Private 5G vs Wi-Fi 7 vs TSN: Factory Floor Wireless Decision (2026)

Every plant that plans a wireless refresh eventually ends up in the same meeting. Someone from IT says Wi-Fi 7 is fast and already on the laptops. Someone from the telecom vendor says private 5G is the only technology with a scheduler worth trusting. Someone from controls says none of it goes anywhere near a servo loop. All three are partly right, and the argument is usually conducted with peak data rates, which are the least useful numbers on the datasheet.

The real question in private 5G vs Wi-Fi 7 is not which is faster. It is which medium-access model gives you bounded, auditable behaviour for the traffic you actually have, at a cost your operations team can sustain, and where Time-Sensitive Networking (TSN) belongs in the picture. TSN is not a wireless rival at all; it is the deterministic wired layer that 5G can now be bridged into and that Wi-Fi 7 only approximates.

This guide works through the mechanisms (scheduled versus contended access), the 3GPP 5G-TSN architecture, spectrum models, mobility for automated guided vehicles (AGVs), a cost structure you can fill in with your own quotes, and a decision matrix by use case.

What this covers: the access-model difference, the 5G-TSN translator architecture, Wi-Fi 7 features and their limits, spectrum and licensing, roaming, cost structure, failure modes, and a use-case decision matrix.

Context and Background

Industrial wireless has three generations of incumbents. WirelessHART and ISA100.11a serve slow process sensors. Wi-Fi (802.11) took over handhelds, scanners, and then AGVs and AMRs because it was cheap and everywhere. Cellular stayed outside the fence, owned by carriers, until two things changed: regulators carved out local spectrum for enterprises, and 3GPP spent Releases 15 to 18 making 5G behave like an industrial network instead of a phone network.

On the regulatory side, Germany’s Bundesnetzagentur reserved 3,700-3,800 MHz for local use, assigned on application in 10 MHz blocks to parties with rights to a specific property, and not usable for public networks (Telecompaper). In the United States the Citizens Broadband Radio Service (CBRS) opened 3550-3700 MHz under a three-tier sharing model managed by a Spectrum Access System (overview). Those two moves are why “private 5G” is a purchasable product rather than a carrier slide.

On the standards side, 3GPP specified 5G support for TSN in Release 16, with further enhancements in Release 17, and the 5G Alliance for Connected Industries and Automation (5G-ACIA) concluded that the two releases together cover what industrial TSN integration needs (5G-ACIA white paper). Release 17 also added reduced-capability (RedCap) devices and enhanced non-public network support (3GPP Release 17); Release 18 is the first labelled 5G-Advanced. Our deep dive on 5G RedCap for industrial IoT covers the device side.

On the Wi-Fi side, IEEE 802.11be-2024 was published on 22 July 2025, after the Wi-Fi Alliance had already started certifying Wi-Fi 7 on 8 January 2024 (IEEE 802.11be summary). Wi-Fi 7 doubles maximum channel width to 320 MHz in the 6 GHz band, adds 4096-QAM (4K QAM, roughly 20 percent more bits per symbol than Wi-Fi 6’s 1024-QAM), and introduces Multi-Link Operation (MLO) (Wi-Fi Alliance). For background on earlier generations, see our Wi-Fi protocols comparison.

The third contender, TSN, is a family of IEEE 802.1 standards (time synchronization in 802.1AS, scheduled traffic in 802.1Qbv, frame replication in 802.1CB, preemption in 802.1Qbu/802.3br) that give Ethernet bounded latency. If you want the wired comparison against PROFINET and EtherCAT, read PROFINET vs EtherCAT vs OPC UA FX. This post asks the adjacent question: how far can TSN-style determinism be extended over the air, and by which radio.

A word on method. Latency and reliability figures for radio systems are conditional: on the traffic mix, the cell load, the antenna design, the TDD configuration and the radio environment. Where I quote a number from a primary source I cite it; where a number is typical or illustrative I label it as such. Treat vendor “up to” figures as marketing until you have measured on your own floor.

Scheduled Access vs Contended Access: The Core of the Decision

The decisive difference is the medium-access model. Private 5G schedules every uplink and downlink transmission centrally from the base station, so delay is bounded by the scheduler and the radio conditions. Wi-Fi 7 still lets stations contend for the channel, softened by OFDMA scheduling and restricted target wake time, so worst-case delay is statistical. TSN is a wired guarantee; 5G-TSN extends it, Wi-Fi 7 approximates it.

