed broadband connections. Between them sits the hybrid the transport standards define: continuous recording stays local on the card, alarm evidence goes up at once as parcels, stored footage goes up on request through the playback query.
The trade-offs divide along four lines. Cost weighs a one-time card against recurring uplink and cloud fees. Capacity measures the radio link against the recording tier, which it runs an order of magnitude below. Access counts how many readers each side can serve at once. Survival asks what happens to each copy when fire, theft or submersion takes the vehicle.
Each line carries its own arithmetic. The four together turn into a placement answer by footage class, then by fleet type, then into the specification language a buyer writes into four documents.
The split serves each class at its cheapest workable place.
The radio link cannot carry the recording. That single fact settles the larger part of the argument before money enters it. A multi-camera terminal lays down tens of gigabytes per day at recording quality, on the figures the consumption pages of this series price per channel. Real 4G uplink in working cells carries single-digit to low-double-digit megabits per second, the band the dropout page quotes. The recording tier’s full output, run through that uplink, would occupy the link continuously and still fall behind on bad cells, before any other vehicle in the cell asks for capacity.
The cell itself ends the argument a second time, for fleets that share roads. Uplink capacity in a cell is pooled across every user in it. Ten vehicles of one fleet in one depot cell, each pushing recording-tier upload, would contend with each other before any other traffic. The supervisory tier was sized so that whole depots coexist in ordinary cells, the design the dropout page’s congestion section assumes. A pure cloud fleet would congest itself at scale, with no carrier change able to fix the geometry.
The data plan ends the same argument a third time. A fleet SIM carries a monthly pool in the gigabytes, sized on the billing pages around supervisory streams and alarm parcels. Continuous full-quality upload multiplies that pool by orders of magnitude. The plan that covers it does not exist on fleet tariffs, the reason no vehicle platform in this market is built as a pure cloud recorder.
The ratio puts numbers on the impossibility. Four channels at recording quality produce on the order of four megabits per second around the clock. A working uplink carries that only in good cells, with nothing left for the supervisory stream, the alarms or any second vehicle. The same four channels held to the supervisory tier of a few hundred kilobits fit comfortably, the reason the live view works at all. Recording-tier upload and watch-tier upload differ by an order of magnitude, the gap the whole placement question stands on.
The asymmetry of the radio link sharpens the same point. Cellular networks allocate more capacity downward than upward, because consumer traffic watches more than it sends. A vehicle camera system runs its heavy traffic upward, into the direction the network gives least. That scarce upward capacity is one more structural reason the upload has to stay an exception. The supervisory stream, the parcels and the pulls all fit inside the thin upward lane.
The capacity line writes the first placement rule on its own: the continuous recording lives where it is produced. The card sits beside the encoder with no link in between, takes the full rate all day and answers to nobody’s coverage map. Everything else in the design is a decision about which selections from that recording cross the link and on what schedule.
The local side costs once. A high-endurance card at fleet capacities costs tens of dollars and carries an endurance rating in the tens of thousands of recording hours, on the published figures the recording-mechanism page cites. Amortised across its rated life, the card’s cost per recorded hour sits near zero. The card adds nothing in the months it is not read, which is the normal fate of routine footage. The replacement schedule is the only recurring element on the local side, one card per rated life per vehicle, logged in the service plan the recording page describes. Read honestly, the local meter is a slow drip of hardware on a schedule the fleet controls. Cloud storage runs two meters at once. The first is the uplink: every gigabyte sent to a server crosses the data plan, at the per-gigabyte reality the billing pages price. The second is the storage itself: published cloud surveillance pricing runs from a few dollars per camera per month at light settings to tens of dollars at full resolution and month-long retention. A multi-camera vehicle multiplies that by its channel count, every month, for the life of the contract. Both cloud meters scale with the same variable, gigabytes moved and held, so every configuration choice that adds upload adds twice. The two sides store different things. The card holds everything at its near-zero amortised rate. Every month, the cloud bills only for what was sent up to it. Most recorded footage is never opened, so pushing all of it up would pay both meters at their maximum for nothing. Sending up only the alarm parcels keeps both meters on the slice with a standing reader behind it, the few megabytes per event the consumption page prices against the card’s daily tens of gigabytes. A worked month makes the gap concrete. A four-channel vehicle records roughly forty gigabytes a day to its card, over a terabyte a month, on the series’ standing figures. The same vehicle’s alarm parcels, at a handful of events per month and megabytes per event, total under a tenth of a gigabyte. The hybrid sends up roughly one ten-thousandth of what it records. The two meters bill that fraction and no more, month after month. The card carries the rest at its near-zero amortised rate, the arithmetic that holds the whole design together. The cost line writes the second placement rule: spend recurring money only where a recurring reader waits. The alarm evidence the safety office and the insurer keep returning to earns its server copy. Routine footage waits on the card until an investigation names a date, then the on-request pull serves it without a standing fee.
