How WiFi Broke Through the Contention Ceiling — And Why It Took 20 Years

Last week we hit a wall. We proved that CSMA/CA — the distributed contention protocol that has governed WiFi since 1997 — degrades sharply once dozens of devices compete for the same channel. We traced the evolution of medium access from ALOHA through CSMA/CD to CSMA/CA and watched State grow at every step while Coordination stayed stubbornly distributed. The ceiling, we concluded, isn’t a state limit. It’s a coordination limit.

So what did WiFi actually do about it? And why did it take twenty years?

Before we trace the history, let’s anchor the framework. WiFi operates in unlicensed spectrum. That single fact is the binding constraint — no single entity has authority over the channel. From that constraint, the invariant answers cascade:

Coordination is locked to distributed. No one can tell anyone else when to transmit.
State is forced to be local. No station has global knowledge of who else is contending.
Time is coarse. No global clock, no synchronized scheduling — only local sensing and random backoff.
Interface carries minimal coordination information — just “busy or idle” from carrier sense, plus ACK presence.

CSMA/CA is not a design choice under these constraints — it is the only viable MAC. And the contention ceiling is its inevitable limit. Everything that follows is the system trying to escape that ceiling without changing the binding constraint — until density forced the constraint itself to change.

Making the pipe wider

WiFi’s first instinct was the obvious one: if CSMA/CA wastes most of the channel under contention, make the channel so large that even 30% of it is enough for everyone. That meant pushing more bits per second through the same radio spectrum.

The tool for this is modulation — how digital bits are encoded onto an analog radio signal. The original mainstream WiFi, 802.11b from 1999, used a scheme called Complementary Code Keying (CCK) and topped out at 11 megabits per second. When 802.11a arrived the same year with Orthogonal Frequency Division Multiplexing (OFDM) — splitting the channel into many narrow subcarriers, each carrying its own data, and far more robust against the multipath reflections that bounce off walls and furniture indoors — it could use modulation up to 64-QAM (Quadrature Amplitude Modulation), packing 6 bits into every transmitted symbol. That got you 54 Mbps.

But higher-order modulation comes at a cost. The constellation points — the discrete signal states the receiver must distinguish — get closer together as you pack more bits per symbol. 64-QAM needs about 26 decibels of Signal-to-Noise Ratio (SNR) at the receiver. The simplest modulation, BPSK (Binary Phase Shift Keying), needs only about 7 dB. Every step up in modulation order demands a progressively cleaner signal.

What this means in practice is that your WiFi speed isn’t fixed — it’s a function of channel quality. As you walk away from the access point, the signal weakens, SNR drops, and the system drops to a simpler modulation: 16-QAM, then QPSK, then BPSK. This is rate adaptation: the transmitter selects the highest modulation order the current SNR can support, using feedback from the receiver — missing ACKs, channel quality reports, or packet reception statistics. The decision is made per-packet, so the “speed” on your WiFi status bar is the output of this feedback loop, varying moment to moment as you move.

By 802.11ac in 2013, WiFi had reached 256-QAM — 8 bits per symbol, requiring about 25 dB SNR. 802.11ax pushed to 1024-QAM — 10 bits per symbol, needing roughly 35 dB. Each step up in modulation order squeezes about 25% more bits per symbol, though total spectral efficiency gains are larger when combined with other PHY changes like narrower subcarrier spacing (more tightly packed frequency slices within the channel, reducing guard-band waste). But the SNR bar keeps rising. So how do you get better SNR without cranking up the transmit power — which is capped by regulation and battery constraints?

The answer is antennas. Multiple-Input Multiple-Output (MIMO) uses arrays of antennas at both transmitter and receiver for two independent benefits. Spatial diversity lets the receiver combine copies of the same signal arriving via different propagation paths — reflections off walls, ceilings, furniture — averaging out fading and improving effective SNR without increasing transmit power. Spatial multiplexing goes further: with enough antennas and sufficient scattering in the environment, you can transmit independent data streams simultaneously on the same frequency. Four antennas can carry up to four independent spatial streams — each stream carries its own data, so four streams quadruple the effective data rate.