Figure 1: Decision flow. Hard real-time loops stay wired on TSN or EtherCAT; mobility and spectrum control push toward private 5G; everything else is a Wi-Fi 7 design problem.

Figure 1 compresses the argument of the whole article. The first branch is the one most vendor decks skip: if the control loop has a cycle time in the sub-millisecond to low-millisecond range, no wireless system in this comparison is the right answer, and the section on trade-offs explains why with numbers. Wireless earns its place for the mobile, the awkward-to-cable and the non-hard-real-time.

Why a scheduler changes the failure model

In 5G New Radio (NR), the gNB (base station) owns the resource grid. A device that needs to send data requests an uplink grant, or in configured-grant operation is given a periodic grant in advance. Because the network decides who transmits in which slot, two devices never collide inside a cell, and the cell can reserve capacity for a flow with a known period. This is the property that makes periodic industrial traffic (a PLC talking to a drive every few milliseconds) a natural fit.

The grid is also finite, and that is where the trade-offs live. At 30 kHz subcarrier spacing, typical for mid-band TDD such as band n78 (3.3-3.8 GHz), a slot lasts 0.5 ms. In a TDD pattern with many downlink slots and few uplink slots, an uplink packet that arrives just after an uplink opportunity waits for the next one. A pattern repeating every 2.5 ms (five slots) with one uplink slot can add up to roughly 2 ms of alignment delay before any retransmission. Those are arithmetic consequences of the frame structure, not vendor claims, and they are why TDD pattern selection is a first-class design decision for private networks.

Wi-Fi works the opposite way. A station senses the channel, and if it is busy it draws a random backoff and waits. Wi-Fi 6 introduced OFDMA so the access point can schedule groups of stations in a trigger-based exchange, and Wi-Fi 7 builds on that. But the basic enhanced distributed channel access (EDCA) mechanism remains the default, and any other device on the channel, including a neighbour’s access point or a rogue hotspot in unlicensed spectrum, can seize the medium for a frame duration.

What “URLLC” actually promises

Ultra-reliable low-latency communication (URLLC) is a 5G service category, not a single feature. The ITU’s IMT-2020 targets for it are commonly quoted as 1 ms user-plane latency and 99.999 percent reliability for a small packet (my recollection of ITU-R M.2410; verify against the recommendation before quoting it in a tender). Ericsson’s TSN integration paper gives the more modest figure of up to 99.999 percent reliability over the radio interface in a controlled environment such as a factory (Ericsson Technology Review).

Two caveats matter. First, 1 ms and 99.999 percent are target conditions for one packet in an unloaded, well-engineered cell, not a promise for every flow in a loaded one. Second, the mechanisms that deliver them (short slots, mini-slots, robust modulation and coding, repetition, packet duplication across carriers or paths) trade capacity for reliability. A cell tuned for 99.999 percent on a handful of control flows will carry far less aggregate data than the same cell tuned for throughput.

Three layers, not three rivals

A mature plant design usually uses all three technologies at different layers. TSN or a fieldbus runs the hard real-time islands (motion control, safety). Private 5G carries mobile and semi-fixed workloads that need predictable behaviour across a large site: AGVs, handheld tools, wireless I/O blocks, camera uplinks. Wi-Fi 7 carries high-throughput, delay-tolerant or short-range traffic: engineering laptops, tablets, AR headsets in a defined zone, bulk file sync. The skill is placing each flow on the layer whose failure mode it can tolerate.

Deeper Analysis: How 5G-TSN Bridging Works

5G-TSN integration makes the entire 5G system look like one or more IEEE 802.1Q bridges to the wired network. Translators at the device and at the user-plane function hide the radio, carry gPTP time across, and map TSN traffic classes to 5G QoS flows. The control plane learns the virtual bridge through the Application Function and hands scheduling information to the 5G core.

Figure 2: The 5G system as a virtual TSN bridge. Translators at the device side and the user plane function adapt Ethernet frames and time; a control-plane function exchanges configuration with the centralized network configuration entity.

The virtual bridge and its translators

According to Ericsson’s architecture paper, the wired network sees the 5G system as a set of TSN bridges, one virtual bridge per user plane function (UPF), and the system can interconnect bridges as well as attach end stations (Ericsson). Translators (TT) sit at the UE and the UPF in the user plane and in the Application Function (AF) in the control plane. They provide bridge ingress and egress port behaviour, including hold-and-forward buffering so that variable radio delay is smoothed into a predictable residence time. The 3GPP specification names the user-plane translators device-side (DS-TT) and network-side (NW-TT); the Ericsson paper itself does not use those labels, so I use them here as the 3GPP terminology (check TS 23.501 for the normative text).