A server copy answers to anyone with credentials, from anywhere, at once: the safety office in the morning, the insurer’s assessor in another city that afternoon, each reaching the same parcel. Nobody waits for the vehicle to return to the depot. Nobody asks where on the road the vehicle currently is. The reading scales without touching the fleet or its drivers, the property that makes the server side the natural place for evidence under active review by several parties.
Reaching the card takes more. A stored clip on a moving vehicle is reachable two ways: the remote playback query over the link, at the pace and patience the current cell allows, or the card in a reader after the vehicle comes home at the end of its run. Both work. Both serve one reader at a time, on one vehicle at a time. The access cost is time and coordination, the quantity the weekly operations meeting spends.
Access control follows the same split. Server copies live behind the platform’s accounts, roles and logs, the governance the operations procedure writes once for everyone. Card access happens in the physical world, the reason card handling needs its own procedure line: who may pull a card, who logs the chain, where the card sits while the vehicle works.
A worked afternoon shows the difference in clock time. An insurer asks for three minutes of forward camera from last Tuesday. If the parcel went up at the event, the safety office forwards a link in the time it takes to find it. If the minutes sit on a card in a vehicle two provinces away, the office runs a playback query against the vehicle’s next coverage window, waits out the transfer at fringe-cell pace, then forwards the file. Both paths end in the same evidence.
The access line writes the third placement rule: footage moves up when its readers multiply. An alarm under review gets its server copy at the moment it fires, because review is the expected next step. A routine Tuesday afternoon stays on the card until someone names it, then moves once, pulled by query and shared from the server thereafter.
The card shares the vehicle’s fate. A cab fire takes the card with the cab. A submersion, a theft, a crash that destroys the mounting corner takes the recording with it. The events are rare across a fleet year. They are also correlated with exactly the moments the footage matters, because the crash severe enough to destroy a terminal is the crash an investigation is certain to follow. A pure local design can lose its best evidence to the same event it documents.
Theft adds a quieter version of the same loss. A stolen vehicle takes its card wherever it goes, with the terminal’s last reported position as the only lead. The alarm parcels that fired before the theft, the door alarm, the ignition event, the geofence exit, sit on the server with their stills and clips, the evidence the report to the police opens with. The card may return with the vehicle, weeks later or not at all. The case did not wait for it.
The standards’ answer is the alarm parcel. The survival line is its strongest justification. The parcel uploads at the event, within the minutes the link allows, on the queue-and-resume mechanism the recording page walks. By the time a fire reaches the card, the collision parcel sits on the server, stills and clip and record together. The cloud copy here is no small thing. On the worst day the system is built for, it is the whole difference between holding the evidence and holding nothing.
The survival line also has an honest boundary. The parcels cover the flagged moments, a sliver of the day. A catastrophic loss still costs the routine hours around the events, the context an investigation sometimes wants. A fleet whose risk profile cannot accept that gap escalates its alarm coverage, wider pre-event windows, more trigger classes, on settings the platform pushes, raising the protected sliver at a priced uplink cost per event.
The transport standards define the hybrid as the architecture, the reason this page’s question is placement inside it. The terminal standard requires continuous local recording and sets its retention. The protocol family defines the alarm parcel that uploads at the event, the stored-footage query that finds minutes on a card from the centre, the playback session that pulls them and the resumable transfer that survives bad cells. Each trade-off line above has its machinery already specified, tested at docking and priced on the pages of this series.
