Wi-Fi: most people's first hop

Everything this course has taught so far assumes a link exists. Encapsulation assumes a frame can be put on a wire. Subnetting assumes a local segment. For most people, in most places, that first link is not a wire at all: it is a few metres of radio between a laptop or a phone and an access point on a ceiling. It is the only hop the user can see the equipment for, the only one they can physically walk away from, and the one with the least evidence attached to it. That combination is why it collects the blame.

Start with what association actually is, because the vocabulary here does more work than any other part of Wi-Fi. A wireless network advertises a name, the SSID, which is the string a user picks from a list. A client does not connect to a name. It connects to one specific radio, identified by a BSSID, which is normally that radio's MAC address. A single office SSID may be served by three hundred BSSIDs across six floors. Users complain about the SSID. Causes live in a BSSID. Almost every unproductive Wi-Fi conversation you will sit through comes from the two sides of the table using the same word for different things: one means the network as a service, the other means one radio in one corridor.

Once associated, the client sends frames. These are link-layer units doing the same job as the Ethernet frames from the encapsulation module, carrying the IP packet inside them and being stripped away at the access point, which bridges the payload onto the wired network. The IP header the client wrote survives that transition untouched, exactly as it survives every other hop. Nothing above the link layer knows or cares that the first hop was radio.

Wi-Fi frames come in three families, and knowing that there are three explains a quantity of otherwise baffling behaviour. Data frames carry user traffic. Management frames set up and tear down associations, and include the beacons every access point broadcasts continuously so that clients can find it. Control frames coordinate access to the medium itself. The consequence is that a wireless network spends part of its capacity talking about itself, before a single byte of anyone's traffic moves. On a quiet network that overhead is invisible. On a crowded one it is a meaningful slice of what is available, and it grows with the number of SSIDs you advertise, which is why publishing eight networks from one radio is a design decision with a cost rather than a free convenience.

Now the property that makes Wi-Fi behave unlike the cable it replaced. A switch port gives a device its own lane and lets it send and receive at the same time. A radio cell gives every device a share of one lane, and only one radio in that cell can transmit at a time. Devices listen before they talk, and when the medium is busy they wait for a random interval before trying again. Radios cannot detect a collision while transmitting the way a wired interface can, so the whole design is built around avoiding collisions rather than recovering from them. Politeness, in other words, is structural.

The scarce resource that falls out of this is not bandwidth. It is airtime: the time each device occupies the channel while it transmits. Two devices sending the same quantity of data do not consume the same airtime. A device close to the access point with a clean signal negotiates a high data rate and finishes quickly. A device at the far edge of coverage negotiates a low rate and takes far longer to say exactly the same thing, and every other device in the cell waits through all of it. This single mechanic explains the most common complaint in the whole discipline, which is that one person's connection got worse because of somebody else's device, in a building where the wired network is entirely healthy.

One more asymmetry completes the picture, and it is the one that most surprises people who come from wired networking. The client decides. The client decides which BSSID to join, at what rate to transmit, when the current radio has become bad enough to leave and where to go next. The network can advertise, hint and refuse, but the decision logic sits in the device driver of a phone or a laptop that the organisation may not own. Wired networking trains people to believe the network is in charge of the network. At the access layer that is simply not true, and the fixes that assume it is are the fixes that fail.

Those four facts, one shared medium, airtime as the currency, an SSID that is not a radio, and a client that makes the choices, are the whole toolkit. The obvious question they raise is what can be done about a shared medium at all, and there are only two honest answers: make each transmission shorter, or find somewhere else to transmit. The next section is about the moment the industry finally got somewhere else.

For most of Wi-Fi's life, adding capacity meant making transmissions shorter. Every generation encoded more bits into the same slice of radio so that a device finished talking sooner and freed the channel for the next one. That works, but it has a hard ceiling: no amount of clever encoding creates a second conversation, and a cell where everyone talks quickly is still a cell where only one device talks at a time.

Opening the 6 GHz band changed the shape of that problem rather than its size. The two bands Wi-Fi had before behave very differently from each other. The 2.4 GHz band travels furthest and through the most walls, which sounds like an advantage until you notice that it is also the band shared with an enormous quantity of non-Wi-Fi equipment, and that it holds very few channels that do not overlap. Its reach is precisely what makes it crowded: your neighbours' networks reach you as easily as yours reaches them. The 5 GHz band carries more usable channels and less foreign traffic, at the cost of range, and it is where serious deployments have lived for years.

What 6 GHz adds is not just more room. It is room that older devices are not permitted to enter. That restriction, which sounds like a limitation, is the entire point. A band with no legacy clients has no legacy overhead: no slow devices negotiating slow rates and consuming airtime out of proportion to what they carry, no compatibility signalling for equipment that has not been sold in a decade. A network operating there starts from something close to a clean sheet, and the practical gain is more simultaneous conversations rather than a faster single conversation. Against the airtime argument from section 8.1, that is worth more than any encoding improvement that has ever shipped.