The consequence for design is that the 5G system is a bridge with a published, bounded residence time, not a transparent cable. A TSN schedule computed by a Centralized Network Configuration (CNC) entity must account for the minimum and maximum delay the virtual bridge reports. If your radio cannot honour the bound on a given link, the right outcome is an admission failure at configuration time, not a late frame at runtime.

Time synchronization across the air interface

TSN depends on a shared notion of time. The 5G system appears as an IEEE 802.1AS (gPTP) time-aware system, forwarding gPTP information through the user-plane translators and accounting for its own residence time in the correction field. The 5G system clock can act as grandmaster for the whole deployment, and radio nodes are synchronized internally using the ITU-T G.8275.1 PTP telecom profile. The 5G-ACIA paper states that 5G supports up to 128 separate gPTP time domains simultaneously, which matters because IEC/IEEE 60802 expects a minimum of three working clock domains for Class A devices and two for Class B (per the Ericsson paper).

Be precise about what this buys. It delivers a common time base to devices on both sides of the bridge, which is what a time-aware shaper needs. It does not make the radio itself cycle-synchronous with your application; the application cycle and the radio slot are still two clocks that drift against each other unless you engineer their phase relationship.

QoS mapping and scheduled traffic

TSN traffic classes are mapped to 5G QoS Identifiers (5QIs) via the AF and the Policy Control Function (PCF); each 5QI is then handled per its own QoS profile, typically as a QoS flow inside a PDU session. For deterministic periodic streams, 3GPP defines Time Sensitive Communication Assistance Information (TSCAI), which lets the network know the burst period and arrival time so the gNB can place configured grants ahead of the data. For Class A virtual bridges the 5G system emulates time-controlled transmission aligned with 802.1Qbv scheduled traffic, using transmission-time information the AF receives from the CNC (per Ericsson).

The practical gain of TSCAI is that the scheduler stops reacting to scheduling requests and starts anticipating the traffic. Reactive uplink costs a scheduling request cycle plus a grant; anticipatory uplink removes most of that. The practical risk is the reverse: if the application’s real period drifts from the declared period, grants either go unused (wasted capacity) or miss the traffic (added delay).

Use cases the standards bodies actually examined

5G-ACIA examined three: controller-to-controller (C2C), controller-to-device (C2D) and device-to-compute (D2Cmp). C2C traffic is cyclic, small and tightly bounded, and mostly stays wired. C2D is where mobile equipment and wireless I/O live. D2Cmp is about offloading heavier compute, such as vision or control-in-the-cloud pilots, to edge servers. The paper reports that Releases 16 and 17 together provide what industrial TSN integration needs, but also that the IEC/IEEE 60802 industrial TSN profile was still evolving, which means 5G must track a moving target (5G-ACIA).

Where OPC UA FX fits

OPC UA FX (Field eXchange) is the controller-to-controller and controller-to-device extension of OPC UA intended to run over TSN Ethernet. Over a 5G-TSN bridge the payload is simply Ethernet, so FX traffic rides the QoS flow mapped to its TSN class. I have not seen a published, independently verified multi-vendor performance report for OPC UA FX across a 5G bridge; treat any vendor claim in that space as demonstration-grade until you have a reproducible test. The architectural point is that the OPC UA information model gives you a vendor-neutral semantic layer across both wired and wireless segments, which is what a digital twin ingest pipeline wants.

What Wi-Fi 7 adds, and where determinism still ends

Wi-Fi 7 is a real step for industrial use, but its improvements are about throughput and latency tails, not about a guarantee. The features that matter on a factory floor are these.

320 MHz channels and 4096-QAM. Wi-Fi 7 doubles maximum channel width to 320 MHz, available in the 6 GHz band, and 4K QAM carries 12 bits per symbol instead of 10. IEEE 802.11be’s theoretical single-band maximum is 23 Gbit/s, and real-world rates are much lower; the final specification dropped the early 16-spatial-stream goal in favour of 8 (summary). Wide channels are a capacity feature. On a floor with dozens of APs, 320 MHz channels are mostly unattainable because the 6 GHz band only holds a few of them; the realistic planning width is 80 or 160 MHz.

Multi-Link Operation. An MLO device can use 2.4, 5 and 6 GHz links at once. The value for reliability is that a frame can be sent on whichever link is clear, or duplicated, which cuts the tail of the latency distribution when one band is congested. MLO is mandatory in the specification per the 802.11be summary above. Which MLO modes a particular client chip supports (simultaneous transmit and receive versus enhanced single-radio variants) varies, and low-cost IoT clients often implement the cheaper modes, so test the client you will actually buy.