802.11n introduced 4×4 MIMO in 2009 — up to four spatial streams, reaching 600 Mbps. 802.11ac went to 8×8 and added downlink Multi-User MIMO (MU-MIMO): the AP uses beamforming — shaping the combined antenna pattern to focus energy toward a specific client by adjusting the phase of each antenna element, using channel measurements the client feeds back after a training exchange — to direct different spatial streams to different clients simultaneously, serving up to four users in parallel rather than one at a time. Channel width grew in lockstep — 20 MHz in 802.11a/g, 40 in 11n, 80 or 160 in 11ac — with wider channels providing more OFDM subcarriers and more data per symbol period.

Multiply it together: 8 spatial streams × 256-QAM × 160 MHz channel = a peak PHY rate of 6.9 Gbps in 802.11ac. The engineering achievement is real, and it required all three advances — modulation, MIMO, and channel width — to work in concert. Higher-order modulation without MIMO can’t meet the SNR requirement. MIMO without higher-order modulation wastes spatial capacity. Wider channels without MIMO create more noise to overcome.

But here’s the paradox. 802.11ac cracked the downlink: the AP wins a CSMA/CA contention round, then beamforms to four users in parallel during that single Transmit Opportunity. Downlink delivery is genuinely parallel. But the uplink is untouched — every client still contends one at a time via CSMA/CA, exactly as before. And the AP itself must fight those same clients just to get a turn to start its MU-MIMO burst. Downlink parallelism exists, but it’s gated by a serial access bottleneck and works in only one direction. The 6.9 Gbps peak assumes the AP has the channel — getting the channel is still a contention problem, and the uplink is still one user at a time.

The overhead tax

This is where the pre-lecture exercise comes in.

Your exercise used 100 microseconds as a round number for protocol overhead. The real overhead — including average backoff — is closer to 160 microseconds. That makes the picture even worse than what you calculated. The protocol overhead per frame — DIFS (34 µs), average random backoff (67.5 µs at CW_min=15 with 9 µs slots), PHY preamble (20 µs), SIFS (16 µs), and the ACK sent at the basic rate (~24 µs) — totals roughly 160 microseconds. This number is essentially fixed regardless of PHY rate, because DIFS and SIFS are set by radio turnaround physics, and the ACK is transmitted at the lowest common rate for backward compatibility.

At 54 Mbps, a 1500-byte frame takes about 222 microseconds to transmit. The 160 microseconds of overhead is substantial but the data still dominates — roughly 58% of airtime carries actual payload. Not great, but workable. At 600 Mbps with MIMO, that same frame takes just 20 microseconds. The overhead is now 8× the payload; only about 11% of airtime is actual data. At 6.9 Gbps peak, the frame takes 1.7 microseconds. The overhead is nearly 100× the payload. Data fraction: barely 1%.

The PHY engineers gave us 6.9 gigabits per second, and the MAC wastes 99% of it on protocol gaps. Making the pipe wider didn’t help — it made the overhead ratio catastrophically worse. The data shrunk to a sliver inside a vast protocol envelope. This is the overhead tax, and it’s distinct from the contention ceiling we studied last week. The contention ceiling comes from collision probability growing with the number of stations. The overhead tax hits even a single station on an empty channel. Both limit throughput, and both stem from the same root: CSMA/CA’s per-frame protocol structure was designed when frames took milliseconds, not microseconds.

Aggregation: the MAC’s first response

802.11n’s MAC engineers understood the overhead tax and designed a direct counter: frame aggregation. Instead of paying the 160-microsecond tax for every frame, bundle multiple frames and pay it once.