The design reads as a decision about defaults. Footage stays local by default, held there by the capacity and cost lines. An event or a named request lifts a selection upward, the exception the access and survival lines pay for, and nothing else moves. A fleet tunes the exception rate, alarm sensitivity up or down, pull discipline tight or loose. The architecture holds at any setting.
The pure designs survive only outside this market’s constraints. Pure cloud recording suits only wired, unmetered links, the home and shop camera case. The capacity arithmetic above rules it out on a metered radio. Pure local recording works where nobody reviews remotely and no catastrophic-loss risk is carried, a shrinking set of operations once insurers and regulators start asking for evidence within days. The hybrid is the one design left standing after the constraints are applied.
Continuous recording lives on the card in every case. The capacity line forbids anything else and the cost line confirms the same answer. Its retention is the card’s circle in days, sized on the storage arithmetic against the route’s worst offline-plus-unvisited run, the line the recording page derives.
Alarm evidence lives in both places by design. The original stays in the card’s pinned queue until sent, protected from overwrite, and in the recording circle after. The server copy exists from minutes after the event, survives the vehicle, serves every reader and carries the retention the safety contract names, commonly months to years at the parcel’s small size. The double placement costs megabytes per event and buys the survival and access lines outright.
Pulled footage lives where the pull put it. A stretch retrieved by query for a named investigation lands on the server and inherits its server retention from the case file it joins. The class is small, named and budgeted per case, the discipline line the recording page writes for the retrieval habit.
Event-adjacent context rides with its event. An investigation that opens a collision parcel often wants the minutes around it, the approach, the loading stop before, the lane history. The context is pulled by query against the event’s own times, lands in the same case file and inherits its retention. The case list bounds the class and prices it per investigation on the pull line’s budget.
Position and telemetry live on the server by default, the one class that reverses the rule. The points are bytes each, the uplink cost is noise in the plan, and the readers are every screen on the platform that draws a map. The class shows placement following size and readership. Small data with many readers goes up as a matter of course, the position stream the clearest case.
Live view lives nowhere. The supervisory stream is watched and discarded, a transit class with no storage on either end beyond the moment. Its cost is uplink minutes alone, at the supervisory tier the codec and consumption pages size per channel. A platform that records live views into cloud storage has turned a transit class into a stored class, a configuration error the monthly bill reports before anyone notices it on screen.
A city distribution fleet leans local. Vehicles come home nightly, cards stay physically reachable at the depot, urban coverage is strong for the alarm parcels that do go up. The depot reader covers the rare deep investigation at no uplink cost. The cloud side stays at the standards’ default: alarms up at once, pulls on request. The data plan stays at the supervisory tier the billing pages size.
A long-haul fleet leans harder on the cloud slice. Vehicles stay out for days at a time, cards are physically unreachable, coverage holes queue the parcels for the next usable cell. The alarm coverage widens, the card’s circle is sized for the route, the dual-SIM line keeps the upload path alive. Only the cloud slice grows with route length. The capacity line keeps routine recording local regardless.
A hazardous-goods or escort operation buys the widest cloud slice. Regulators and clients want evidence off the vehicle fast, sometimes streamed live in incident windows. Pre-event windows widen, trigger classes multiply, the uplink tier and the data plan grow to match the contracted duty. The placement framework holds. The exception rate is set high, with each widening priced on the same per-event arithmetic.
A site or construction fleet adds a coverage twist to the local lean. Quarries, basements and new corridors sit in coverage holes for whole shifts, so the alarm parcels queue on the card and arrive in the evening burst the recording page describes. The placement framework needs no change. The card’s circle is sized for the offline days, the queue pins the parcels through them, the cloud copies arrive late and complete. The only line that moves is the card capacity line, priced in tens of dollars once.
A mixed fleet runs all of these on one platform at once. The placement framework is a per-vehicle configuration, so the city vans hold the default profile, the long-haul tractors carry the widened alarm set and the escort vehicles run their high-exception profile, side by side on the same screens. The monthly per-vehicle figures then audit each profile against its own expectations, the reading habit the billing pages set.
