Communications Systems: A Practical Reference

Radio spectrum is a legal object as much as a physical one. The ITU Radio Regulations are a treaty that divides everything from 8.3 kHz upward into allocations by service, separately for three world Regions, and the treaty is renegotiated at a World Radiocommunication Conference roughly every four years. National regulators then assign licenses inside those allocations, which in the United States means the FCC for non-federal users and NTIA for federal ones, both printed on the same frequency allocation chart. Physics decides what each band is good for. Below about 1 GHz a wave diffracts around buildings and passes through ordinary walls, so one site covers many kilometers; above about 20 GHz it does neither, so a site covers a few hundred meters and needs line of sight. Channel width runs the other way, because 100 MHz is a seventh of all the spectrum below 700 MHz and an ordinary block at 28 GHz, so low band gives coverage, mid band from 1.7 to 4.2 GHz carries most of the traffic, and millimeter wave gives peak rate in a small area.

The system's real product is harmonization. When most countries allocate the same band to the same service, a chipmaker builds one radio for a global market, and that scale is why a 5G modem covering dozens of bands is a commodity part costing tens of dollars rather than a separate design per country. Allocation also gives interference protection with legal force, which is what lets an operator plan a network against a known noise floor instead of whatever the neighbors do. The weakness is that moving a band takes about a decade and the incumbents have to be paid. Clearing 280 MHz of C-band took an FCC proceeding opened in 2018, an auction that closed in February 2021, and $9.7 billion in accelerated relocation payments to satellite operators on top of the auction price. The other weakness is path dependence: a band is only worth what handsets already support, so a newly allocated band with no device ecosystem sells cheap even when the physics is good.

If you need coverage per site, pick low band at 600 to 900 MHz and accept 5 to 20 MHz channels. If you need capacity across a city, pick mid band at 1.7 to 4.2 GHz, where 100 MHz channels are available and a macro site still reaches roughly a kilometer. If you need peak rate in a stadium or a plaza and can afford a site every few hundred meters, use millimeter wave at 24 to 47 GHz, where 400 to 850 MHz blocks are routine. If you cannot buy an exclusive license at all, use unlicensed spectrum at 5 and 6 GHz, where the FCC opened 1,200 MHz in 2020, or shared spectrum such as CBRS, and design for interference you do not control. For satellite, C-band at 4 and 6 GHz survives rain, Ku at 11 and 14 GHz is the volume band for consumer and enterprise service, and Ka at 20 and 30 GHz is where high-throughput capacity is, so pick by rain climate and by how much fade margin the link budget can carry. Below 30 MHz, HF is the only band that reaches over the horizon with no infrastructure, at a few kilobits per second.

Low band 600–900 MHz, mid band 1.7–4.2 GHz, millimeter wave 24–47 GHz · Auction 107 sold 280 MHz of C-band for $81.2 billion in gross bids, about $0.88 per MHz-POP · Auction 1002 sold 70 MHz at 600 MHz for $19.8 billion, about $0.86 per MHz-POP · Auction 110 sold 100 MHz at 3.45 GHz for $22.5 billion, about $0.68 per MHz-POP · Auction 103 sold 3,400 MHz at 37, 39 and 47 GHz for $7.6 billion, under one cent per MHz-POP · CBRS priority access licenses cleared about $0.21 per MHz-POP · winners paid $9.7 billion in accelerated relocation payments on top of the C-band price (per-MHz-POP figures derived here from FCC gross bid totals).

FCC Auction 107 (3.7–3.98 GHz C-band, $81.2 billion, 5,684 licenses, closed February 2021), Auction 1002 (the 600 MHz forward auction, $19.8 billion for 70 MHz, closed 2017, funded by $10.1 billion paid to broadcasters in the paired reverse auction), Auction 110 (3.45 GHz, $22.5 billion), Auction 105 (CBRS priority access, $4.9 billion for 20,625 county-level licenses) and Auction 103 (upper 37, 39 and 47 GHz, $7.6 billion for 3,400 MHz). Those totals are gross bids; net proceeds run several percent lower after bidding credits, so Auction 105 appears as $4.86 billion gross and $4.54 billion net. The NTIA frequency allocation chart and ITU Radio Regulations Article 5 are the two reference tables everything else points at.

Spectrum is a balance sheet asset with no maintenance cost and no depreciation in the ordinary sense, which is why carriers borrow against it and why auction totals dwarf the equipment budgets they enable. The arithmetic that matters is dollars per megahertz per covered person: $81.2 billion for 280 MHz across roughly 330 million people is about $0.88 per MHz-POP, and $19.8 billion for 70 MHz across the same population is about $0.86. Those two numbers landing within a few cents of each other is the useful result, because it says buyers valued coverage spectrum and capacity spectrum about equally while millimeter wave, at under a cent per MHz-POP in Auction 103, was valued at roughly a hundredth as much. Between those poles, 100 MHz at 3.45 GHz cleared for $22.5 billion, about $0.68 per MHz-POP, and CBRS priority access licenses cleared at about $0.21, discounted because those licenses are county-sized, ten years long, and shared with naval radar. The money flows to the US Treasury, minus what goes to incumbents: broadcasters received $10.1 billion to vacate 600 MHz and satellite operators $9.7 billion to clear C-band early. For an operator the practical consequence is that spectrum is usually the largest single line item in a decade of capital spending, it is paid years before the first customer, and buildout conditions attached to the license mean it cannot simply be held.