The Aggregated MAC Protocol Data Unit (A-MPDU) takes up to 64 individual frames, each with its own MAC header, and wraps them under a single PHY preamble — the 20-microsecond training sequence the receiver needs for synchronization and channel estimation. Without aggregation, that 20 µs preamble is paid for every single frame regardless of size. With aggregation: one preamble, one DIFS, one backoff, one channel access — then a burst of data frames. The receiver confirms the entire burst with a single Block ACK — a bitmap where each bit corresponds to one frame in the burst, marking it as received or lost — rather than 64 individual acknowledgments. The transmitter can then selectively retransmit only the frames that failed. The earlier 802.11e amendment had introduced Transmit Opportunity (TXOP), which lets a station hold the channel for a burst duration after winning contention, instead of re-contending after every frame.

Together, aggregation, Block ACK, and TXOP amortize the 160-microsecond overhead across many payloads. Aggregate 64 frames and the per-frame overhead drops from 160 µs to about 2.5 µs. The data fraction recovers to healthy levels. In the framework’s language, aggregation is a Time invariant fix. It changes the effective timescale of the protocol — amortizing overhead across a burst rather than paying it per frame. But Coordination remains distributed. State remains local. The binding constraint — unlicensed spectrum, no authority — hasn’t changed. The ceiling is still there; aggregation just raised the floor beneath it.

This was enough for the 802.11n and 802.11ac era — homes, small offices, coffee shops with five to ten devices per AP. Aggregation gave MIMO a way to actually deliver its PHY gains to users.

But there’s a deployment scenario where aggregation still fails.

Density breaks the fix

Aggregation fixed the per-frame overhead. It did not fix contention. Let’s quantify why.

In L6 we described the random backoff: a station picks a random number from [0, CW-1] and counts down that many slots before transmitting. Define τ as the probability that a station transmits in any given slot — not “per attempt,” but per slot. If I freeze time at a random slot boundary, τ is the chance that station i’s backoff counter happens to be at zero.

If every station is at CW_min (no collisions have occurred yet), each station picks a backoff uniformly from [0, CW_min], waits on average CW_min/2 slots, and transmits:

τ ≈ 2 / (CW_min + 1)

For 802.11a/g with CW_min = 15: τ ≈ 2/16 = 0.125. In any given slot, there’s a 12.5% chance a station transmits.

Now ask: in this one slot, what’s the probability that exactly one station transmits and all others stay silent? That’s a successful, collision-free slot:

P(success) = n · τ · (1 - τ)^(n-1)

The logic: pick one station to be the transmitter (probability τ), require all other n-1 stations to stay silent in this same slot (each with probability 1-τ), and since any of the n stations could be the one, multiply by n.

At n = 10: P(success) = 10 × 0.125 × (0.875)^9 ≈ 0.40. About 40% of slots succeed — workable.

At n = 50: P(success) = 50 × 0.125 × (0.875)^49 ≈ 0.006. Less than 1% of slots succeed.

At n = 250: P(success) ≈ 0. Virtually no slot produces a successful transmission.

Now, an important caveat about these numbers. We used τ = 0.125, which assumes every station is sitting at CW_min — meaning no collisions have happened. That’s only realistic at low load. At n = 250, collisions are constant, BEB doubles CW after each one, and the average window across all stations grows far beyond CW_min. The real stationary τ is much smaller than 0.125. Our calculation overestimates how aggressively stations transmit, making the collision picture look worse than steady-state reality.

But the qualitative conclusion survives. Bianchi’s full model solves for the stationary τ where collision probability and BEB backoff growth reach equilibrium — a system of coupled equations. Even at that equilibrium, throughput at n = 50 drops to roughly 40-55% of channel capacity, and continues falling at higher density. BEB slows the collapse but cannot prevent it. The curve’s shape — graceful at low n, steep decline above ~30 stations — is structural, not tunable.

And there’s an asymmetry that makes it worse. The AP carries the downlink traffic for all clients — in typical web browsing, 10 to 100 times more data than any single client sends upstream. But under CSMA/CA, the AP contends on equal terms with every client. One station, one contention slot, same probability as everyone else. The AP, holding 95% of the data, gets 1/(n+1) of the channel access opportunities. Downlink starves. Users experience massive latency and packet loss, even on a PHY theoretically capable of gigabits.