Preamble puncturing. A device can blank the part of a wide channel hit by interference and keep using the rest. It is mandatory for Wi-Fi 7 certification (same summary). Useful when a narrowband interferer sits in your channel.

Restricted TWT. Target wake time (TWT) was introduced in Wi-Fi 6 mainly for power saving. Restricted TWT (R-TWT) in 802.11be is aimed at latency-sensitive traffic such as industrial IoT and virtual reality: the AP establishes protected service periods during which only designated low-latency streams may transmit, shielding them from contending best-effort clients (the Rohde and Schwarz Wi-Fi 7 white paper describes it that way). The Wi-Fi Alliance Wi-Fi CERTIFIED 7 page, as I read it, does not list R-TWT among certified features, so do not assume your “Wi-Fi 7 certified” AP and client implement it; ask for the specific capability in writing.

Figure 3: The contention problem. A station that finds the channel busy draws a random backoff, so access delay is a distribution, not a bound.

Figure 3 shows why, even with R-TWT, determinism stops at the edge of your own BSS. R-TWT protects a service period against stations that obey the AP’s schedule. It cannot silence a neighbouring operator’s overlapping BSS on the same channel, a visitor’s hotspot, or a non-Wi-Fi emitter. Where the 6 GHz band is available and you control the building, interference exposure drops sharply compared with 2.4 and 5 GHz, which is the main reason Wi-Fi 7 on 6 GHz can reasonably host semi-critical traffic. But that is an environmental advantage, not a protocol guarantee.

6 GHz regulation is not uniform

The 6 GHz band that gives Wi-Fi 7 its clean 320 MHz channels is regulated differently by region: the United States opened roughly 1.2 GHz of it (5.925-7.125 GHz), while Europe made only the lower portion (about 5.945-6.425 GHz) available for unlicensed use at the time I last checked (verify current CEPT and national decisions before designing a multi-country rollout, since this is under review). The channel plan, maximum power and indoor-only restrictions differ. A fleet of AGVs that crosses an outdoor yard may be forbidden from using 6 GHz on that stretch, so the 6 GHz advantage frequently does not extend to the whole route.

Wi-Fi 6E/7 determinism limits, in one list

Medium access is contention-based by default; scheduled elements (OFDMA triggers, R-TWT) cover only the clients that participate.
Roaming is client-driven: the station decides when to leave an AP, and fast-transition mechanisms (802.11r/k/v) help but do not make handover make-before-break for every client.
Link-layer retries hide loss at the cost of delay; the retry limit and rate adaptation algorithms are vendor-specific.
Unlicensed spectrum offers no regulatory protection from other users.
There is no standardized, end-to-end equivalent of the 5G-TSN bridge for 802.11be; IEEE 802.11 working groups have discussed TSN-aligned scheduling extensions, but I am not aware of a mature, widely shipped industrial profile. Label this as my assessment, not a verified market survey.

Spectrum Models: Who Gets to Transmit, and What Protection You Buy

Spectrum is where private 5G and Wi-Fi 7 diverge most in risk terms. With Wi-Fi you share the commons. With private 5G you either hold a local license, rent a slice from an operator, or participate in a managed sharing scheme. The model you can obtain depends on your country, so verify with the national regulator before architecture work starts.

Germany: local licences in 3.7-3.8 GHz

The Bundesnetzagentur assigns 3,700-3,800 MHz on application in 10 MHz blocks to applicants with ownership or rights to a specific property; the spectrum may not be used for public networks, and licensees must coordinate with mobile operators to avoid interference (Telecompaper). The practical effect is an exclusive, interference-protected channel on your campus, which is the closest a plant gets to the “wired-like” property in a radio. The same source does not cover the 26 GHz local band or the fee schedule, so I do not state them here; check the regulator for current fees and for how the band has been extended since 2019.

United States: CBRS shared spectrum

CBRS is 150 MHz wide, 3550-3700 MHz, in three tiers: Incumbent Access (federal users such as Navy radar, protected from harmful interference), Priority Access (licensed users in 3550-3650 MHz) and General Authorized Access (GAA), licence-by-rule use of whatever is not assigned higher, plus opportunistic use of idle Priority channels. A Spectrum Access System (SAS) manages it dynamically, and users pay the SAS provider a fee even for GAA-only operation (CBRS overview). A Priority Access License (PAL) is a non-renewable authorization for a 10 MHz channel in one census tract for three years, with up to seven PALs per tract and no more than four to any one applicant in the first auction. 3GPP created Band 48 for TD-LTE across the full CBRS range.