A taxi or ride fleet sits between, with privacy as the extra line. In-cab audio and video of passengers carry consent and retention rules that vary by jurisdiction and by contract. The local circle’s short window is an asset there: routine footage expires on its own schedule unless a named event pins it first. The cloud slice carries only event evidence, with retention written from the local rulebook, with no default copied in.
The placement decisions compress into five written lines, each short enough to read aloud at the contract table. The card line: high-endurance, sized in days against the worst run, swapped on rated hours. The alarm line: trigger classes and pre-event windows named, parcel destinations and cloud retention stated in months or years. The pull line: who may request stored footage, at what tier, billed to which case. The live line: supervisory tier by default, full tier on incident, nothing stored. The plan line: the data pool sized to the alarm rate and pull habit, on the billing pages’ arithmetic, reviewed against the per-vehicle figures monthly.
Each line lands in a different document, the reason the compression matters. The card line sits in the hardware specification. The alarm and live lines sit in the platform configuration profile. The pull line sits in the operations procedure. The plan line sits in the carrier contract. A fleet that writes all five has decided the local-versus-cloud question completely, without ever phrasing it as a debate.
Size the card against the route, on the recording page’s line. Keep continuous recording local and confirm the platform does not silently store live views. Name the alarm classes, their pre-event windows and their cloud retention in the contract. Write the pull procedure with a case number on every request. Size the data plan to the resulting event rate and review it monthly per vehicle against the figures. The trade-offs are then settled where they belong, in five lines of paperwork, with the architecture the standards already built doing the rest on every vehicle, every day.
The radio link cannot carry it. A multi-camera terminal produces tens of gigabytes per day at recording quality, against a 4G uplink of single-digit to low-double-digit megabits per second and a fleet data plan pooled in gigabytes per month. Continuous upload would saturate the link, multiply the plan by orders of magnitude and congest the depot’s own cells at fleet scale. Pure cloud recording belongs to wired, unmetered settings like home and shop cameras. A vehicle’s radio is metered and mobile. Recording-tier output and watch-tier upload differ by an order of magnitude, the gap the placement question stands on.
Rarely. The card shares the vehicle’s fate, so a fire, theft or destructive crash takes the recording with the vehicle at exactly the moment evidence matters. Remote review also becomes slow: one reader, one vehicle, by playback query at link pace or by waiting for the card to come home. The standards’ hybrid sends alarm parcels up at the event for survival and shared access, at megabytes per event. Local-only fits the narrow case of home-based vehicles with no remote review and an accepted loss risk, a case that shrinks as insurers ask for evidence within days.
Two meters run. The uplink crosses the data plan at the per-gigabyte rates the billing pages price. The storage itself, on published surveillance-cloud pricing, runs from a few dollars per camera per month at light settings to tens of dollars at full resolution and month-long retention. The hybrid keeps both meters small by sending only alarm parcels and named pulls, megabytes per event against the card’s daily tens of gigabytes. A worked four-channel month sends up roughly one ten-thousandth of what it records, with the card carrying the rest at its near-zero amortised rate.
In both places, by design. The original sits pinned on the card until uploaded, protected from overwrite, and stays in the recording circle after. The server copy exists minutes after the event, survives the vehicle, serves the safety office and the insurer at once and carries the contract’s retention, commonly months to years at parcel sizes. The double placement is the survival line’s whole answer: a cab fire or a theft takes the card, by which time the parcels are already off the vehicle.
The card keeps days: its circle is sized by capacity and recording bitrate against the route’s worst offline-plus-unvisited run, then overwrites the oldest routine footage on its own schedule. Cloud retention is whatever policy names: alarm parcels for the months or years the safety contract states, pulled footage for the life of its case file, nothing at all for live views. The two windows are set independently, in different documents: the card line in the hardware specification, the cloud line in the platform contract.
No. The split is the transport standards’ own architecture: continuous local recording with retention, alarm parcels that upload at the event, a stored-footage query and playback session for named pulls, resumable transfers for bad cells. A compliant terminal and platform carry all of it, tested at docking. The fleet’s work is configuration: alarm classes, retention lines, pull procedure and plan size, four lines in four documents, written once and reviewed against the monthly figures.