This is not just “more devices.” Density changes the binding constraint itself. The original constraint was: no authority over the spectrum, therefore coordinate distributedly. The new constraint is: coordination overhead dominates capacity. When the cost of distributed coordination — backoff, collision, re-backoff — exceeds the cost of the data itself, distributed coordination is no longer viable regardless of how clever the protocol is. The binding constraint has shifted from “no authority” to “coordination cost exceeds capacity.” That shift forces the Coordination invariant to change — from distributed to centralized. Not as a design preference, but as a structural necessity.

This wasn’t just a theoretical prediction — it was observable in the wild. In 2012, while I was a grad student at NC State, we were studying WiFi performance in large lecture halls and conference venues. The symptoms were obvious to anyone who’d tried to use WiFi at a packed conference: pages wouldn’t load, video buffered endlessly, everything felt broken — even though the AP’s PHY rate was nominally hundreds of megabits per second.

We traced the root cause to exactly the asymmetry I just described. The AP was starving for channel access. It had a massive downlink queue but could only transmit when it won a contention round against hundreds of clients — each of whom had almost no data to send but held equal contention rights. The protocol’s fairness, designed for a world of five devices, was the source of failure in a world of five hundred.

Our solution, WiFox (published at ACM CoNEXT 2012), was a software-only mechanism that dynamically adjusted the AP’s channel-access priority based on its queue depth. When the AP’s buffer was growing — a signal that downlink was starving — WiFox shortened the AP’s backoff window, effectively giving it priority over clients. The results were dramatic: 400 to 700 percent improvement in downlink throughput, 30 to 40 percent reduction in response time, deployable as a firmware update with no hardware or client changes.

But notice what WiFox actually was, in the framework’s terms: a partial Coordination shift. We gave the AP privileged access — it no longer competed on equal terms. We broke CSMA/CA’s per-station fairness because that fairness was the source of failure. WiFox was a step toward centralization. The full step came with 802.11ax.

WiFi centralizes

In 2021, 802.11ax shipped — marketed as WiFi 6 — and changed the MAC architecture fundamentally. Instead of stations contending for the whole channel one at a time, the AP divides the channel into Resource Units (RUs) — think of them as time-frequency slots that the scheduler can assign independently, like cells in a spreadsheet where each column is a frequency slice and each row is a time interval. Each RU consists of a group of OFDM subcarriers. The smallest RU is 26 subcarriers, about 2 MHz of bandwidth. In a 160 MHz channel, the AP can create dozens of RUs and assign each to a different client.

The mechanism is a Trigger Frame — a specialized downlink control frame that the AP broadcasts to coordinate uplink transmissions. It carries per-client metadata: which client is assigned to which RU, what modulation and coding scheme to use, how much transmit power to apply, and how long the transmission window lasts. Clients cannot transmit on OFDMA RUs without receiving a Trigger Frame first — that’s the centralization made concrete. Upon receiving it, all assigned clients begin transmitting exactly one SIFS (16 µs) later, synchronized in time and frequency across their respective RUs. Multiple clients transmit simultaneously on non-overlapping frequency slices. Zero contention on the data path.

Follow the logic. Coordination must be centralized. Someone needs a global view — that’s the AP. The AP must allocate resources deterministically, eliminating contention. It must partition the channel and assign slices. That structure is Orthogonal Frequency Division Multiple Access (OFDMA). WiFi didn’t borrow OFDMA from cellular LTE (which has used it since 2009) out of convenience. It arrived at OFDMA because the constraint shift forced the same invariant answers that cellular had reached from a different starting point. The AP is now a scheduler, not a contender.

CSMA/CA is retained for exactly two purposes: serving legacy clients that don’t speak 802.11ax, and the initial association process — the moment when a new device joins the network and the AP doesn’t yet know it exists. For everything else, it’s scheduled access. The throughput improvement is real: per-user throughput under dense contention improves by several times compared to CSMA/CA, because collision overhead is eliminated entirely on the OFDMA path.