For a plant, the important operational point is that GAA is shared. Your SAS can move your radios off a channel if a higher tier appears or if the SAS’s coexistence algorithm reassigns spectrum, and a neighbouring GAA user can use the same channel. In rural or isolated sites that is a theoretical risk. In a dense industrial park with several CBRS deployments, it is a design input. PAL is the fix, at a price and with a three-year term that matters for capital planning. I did not verify CBRS rule changes proposed or adopted in 2024-2026, so check the FCC for the current state before committing.

Other countries

Several other regulators have local-licence or shared-access schemes for industrial private networks, with differing bands and fees. I have not verified the current status country by country for this post, so use the following as a checklist, not as facts: does your regulator offer (1) a local licence in the 3.3-4.2 GHz range, (2) a millimetre-wave local band, (3) a shared-access model like CBRS, or (4) only operator-provided network slicing or “private network as a service”? The answer determines whether you can own the radio spectrum, and therefore whether you can make a hard commitment on interference.

The operator-slice alternative

If no local spectrum is available, a mobile operator can offer a dedicated or sliced network on its public spectrum with on-premises UPF and radio. This keeps spectrum protection but gives up some control: upgrade windows, radio parameters and even the TDD pattern are the operator’s, not yours. Contract terms on SLA, outage credits and parameter-change rights are what you are buying, so negotiate them like a safety-relevant utility.

Why spectrum protection matters more than peak rate

An illustrative comparison: suppose a plant has 40 AGVs sending a 200-byte status message every 100 ms and a 100 kbit/s telemetry stream each. Total payload is a few hundred kbit/s, trivially small for either technology. What breaks such a deployment is not capacity; it is the tail event, say the 1-in-10,000 handover that takes 400 ms on the radio and trips a safety stop. Spectrum control (and the radio network’s behaviour in that tail) is the product you are buying, not megabits.

Figure 4: A private 5G reference architecture. Radio units connect to an on-premises user plane function that keeps OT traffic local; the control-plane core may be on site or hosted, and SIM provisioning is a first-class subsystem.

Figure 4 shows the shape of a typical on-premises deployment. The local UPF is the piece that makes private 5G different from a plain cellular subscription: user-plane traffic is broken out inside the fence, so latency is not hostage to a backhaul and OT data never leaves the site. The control plane (AMF, SMF, UDM) can be a compact on-premises core or a hosted one; hosting reduces operational burden but ties session setup and mobility signalling to the WAN. Whichever you choose, define what happens when the WAN link fails: an on-site core can keep existing sessions and admit known SIMs, a hosted core may not.

Mobility, Handover and AGV Behaviour

AGVs and autonomous mobile robots (AMRs) are the workload that most often justifies private 5G, and also the one where marketing claims outrun measurements. The relevant property is not average latency but what happens at the moment a vehicle changes cells.

Handover in 5G

In 5G NR the network commands the handover. The serving gNB configures measurement reports, the UE reports neighbour signal levels, and the network decides and prepares the target cell before the UE moves. A conventional NR handover is still break-before-make at the radio level: the UE detaches from the source and attaches to the target, typically producing an interruption on the order of tens of milliseconds (a typical range, which varies by implementation; measure it). Release 16 specified enhancements aimed at reducing interruption time, such as dual active protocol stack handover, where the UE keeps the source link while synchronizing to the target, with the goal of close-to-zero interruption. Vendor support for those options varies, and I did not verify which commercial private-5G radios implement them, so ask for it in the RFP and test it.

What the scheduler gives you across a cell edge is predictability: the network knows which UEs are near the edge and can handle them deliberately, including packet duplication across two cells in some configurations. What it does not give you is a free pass for coverage holes. An AGV that drives behind a steel rack into a shadow will lose uplink regardless of the standard.

Roaming in Wi-Fi

In Wi-Fi the client decides to roam. A well-behaved fleet client scans, authenticates and reassociates to a new AP. With 802.11r fast BSS transition, the key exchange is pre-computed, shortening the authentication step, and 802.11k and 802.11v help the client discover and be steered to better APs. Even so, the gap can range from a few milliseconds to several hundred milliseconds depending on client firmware, scan strategy and whether the environment is a single-channel or multi-channel deployment (typical range; the variance is the point). MLO helps in principle, because a client holding links in two bands can keep one alive while another reassociates, but only on clients and APs implementing the necessary multi-link modes.