This is also where channel width has to be understood honestly, because it is the number vendors put on the box. Widening a channel raises the peak rate a device can reach, and it does so almost linearly, which is why the marketing loves it. It also does two other things. It reduces how many non-overlapping channels fit in the band, so wide channels mean fewer independent cells in the same building. And it makes the radio listen to a wider slice of spectrum, so it hears more interference and defers more often. In a home with one access point and a clear neighbourhood, wide channels are close to free. In a dense office with radios on every floor, configuring the widest channel available is one of the most reliable ways to make a network worse while every specification sheet says you improved it.

The design principle to carry forward is that capacity in a busy building comes from channel reuse, meaning many small cells on different channels serving fewer devices each, and not from the largest channel the equipment will accept. It is worth adding that how much of the 6 GHz band is available, and on what terms, is a regulatory matter that differs between countries, so a design that is optimal in one market is not automatically legal or sensible in another. Check the rules for the country you are deploying in before you copy a channel plan out of a vendor guide.

The band arrived first and the generation that knows what to do with it arrived second. Wi-Fi 7 is where 320 MHz channels become possible at all, and it is also where the gap between what is advertised and what a user experiences becomes widest, which makes it the right place to be careful.

Wi-Fi 7 brings three headline changes. Two of them are refinements of the old make-it-shorter strategy. The third is a structural change to what an association is, and it is the one that matters for how a network feels.

Take 4096-QAM first, because it is the purest example of a number that does not mean what a buyer thinks it means. Modulation is how many bits a radio packs into each transmitted symbol, and a denser scheme carries more bits in the same time. The trade-off is that the receiver must distinguish between more, and more similar, signal states, which demands an exceptionally clean signal. In practice that means a device close to the access point in good conditions gets the benefit, and a device in the next room does not, because it will have fallen back to a lower rate long before it gets near the top of the range. The headline rate is a laboratory condition. It is not a corridor.

320 MHz channels are the second, and section 8.2 has already supplied the judgement: double the width, double the peak rate, halve the number of independent cells you can build and widen the slice of spectrum in which any interference will stop you transmitting. These channels only fit in the 6 GHz band, so this headline feature is in practice a 6 GHz feature. Whether it helps depends entirely on how many radios you need to fit in the same building, which is a question about the building rather than about the standard.

Multi-link operation is the change of a different kind. Moving an association from one access point to another inside the same network has a name, roaming, and until this point it was always the same manoeuvre: a client held an association with one radio on one band, and going anywhere else meant tearing that association down and building a new one. Multi-link operation lets a single association hold links on more than one band at the same time. The client and the access point can spread traffic across those links for throughput, or hold a session on one while the other is unusable, and neither move requires re-associating.

That is why the board at the top of this page is a roaming map rather than a speed chart. On a single link, the handover pattern is break then make: the association with the fading radio has to end before the next one begins, and the gap between them is the stall people report. Applications experience that gap as a hung page or a video call that freezes for a second and recovers. Multi-link operation attacks the gap directly by making sure a working link exists throughout, which converts the most visible failure in wireless networking into something a user does not notice at all. No amount of extra modulation density does anything for that failure.

The adoption picture is where judgement is required, and where measurement literacy earns its place in a networking course. Wi-Fi 7 has been certified since January 2024. About 1.8 percent of global Wi-Fi connections were Wi-Fi 7 in the first quarter of 2026, with Singapore the leading national figure at 25.1 percent. Read those two numbers together rather than separately. A national figure more than ten times the global one does not mean the global figure is wrong; it means adoption is concentrated, in the places where devices are replaced fastest and networks are newest. It also means that any argument built on the global average will be wrong for some specific building, in both directions.

Note as well what the 1.8 percent is a share of. It counts connections, which is not the same as a share of devices sold, a share of access points installed, or a share of traffic carried. Each of those four questions has a different answer and each supports a different decision. Set that against the roughly 118 million access points forecast to ship in 2026 and the position becomes coherent rather than contradictory: Wi-Fi 7 is simultaneously a small fraction of what is connected today and the default thing being bought, because installed base and purchasing default are answers to different questions on different timescales.

For a working recommendation, that resolves cleanly. Design the network you are operating this year for the devices that exist this year, which means not assuming multi-link operation will rescue a coverage plan. Buy for the life of the installation, which means not paying to install a generation you will want to replace early. Neither of those requires believing the marketing, and both require knowing which number you are quoting.