But let’s be honest about the cost. WiFi lost its defining character: any station can transmit whenever it wants, no permission needed, no infrastructure required. Under 802.11ax, the high-throughput path requires an AP scheduler. No AP, no OFDMA. Ad-hoc networks, mesh networks, infrastructure-free operation — all fall back to CSMA/CA and its limitations. In exchange, WiFi gained channel utilization above 70%, multi-user parallel transmission, more deterministic latency, and elimination of hidden terminals. The invariant shift is clean: State moved from local-only to AP-aggregated; Coordination moved from fully distributed to centralized. The AP became a base station.

Why did this take twenty years?

This isn’t a story of engineers being slow. Three non-technical forces held centralization back.

Unlicensed spectrum means no operator. In cellular, a carrier owns the spectrum and deploys the scheduler. WiFi operates in unlicensed bands — anyone can buy an AP, and APs from different manufacturers must interoperate. There’s no single entity to run a scheduler. The answer — let the AP be the scheduler — only became obvious when APs became ubiquitous infrastructure rather than optional accessories.

Backward compatibility is the second force. An 802.11ax AP must still serve an 802.11b client from 1999. You cannot remove CSMA/CA. You can only layer scheduling on top. This adds complexity and constrains what the scheduler can do.

The third force is the most important: deployment context. Before roughly 2012, a home AP served three to five devices. CSMA/CA’s limitations were irrelevant. The contention ceiling was a theoretical result, not a lived experience. It took lecture halls, airports, and stadiums — environments with hundreds of devices per AP — to turn the math into a visible problem. Technology adoption follows deployment pressure, not theoretical possibility. The engineers could have centralized in 2005. The market didn’t need it until 2015.

The other path

WiFi’s twenty-year journey from distributed contention to centralized scheduling has a mirror image: cellular never had to make that journey at all.

Licensed spectrum is the mirror binding constraint. Where WiFi’s constraint — unlicensed spectrum — locked Coordination to distributed, cellular’s constraint — licensed spectrum, exclusive operator ownership — locks Coordination to centralized from day one. One entity owns the channel, having paid billions at auction for exclusive access. The base station is the sole authority over every transmission. There is no contention for data, ever. State can be global (the BS knows every registered device). Time can be precise (the BS synchronizes all transmissions). Interface can carry rich scheduling information (because the BS controls both ends).

But cellular had its own evolution. The question was never “when to centralize” — it was “how to schedule efficiently as traffic changed from voice to data.”

First generation cellular, FDMA (Frequency Division Multiple Access) in the 1980s, divided the spectrum into 30 kHz channels and assigned one per call — about 416 channels across the system, roughly 60 per cell after frequency reuse. Hard capacity: call number 61 is blocked, not degraded. And average voice activity is only about 35%, so two-thirds of allocated spectrum carries silence that nobody else can use.

Second generation GSM used TDMA (Time Division Multiple Access): eight time slots per 200 kHz carrier, each slot 576.9 microseconds in a 4.615 ms frame. More users per carrier, better voice codecs. But still hard capacity — once every slot is assigned, the next caller is rejected.

CDMA (Code Division Multiple Access) changed the paradigm entirely. All users share the same frequency at the same time, separated by orthogonal spreading codes — unique high-rate bit sequences assigned to each user. “Orthogonal” means these codes have zero cross-correlation: when the receiver multiplies the combined airwave signal by a specific user’s code, all other users’ contributions average to near-zero, leaving only the intended data. The result is that each user’s signal looks like low-level noise to every other user. Adding a user doesn’t block a channel — it raises the noise floor for everyone. Soft capacity instead of hard. The tradeoff: power control becomes the medium access mechanism. Without it, a mobile 100 meters from the base station arrives 10,000× stronger (40 dB) than one a kilometer away, masking the distant user entirely — the near-far problem. The base station compensates by adjusting each mobile’s transmit power 800 times per second to keep all signals arriving at equal strength. If one user’s power control fails, it drowns everyone. The soft-capacity model was the conceptual precursor to data scheduling: if resources can be shared dynamically rather than allocated in fixed blocks, you can match allocation to demand on short timescales.