3 min5G Frequency Bands: What Are They Used For?Wireless Future · 5k+ views

17 minUnderstanding the Radio Frequency Spectrum (#715)David Casler Ask Dave · 100k+ views

United States Frequency Allocation Chart (National Telecommunications and Information Administration) · Auction 107: 3.7 GHz Service - Lower C-Band (Federal Communications Commission)

Free-space path loss is the baseline every radio link starts from: loss in decibels equals 32.44 plus 20 log10 of the distance in kilometers plus 20 log10 of the frequency in megahertz. Doubling the distance costs 6 dB and doubling the frequency costs another 6 dB, so a 1 km link loses about 89 dB at 700 MHz and about 121 dB at 28 GHz. The frequency term comes from the receiving antenna's effective area shrinking with the square of the wavelength rather than from the wave itself weakening, which is why a fixed-size dish actually gains at high frequency while a fixed-gain handset antenna loses. Real environments are worse than free space in most cases and better in a few. The 3GPP channel models fit a distance exponent to measured data: 2.0 in free space, about 2.2 for an urban macro cell with line of sight, 3.9 once the line of sight is blocked, 3.5 in an urban street canyon without line of sight, and 1.7 indoors along a corridor, where the walls act as a waveguide. On top of that sits shadow fading, a log-normal spread with a standard deviation of 4 to 8 dB that describes how much two locations at the same distance differ.

The models are good enough to plan a network with, and that is a real achievement: an operator can predict median coverage over a city to within a few decibels and size a site grid from a map. What they do not predict is any individual location, because the 4 to 8 dB shadow fading standard deviation means a house behind a hill or a desk behind a metal partition can be 15 dB off the median. The sharpest failure is material penetration, which gets far worse with frequency. Using the 3GPP figures, concrete costs 5 + 4f dB with f in gigahertz, so a wall takes about 8 dB out of a 700 MHz signal and about 117 dB out of a 28 GHz one, which is the whole reason millimeter wave does not serve indoor users from outdoors. Ordinary glass is transparent by comparison at 2 + 0.2f dB, about 8 dB at 28 GHz, while the infrared-reflective glass in modern buildings costs 23 + 0.3f dB, about 31 dB. Above roughly 10 GHz weather joins in, and above 50 GHz the atmosphere itself does.

Use free-space path loss only for satellite links, microwave hops with clear line of sight, and sanity checks. For terrestrial cellular planning, use the 3GPP TR 38.901 urban macro, urban micro, rural and indoor models, pick the non-line-of-sight branch as the default, and add shadow fading margin for the coverage percentile you actually promise rather than the median. If your link runs above about 10 GHz, budget rain: at a 25 mm/h rain rate the ITU-R model gives roughly 0.7 dB/km at 10 GHz, 2.8 dB/km at 20 GHz, 4.6 dB/km at 28 GHz and 7.2 dB/km at 40 GHz, so a 3 km 28 GHz hop needs about 14 dB of margin for that rain rate alone. Avoid 60 GHz for anything longer than a few hundred meters, because oxygen absorption there costs roughly 15 dB/km on top of everything else, though that same absorption is useful when you want a link that does not travel far enough to interfere. Whatever the band, the link budget is the arithmetic that decides the design: add transmit power and antenna gains, subtract path loss and everything the air and the walls take out, then compare the result against a noise floor that starts at −174 dBm per hertz and works out to −94 dBm across a 100 MHz channel before the receiver's own noise figure. The honest test of a plan is whether the remaining margin survives rain, a wall, and a body at the same time.

Free-space loss = 32.44 + 20 log10(d in km) + 20 log10(f in MHz), so 89 dB at 1 km and 700 MHz against 121 dB at 1 km and 28 GHz · path loss exponent 2.0 free space, about 2.2 urban macro line of sight, 3.9 urban macro blocked, 1.7 indoor corridor · shadow fading standard deviation 4–8 dB · concrete penetration 5 + 4f dB, about 8 dB at 700 MHz and 117 dB at 28 GHz · rain at 25 mm/h roughly 0.7 dB/km at 10 GHz, 2.8 at 20 GHz, 4.6 at 28 GHz, 7.2 at 40 GHz · oxygen absorption roughly 15 dB/km at 60 GHz · thermal noise floor −174 dBm/Hz, so −94 dBm in a 100 MHz channel (rain figures computed here from ITU-R P.838-3).

3GPP TR 38.901, the channel model behind every 5G coverage tool, covering 0.5 to 100 GHz with urban macro, urban micro, rural, indoor office and indoor factory cases; ITU-R P.838 for rain attenuation and P.676 for atmospheric gases; the Longley-Rice irregular terrain model still used for broadcast and federal coordination; the NYU Wireless and Nokia Bell Labs 28 and 73 GHz measurement campaigns that produced the millimeter wave numbers 38.901 fits.