The pragmatic AGV approach is application-layer tolerance. VDA 5050, the interface standard between AGV fleets and master control, expects vehicles to continue executing already-sent orders when connectivity drops and to report on reconnect, so brief gaps do not stall the fleet. Our VDA 5050 fleet management architecture post covers that design. If your AGV software is VDA 5050-style, a 300 ms radio outage is usually a non-event; if the vehicle streams safety-rated remote control over the radio, it is a different class of problem and probably should not be on any shared wireless link.

Worked example: where the radio design bites

Consider a 60,000 square metre plant with 80 AGVs. Illustrative assumptions: each AGV sends 1 kB of pose and status every 50 ms (160 kbit/s, 12.8 Mbit/s fleet-wide uplink), plus a 2 Mbit/s camera stream on 20 vehicles (40 Mbit/s). Both radios handle that load in an unloaded cell. The design questions are then:

Coverage geometry. Mid-band 3.5 GHz propagates better than 6 GHz Wi-Fi through obstacles; a rough rule is that the 6 GHz Wi-Fi design needs noticeably more access points for the same coverage (the actual ratio depends on path-loss exponent and clutter, so run a survey).
Cell-edge density. Every cell boundary is an event for every vehicle that crosses it. Fewer, larger 5G cells mean fewer boundaries but more devices per cell.
Uplink share. Typical TDD patterns are downlink-heavy. An uplink-heavy camera fleet may need an uplink-weighted pattern, which reduces downlink for the rest of the plant. This is a trade-off Wi-Fi does not have, because Wi-Fi time-shares the channel dynamically.

The result is that 5G wins on controlled behaviour at boundaries and on spectrum protection; Wi-Fi wins on uplink flexibility and ecosystem breadth. Neither wins everywhere.

Cost Model Structure

I cannot give you prices; they vary by vendor, region and volume, and published numbers go stale in a quarter. What I can give you is the structure of the cost model so you can fill it with real quotes. Every figure below is a variable, not data.

Private 5G CAPEX. Radio units and baseband, core (software licence or appliance, or hosted subscription), transport and power over Ethernet or fibre fronthaul, GPS or other time sources for synchronization, edge compute for the local UPF, site survey and RF planning, and integration with identity and OT security. Recurring costs: spectrum fees or SAS fees, support contracts, SIM lifecycle management, staff or managed-service fees, upgrades.

Wi-Fi 7 CAPEX. Access points (more of them for equal coverage at 6 GHz), controllers or cloud management licences, switching with sufficient PoE budget (Wi-Fi 7 APs draw more power than older generations; check the specific model’s PoE class), survey and cabling. Recurring: licences, support, periodic re-survey, replacement of client radios.

Device cost. This is often the dominant term at scale. Wi-Fi modules are cheap and abundant. Private-5G modules cost more and have fewer SKUs, although RedCap devices in Release 17 are specified to be lower complexity than full-capability NR, which should help over time (3GPP Release 17). For fleets of thousands of low-rate sensors, the per-device premium dominates; for dozens of AGVs it vanishes.

Hidden costs. Wi-Fi: interference investigations and repeated tuning. Private 5G: the skills gap (RAN planning, core operations, SIM and eSIM provisioning), and integration with enterprise identity. A realistic total-cost comparison assigns both a loaded staff cost, a refresh cycle (radio generations change faster than PLCs), and a downtime cost for the worst plausible outage. The last term often decides the question for plants where a line stop costs more per hour than the radio costs per year.

Decision Matrix by Use Case

The matrix below reflects my engineering judgement from the mechanisms above, not a vendor benchmark. “Fit” means the technology can be justified on its merits for the workload under typical conditions; always validate on your site.

Use case	Wired TSN or fieldbus	Private 5G (incl. 5G-TSN)	Wi-Fi 7 (6 GHz)	Comment
Servo and motion control, sub-ms cycles	Primary	Not recommended	Not recommended	Radio retry and slot alignment break tight cycles
PLC to PLC, cyclic few-ms traffic	Primary	Possible with 5G-TSN, pilot only	Weak	Use wired unless cabling is impossible
AGV and AMR fleet, VDA 5050 style	n/a	Strong	Good with careful roaming design	Cell-edge behaviour decides it
Wireless I/O and remote I/O blocks	Wired if feasible	Good	Fair	Needs bounded latency and clear spectrum
Condition monitoring, thousands of sensors	Impractical	Good with RedCap or LTE-M style devices	Good if devices are cheap	Device cost drives the choice
AR and tablet work instructions	n/a	Good	Strong	Throughput and ecosystem favour Wi-Fi
Machine vision uplink	Wired	Good with uplink-weighted TDD	Strong in a defined zone	Check uplink capacity
Mobile handhelds and scanners	n/a	Good	Strong, already deployed	Legacy fit favours Wi-Fi
Safety functions (e-stop, safe zones)	Primary, certified safe protocol	Only with a certified safety layer	Only with a certified safety layer	Do not rely on radio performance claims