All three systems — FDMA, TDMA, CDMA — were designed for voice: steady, symmetric traffic well-served by dedicated circuits. What happens when the traffic becomes bursty? Loading a webpage involves a few hundred milliseconds of data transfer, then seconds of idle reading. Under the circuit model, you hold a dedicated channel the entire time. The channel carries data for a fraction of a second, then silence for ten seconds, and nobody else can use it.

HSDPA (High Speed Downlink Packet Access), deployed around 2005, was the bridge. It replaced dedicated per-user channels with a shared pool that the base station schedules every 2 milliseconds — a Transmission Time Interval (TTI). Think of TTI as the cellular answer to the Time invariant: how often does the scheduler re-evaluate? Each TTI, the scheduler picks the users with the best channel conditions — reported via Channel Quality Indicator (CQI) feedback, the cellular answer to the State invariant — and gives them the pool’s resources. This proved the concept: per-TTI packet scheduling works at millisecond timescales. LTE built its entire architecture around OFDMA — the same OFDMA that WiFi would borrow in 802.11ax twelve years later.

The universal architecture

Here’s one of the most satisfying observations in this whole landscape. GSM is a fully centralized system — zero contention for data. But there’s one moment when centralization fails: when a brand-new device powers on. The base station doesn’t know it exists. It can’t schedule what it hasn’t discovered. So the device announces itself in a special slot called the Random Access Channel (RACH), using slotted ALOHA — the 1972 protocol, running inside a 1991 TDMA system.

Could you eliminate RACH entirely? Could you build a system with zero contention at every level? No. Any system with dynamic membership needs at least one contention-based moment: the discovery moment. You can centralize everything after discovery, but discovery itself is irreducibly contention-based. 802.11ax has the same pattern: OFDMA for data, CSMA/CA for initial association. Contention is the tax you pay for letting new participants join without prior arrangement.

Beyond that one contention moment, everything in cellular is scheduled. And the question becomes: how fast can the scheduling loop run? That question turns out to unify the entire landscape.

Every wireless MAC is a control loop: measure the channel, allocate resources, transmit, measure the outcome, reallocate. The speed of that loop — how fast the system can observe and react — determines the architecture’s ceiling. This is different from the overhead tax, which is about how much airtime each individual attempt wastes. The loop speed is about how quickly the system adapts to changing conditions.

CSMA/CA measures the channel via carrier sense, a local binary observation, with a loop period of roughly 12 milliseconds at legacy rates. GSM’s TDMA loop runs at about 4.6 milliseconds. CDMA’s power control loop runs at 800 Hz — every 1.25 milliseconds, fast enough to track vehicular fading. LTE OFDMA measures Channel Quality Indicator (CQI) — each device reports how good its channel is on each frequency slice — and the base station reallocates resource blocks (the LTE equivalent of 802.11ax’s RUs) every millisecond. 802.11ax uses a similar feedback mechanism at comparable timescales.

Faster loop, tighter coordination, higher utilization. CSMA/CA had the slowest loop in the landscape. That’s the structural reason contention broke — and why centralization, with its faster and richer feedback, was the destination for both WiFi and cellular.

Convergence

After thirty years of divergence, WiFi and cellular arrived at the same architecture. WiFi started fully distributed, fought through twenty years of PHY improvements and MAC patches, and finally centralized under density pressure. It borrowed OFDMA from LTE. Cellular started centralized from day one, but when it entered the unlicensed 5 GHz band through Licensed-Assisted Access (LAA) in 2016, it had to coexist with WiFi devices it couldn’t control. It couldn’t schedule transmissions in spectrum it didn’t own. So it adopted Listen Before Talk (LBT) — carrier sensing, the WiFi mechanism.

Each side borrowed the other’s core technique, for the exact reason the other side had adopted it originally. WiFi borrowed centralized scheduling because density demanded it. Cellular borrowed contention because shared spectrum demanded it. And both retain contention for exactly one purpose: bootstrap. Cellular uses RACH for device discovery; WiFi retains CSMA/CA for initial association. The universal architecture is: schedule everything you can, contend only for discovery.