Multi-link operation is the tell for where the standards work has gone. It buys no headline rate at all and it fixes the thing users actually complain about. The generation now being written makes that shift explicit rather than incidental.

The IEEE 802.11 working group is developing the next amendment, 802.11bn, under the name Ultra High Reliability. The name is the specification of intent. Certification is planned for late 2027 and ratification is expected in 2028, so nothing here is a purchasing decision today. It is an expectation-setting decision, and that is worth more than it sounds.

The stated targets are roughly 25 percent gains in three quantities: effective throughput, tail latency and roaming loss. Read that list slowly, because every item on it is deliberately not a peak number.

Effective throughput is what a device actually achieves in a real room with real neighbours and real interference, as opposed to the rate two devices can reach in ideal conditions. It is the difference between the figure on the box and the figure you would measure at a desk. Choosing to improve the second is a decision to be judged on the harder measurement.

Tail latency is the worst few percent of packet delays rather than the average. This is the single most useful idea in the section for a non-technical reader, because average latency is close to useless as a description of user experience. A video call is judged by its worst second, not its typical one. A voice call with an excellent average and a bad tail is a call that keeps breaking up. Averages hide exactly the events that generate complaints, which is why a network can look healthy on a dashboard while the people using it are unhappy, a theme the practice stage of this course returns to at length.

Roaming loss is the packets lost while a device moves from one access point to another, which is the same gap the multi-link diagram in section 8.3 illustrates. Naming it as a headline target of a whole generation is an admission by the industry that this, and not peak speed, has been the limiting factor on how good wireless networks feel.

Take those three together and the direction is unambiguous. A generation number used to be a promise about speed. It is now a promise about consistency: how much of the advertised capability survives contact with a real building, how bad the worst moments are, and whether moving around costs anything. If a supplier's pitch for the next generation leads with a bigger top-line rate, they are quoting the least interesting claim available about it, and a fair question in response is what happens to the tail.

Knowing that the standards are aimed at reliability is only useful if you can tell, on a given Tuesday, whether the complaint in front of you is a reliability problem at the access layer at all. That judgement comes down to three symptom shapes and the one piece of evidence that separates each of them from the other two.

Wi-Fi complaints arrive in a single undifferentiated form: the Wi-Fi is bad. Almost all of them resolve into one of three mechanisms, and each mechanism has a symptom shape and a piece of evidence that confirms it. Learning the three shapes is worth more than any tool, because the tool will tell you a hundred things and only the shape tells you which one to read.

The first is roaming stickiness. The symptom is location-dependent and movement dependent: fine at the desk, bad while walking between two places, and often cured by turning Wi-Fi off and on again, which forces a fresh choice of radio. Roaming is a client decision, so the usual cause is a device holding an association with a distant access point long after a nearer one became available, because its own logic has not yet judged the current link bad enough to leave. The evidence is which BSSID the client is associated with and at what signal level, set against what it could hear instead. The fix is usually some combination of device or driver behaviour and coverage design, and it is very often not more access points.

The second is airtime saturation. The symptom is area-wide and time-of-day-shaped: everyone in one part of a floor is slow at the same time, and a wired connection in the same room is perfectly healthy. Section 8.1 supplies the mechanism, including the fact that a single distant or slow device can consume airtime out of all proportion to the bytes it moves. The evidence is airtime or channel utilisation on the specific radio, the count of associated clients, and the data rates those clients are actually negotiating. A cell with modest client numbers and terrible rates is telling you about coverage; a cell with excellent rates and too many clients is telling you about density.

The third is interference. The symptom is intermittent and stubbornly location-specific rather than user-specific: it follows a place, not a person, and it often has no relationship to how many people are working. The evidence that separates it from saturation is the split of channel utilisation between your own traffic and everything else. A channel reported as busy while your own radios are barely transmitting is interference, and no amount of tuning your own equipment addresses somebody else's transmitter.

Before reaching for any of the three, ask the cheaper question: is this an access-layer problem at all? The paraphrase in section 8.1 is the reason a single test settles it. Because the frame is the only thing that changes when the first hop changes, repeating the same request from a wired port in the same room, at the same moment, splits the problem in half. If the wired attempt is fine, the fault is at or below the access layer and the three shapes above apply. If the wired attempt fails identically, the access layer is exonerated and the rest of this course applies instead: name resolution, routing, the transport, the server. Two further tests refine it almost as cheaply. A second device in the same seat separates a client-specific fault from an area-wide one. The same device on a different band or a different SSID separates a channel or band problem from a network-wide one.

A note on scope before this closes. Some specialist sites, large industrial estates and similar, deploy private cellular networks at the access layer instead of or alongside Wi-Fi; that is a real option with a different economic and operational profile, and it is out of scope for this course, whose audience meets the access layer as Wi-Fi.