Two lines in the matrix deserve emphasis. Safety signals over wireless are possible in principle through a safety protocol such as PROFIsafe or CIP Safety that tolerates message loss and fails to a safe state, but the radio’s job then is availability, not safety. And condition monitoring is the use case where neither premium technology is obviously required: LoRaWAN, Wi-Fi HaLow or WirelessHART can be cheaper, though they fall outside this comparison.

When 5G-TSN is worth the integration effort

Choose 5G-TSN when three conditions hold simultaneously: the application already speaks TSN or OPC UA FX on a wired backbone, a portion of the endpoints must be wireless and mobile, and a configured TSN schedule must reach those endpoints end to end. If only the third condition is missing, a plain private 5G network with a prioritized QoS flow is much simpler. In practice, in 2026, a plain QoS-mapped private network is by far the more common deployment. Whether 5G-TSN functions (translators, CNC interworking, gPTP across the air) are present in the commercial core and device products you can buy is a question for each vendor; the 3GPP specification support does not imply product availability, and I have not independently verified shipping product capabilities.

Interplay with digital twins

A digital twin of a plant is only as truthful as its inbound data. The relevant property is timestamp quality: if the AGV position arrives with a 300 ms jitter and the twin renders it as real time, the twin’s collision predictions are wrong. 5G-TSN’s gPTP bridging and Wi-Fi’s timing measurement features both exist to help; neither is automatic. Stamp at the source with a disciplined clock, carry the stamp in the payload (OPC UA SourceTimestamp, MQTT message properties), and treat network arrival time as a separate, secondary signal.

Trade-offs, Gotchas, and What Goes Wrong

Radio is not a cable. Even a perfectly engineered private 5G cell has fading, interference from the plant’s own machinery (arc welding and variable-frequency drives emit broadband noise), and moving metal. Hard real-time loops should be wired. Do the arithmetic: a 1 ms cycle with a required update every cycle and a 99.999 percent per-cycle reliability implies an expected loss event about every 100,000 cycles, which is roughly every 100 seconds at that cycle time. Whether a control loop can ride through one lost update per 100 s is an application question, but most servo loops cannot, which is why the answer for motion is still copper or fibre.

Packet duplication has a price. Duplicating packets across two paths raises reliability by assuming independent failures. In a factory, failures are correlated: the same steel rack shadows both antennas. Independent paths require independent antennas, independent power and independent spectrum.

Wi-Fi 7 client reality. The headline features depend on both ends supporting them. Many industrial clients are two to three years behind the access point generation. If your AGV client radio is Wi-Fi 5, an access point upgrade changes little for those devices and merely adds more contention-free service for the laptops.

The skills trap. Private 5G pushes a plant into running a mobile core, SIM inventory and RAN tuning. That is a different discipline than OT network engineering. Plants that treat it as just another switch end up with undocumented configuration drift. Either budget for a managed service with clear boundaries or build a team.

Lifecycle mismatch. A PLC lives 15 to 20 years; a radio generation lives five to seven. Design for radio replacement without touching controllers, by keeping the application protocol (OPC UA, MQTT) independent of the radio.

Security surface. Private 5G brings SIM-based mutual authentication and a mature standard security framework. Wi-Fi 7 brings WPA3 and 802.1X with certificates. Both are good if configured properly. The failure mode is the weak default: shared pre-shared keys on Wi-Fi, flat networks behind the UPF on 5G. Segment per IEC 62443 zones and conduits regardless of the radio.

Spectrum policy risk. Sharing rules can change. GAA channels can be reassigned, 6 GHz rules vary by region, and a local licence is tied to a property and a term. Put regulatory change in the risk register.

Measurement discipline. Vendors quote best-case latency. Ask for a latency histogram (p50, p99, p99.99) at stated load, packet size, TDD pattern and number of devices, measured end to end from application to application. If a vendor cannot or will not provide it, assume the tail is worse than the headline.