The destination was determined by physics and density. The path depended on where you started.

Exercise solutions: the overhead tax (from the pre-lecture email)

The email gave students the following setup: fixed protocol overhead per frame ≈ 100 µs, regardless of PHY rate. Frame size: 1500 bytes. Compute the data fraction at three PHY rates.

The 100 µs was a deliberate simplification — it omits average backoff. The precise breakdown at 802.11a/g rates:

Component	Value	Why it’s fixed
DIFS	34 µs	Radio turnaround physics
Average backoff	67.5 µs	CW_min=15, avg=7.5 slots × 9 µs/slot
PHY preamble	20 µs	OFDM training symbols
SIFS	16 µs	Radio RX-to-TX switch
ACK at 6 Mbps basic	~24 µs	14 bytes + preamble at lowest mandatory rate
Total	~162 µs	None of these scale with data-rate PHY improvements

For the email exercise (using 100 µs as given), here are the precise solutions:

Q1 — 54 Mbps (802.11g):

Frame TX time = (1500 × 8 bits) / (54 × 10⁶ bits/s) = 12,000 / 54,000,000 = 222 µs

Total airtime per frame = 222 + 100 = 322 µs

Data fraction = 222 / 322 = 68.9%

About two-thirds of airtime carries payload. The overhead is significant but the data dominates.

Q2 — 600 Mbps (802.11n, 4×4 MIMO):

Frame TX time = 12,000 / (600 × 10⁶) = 20.0 µs

Total = 20 + 100 = 120 µs

Data fraction = 20 / 120 = 16.7%

The PHY is 11× faster, but data fraction dropped from 69% to 17%. Five-sixths of airtime is now overhead.

Q3 — 6.9 Gbps (802.11ac peak, 8×8 MIMO, 256-QAM, 160 MHz):

Frame TX time = 12,000 / (6.9 × 10⁹) = 1.74 µs

Total = 1.74 + 100 = 101.74 µs

Data fraction = 1.74 / 101.74 = 1.7%

Over 98% of airtime is protocol overhead. The PHY can push 6.9 Gbps but the station spends virtually all its time in protocol gaps.

Q4 — Does faster PHY solve the contention ceiling?

No. The overhead is fixed in microseconds and doesn’t scale with PHY rate. As the PHY gets faster, the data shrinks relative to the protocol envelope. The per-frame overhead was designed when a frame took milliseconds; at gigabit rates, the frame takes microseconds but the overhead doesn’t shrink.

Two fixes exist: (1) amortize the overhead across multiple frames (aggregation — A-MPDU bundles up to 64 frames under one preamble), or (2) eliminate per-frame contention entirely (centralized scheduling — OFDMA).

With the corrected 160 µs overhead, the numbers are even starker:

PHY rate	Frame TX	Overhead	Data fraction
54 Mbps	222 µs	160 µs	58%
600 Mbps	20 µs	160 µs	11%
6.9 Gbps	1.74 µs	160 µs	1.1%

The qualitative insight is identical; the corrected numbers make the case stronger.

In-class exercise: CSMA/CA vs OFDMA at scale

Setup. You’re an AP in a stadium section. 250 phones in your coverage area. 160 MHz channel. PHY rate: 600 Mbps.

Part 1 — Compute τ and collision probability.

Each station’s transmission probability per slot: τ ≈ 2/(CW_min + 1). For CW_min = 15: τ = 2/16 = 0.125.

The probability of a successful (collision-free) slot with n stations is: P(success) = n · τ · (1-τ)^(n-1).

Compute P(success) at n = 10, n = 50, and n = 250.