Practical Recommendations

Start from traffic, not technology. Inventory every wireless flow in the plant with its packet size, period, tolerance for a lost or late message, mobility, and who owns the endpoint. Sort the flows into four buckets: hard real time (stay wired), mobile with bounded latency (candidate for private 5G), high-throughput local (candidate for Wi-Fi 7), and massive low-rate (candidate for RedCap or a low-power network). Most plants discover that the bucket needing private 5G is smaller than the vendor pitch implies, and also that it is the one with the highest cost of failure.

If you have a license option, run a time-boxed pilot on one real flow, typically an AGV zone, with a defined acceptance test: end-to-end latency histogram at p99.99, handover interruption measured on the actual vehicles, and a deliberate interference test during real production. Run the same test on a well-surveyed Wi-Fi 7 deployment with the same clients. Decide on data from your floor.

Keep the application layer radio-agnostic. Use OPC UA and MQTT with source timestamps, make AGV software tolerate connection gaps as VDA 5050 does, and avoid embedding radio-specific behaviour in controllers. That preserves your freedom to replace the radio generation.

Classify flows: hard real time, mobile bounded, high throughput, massive low-rate.
Verify spectrum options with your national regulator before architecture; record term and renewal risk.
Demand p50, p99 and p99.99 latency at stated load and TDD pattern, measured application to application.
Confirm in writing that your client devices support the Wi-Fi 7 features you rely on (MLO modes, R-TWT) or the 5G features you rely on (RedCap, TSCAI, dual-connectivity handover).
Test handover and roaming with the real vehicles at working speed.
Segment with IEC 62443 zones and conduits; keep SIM and certificate lifecycle owned by a named team.
Define WAN-failure behaviour for the core and for the AGV fleet.
Keep motion and safety loops on wired TSN or a fieldbus.

Frequently Asked Questions

Is private 5G better than Wi-Fi 7 for a factory?

Neither is universally better. Private 5G schedules transmissions centrally, offers managed mobility and can use protected local spectrum, which suits AGVs and bounded-latency mobile flows. Wi-Fi 7 offers cheaper hardware, a vast device ecosystem and high throughput, which suits tablets, AR and defined zones. Wi-Fi’s contention-based access gives statistical rather than bounded latency. Choose by workload, spectrum availability and the skills you can sustain, and verify with a pilot on your floor.

Can private 5G replace TSN or industrial Ethernet?

No. TSN and industrial Ethernet remain the right choice for sub-millisecond cycles, motion control and safety. 3GPP Release 16 and 17 let a 5G system act as a virtual TSN bridge so that wireless endpoints can join a TSN network, with translators carrying gPTP time and mapping traffic classes to 5G QoS flows. That extends TSN to mobile endpoints; it does not make the radio equal to copper for the tightest loops.

What is 5G-TSN integration?

5G-TSN integration lets the 5G system appear to a TSN network as one or more IEEE 802.1Q bridges. Translators at the device and at the user plane function hide the radio, hold and forward frames to smooth delay, and pass 802.1AS gPTP time. A control-plane function exchanges configuration with the centralized network configuration entity, and TSN traffic classes map to 5G QoS flows. The 5G system can support up to 128 gPTP time domains according to 5G-ACIA.

Does Wi-Fi 7 support deterministic latency?

Partly. Multi-Link Operation reduces latency tails by letting a device use multiple bands, and restricted target wake time creates protected service periods for latency-sensitive streams. But access is still contention-based for stations outside the schedule, so interference or overlapping networks can break bounds. Wi-Fi 7 on 6 GHz in a controlled building can host semi-critical traffic well, but it does not provide the guarantee a scheduled system does. Verify R-TWT support on both access point and client.

What spectrum can I use for private 5G?

It depends on the country. Germany assigns local licences in 3,700-3,800 MHz in 10 MHz blocks to parties with rights to a property. The United States has CBRS in 3550-3700 MHz, a three-tier model coordinated by a Spectrum Access System, where licensed Priority Access covers 3550-3650 MHz and General Authorized Access is licence-by-rule. Other countries have their own models or only operator-provided slices, so check with your national regulator.

Is 5G or Wi-Fi 7 better for AGVs?

For large sites with many vehicles, private 5G’s network-controlled handover and protected spectrum are attractive, particularly where the fleet software tolerates brief gaps. Wi-Fi 7 can serve smaller or already-covered sites well if roaming is engineered carefully and clients support fast transition and multi-link modes. Handover interruption is the deciding metric: measure it on your own vehicles rather than trusting datasheet values, and design the fleet software to tolerate gaps.