Expected answers:

n=10: P = 10 × 0.125 × (0.875)⁹ = 1.25 × 0.316 = 0.40 — 40% of slots succeed.
n=50: P = 50 × 0.125 × (0.875)⁴⁹ = 6.25 × (0.875)⁴⁹. Compute (0.875)⁴⁹: ln(0.875) = -0.1335, so 49 × (-0.1335) = -6.54, e^(-6.54) ≈ 0.00144. P = 6.25 × 0.00144 = 0.009 — less than 1% of slots succeed.
n=250: (0.875)²⁴⁹ ≈ e^(-33.3) ≈ 4×10⁻¹⁵. P ≈ 0 — virtually no successful slot.

Note: these are first-attempt probabilities at CW_min. BEB responds by growing CW after collisions, reducing τ and reaching a balance — but at the cost of long backoff waits and sharply reduced throughput.

Part 2 — Compute OFDMA’s capacity.

We described how 802.11ax uses Trigger Frames to assign Resource Units to individual users. Now quantify the parallelism.

Same 160 MHz channel. Each RU is 26 subcarriers (~2 MHz). How many parallel RUs? A naive division gives 160/2 = 80, but the actual number is 74. The missing six RUs’ worth of subcarriers are consumed by physics: guard bands at the channel edges prevent spectral leakage, DC null subcarriers at the center protect against hardware impairments, and inter-RU gaps prevent adjacent frequency slices from interfering with each other.

So each Trigger Frame serves up to 74 users simultaneously — zero contention, zero collision.

Now put the two side by side. Under CSMA/CA with n=250 (Part 1), virtually no slot produces a successful transmission — the system is choked by collisions. Under OFDMA, every single Trigger Frame successfully serves up to 74 users in parallel. CSMA/CA gets 1 user per successful access (which almost never happens at this density). OFDMA gets up to 74 per access (which succeeds every time, because there’s no contention).

There’s a tradeoff to note: under CSMA/CA, the winning user gets the full 160 MHz channel. Under OFDMA with 74 RUs, each user gets only ~2 MHz. Parallelism increases but per-user bandwidth shrinks. In sparse scenarios — five devices on a home AP — CSMA/CA’s “winner takes all” is actually fine, and OFDMA’s scheduling overhead may not be worth it. But at 250 users, 2 MHz of guaranteed, collision-free access per trigger beats fighting over 160 MHz and getting nothing. Density is what flips the tradeoff.

Part 3 — The insight.

The difference isn’t incremental. It’s structural. CSMA/CA at n=250 produces effectively zero throughput because distributed coordination collapses. OFDMA at n=250 serves everyone in milliseconds because centralized scheduling eliminates contention entirely.

Which invariant changed? Coordination — from distributed to centralized. What forced it? Density shifted the binding constraint from “no authority” to “coordination cost exceeds capacity.” Could you have predicted the need for something like OFDMA from the constraint shift alone? Yes — once coordination must be centralized, the AP needs a global view and must partition resources deterministically. That structure is OFDMA.

The invariant shift

WiFi’s invariant answers changed across this story. State went from local-only — each station tracking its own backoff counter — to AP-aggregated, with the AP holding per-client channel quality, buffer status, and scheduling decisions. Time went from frame-time-bounded feedback at roughly 12 milliseconds to scheduling-cycle feedback at 1 to 5 milliseconds. Coordination — the invariant that stayed fixed through the entire evolution from ALOHA to DCF — finally changed. Fully distributed became centralized. Breaking this invariant is what broke the ceiling. Interface didn’t change. 802.11ax still speaks 802.11 frames. Backward compatibility locked the interface, and it’s the reason the transition took twenty years, not five.

The lesson is structural, not historical. When density shifts the binding constraint, the invariant answers converge — regardless of starting point. WiFi and cellular arrived at the same State (AP/BS-aggregated), the same Time (millisecond scheduling loops), the same Coordination (centralized), and the same Interface (OFDMA resource allocation). Two different binding constraints, two different twenty-year paths, one destination. The framework predicted it: identical constraints force identical invariant answers.

But centralization creates its own problem: the scheduler is now a single point of complexity — and of failure. How do you build, scale, and open up that scheduler? That’s the infrastructure question, and it’s next.