Propagation sets site count, and site count is most of a wireless network's capital cost. Coverage area goes as the square of cell radius, so if a 700 MHz macro cell reaches 5 km and a 28 GHz cell reaches 200 m, covering the same ground takes about 625 times as many sites, which is the arithmetic that confined millimeter wave to stadiums, airports and dense downtown blocks rather than nationwide coverage. That is why low-band spectrum sells for roughly what mid-band sells for while millimeter wave sells for a hundredth as much: buyers are pricing sites they will not have to build. Each macro site carries a ground lease, power, backhaul and a zoning process, so the recurring cost keeps running whether the site is busy or not, and adding margin in software is free while adding it in sites is not. The practical lever is antenna height and gain, both of which buy decibels far more cheaply than transmit power does, since power costs energy continuously and is capped by emission limits. When a coverage plan fails, the usual cause is a penetration or shadowing assumption rather than a mistake in the free-space term.

5 minWhat is Path Loss?Carl Oliver · 50k+ views

14 minWireless propagation losses [Part 2, Fundamentals of mmWave communication]Wireless Future · 5k+ views

ETSI TR 138 901: Study on channel model for frequencies from 0.5 to 100 GHz (ETSI / 3GPP) · Recommendation ITU-R P.838-3: Specific attenuation model for rain for use in prediction methods (International Telecommunication Union)

Shannon's 1948 result gives the ceiling: capacity equals bandwidth times log2 of one plus the signal-to-noise ratio. At 20 dB signal-to-noise that is log2(101), about 6.7 bits per second per hertz, so a 100 MHz channel cannot carry more than roughly 670 Mbps no matter what the radio does. Modulation is how a transmitter approaches that ceiling, by mapping bits onto the amplitude and phase of a carrier: QPSK carries 2 bits per symbol, 16QAM carries 4, 64QAM 6, 256QAM 8, and Wi-Fi 7's 4096QAM carries 12. Each step to a constellation four times denser halves the distance between neighboring points, so it needs roughly 6 dB more signal-to-noise for the same error rate. Coding is the other half, adding redundancy so the receiver can correct the errors that remain, and a long-block LDPC code typically lands within about a decibel of the Shannon bound. In 5G NR the two are chosen together as a modulation and coding scheme, and the standard tables run from 0.2344 bits per second per hertz at the bottom to 5.5547 with 64QAM and 7.4063 with 256QAM.

Adaptive modulation and coding is what lets one radio serve a user 100 m away at 8 bits per symbol and a user at the cell edge at a fraction of a bit, reusing the same hardware. The weakness is that high-order modulation needs a signal-to-noise ratio that only exists near the transmitter, so 256QAM and above are useful across a small part of a cell and contribute little to average throughput. High-order QAM also demands linearity: the amplifier has to reproduce amplitude faithfully, which means backing it off from saturation and losing efficiency, and it demands phase noise low enough that the constellation points do not smear, which gets harder as carrier frequency rises. On the coding side, redundancy costs both spectrum and latency, since a long code block cannot be decoded until it arrives, and that delay is exactly what a low-latency service cannot spend. Retransmission schemes partly hide this by sending extra parity only when the first attempt fails, at the cost of a round trip.

Do not pick one modulation and coding scheme. Measure the channel, feed the quality report back, and let the scheduler choose per user per slot, because the useful range of operating points spans about 30 dB. If your link is power-limited rather than bandwidth-limited, such as a deep-space or satellite Internet-of-things link, use low-order modulation with a very strong code and accept fractions of a bit per symbol. If it is bandwidth-limited, such as a fixed microwave hop or an indoor Wi-Fi link, spend the signal-to-noise on constellation order instead. For the code itself, use LDPC for data at high rates, because its decoder parallelizes and holds throughput at gigabit rates where turbo decoding does not, and use polar codes for short control messages of tens of bits, where LDPC and turbo both perform badly. 5G NR does exactly this: the standard assigns LDPC to the uplink and downlink shared channels and polar codes to the broadcast channel and to downlink control information, which is why turbo codes, the workhorse of 3G and LTE, do not appear in NR at all.

Shannon capacity = bandwidth × log2(1 + signal-to-noise), so 6.7 bits/s/Hz at 20 dB and about 670 Mbps in a 100 MHz channel (derived here) · QPSK 2 bits per symbol, 16QAM 4, 64QAM 6, 256QAM 8, 4096QAM 12 · roughly 6 dB more signal-to-noise per fourfold step in constellation density · 5G NR modulation and coding tables span 0.2344 to 5.5547 bits/s/Hz with 64QAM and reach 7.4063 with 256QAM · LDPC for the shared data channels, polar codes for broadcast and downlink control · long-block LDPC codes typically land within about 1 dB of the Shannon limit.

5G NR as specified in 3GPP TS 38.212, which assigns LDPC to UL-SCH and DL-SCH and polar codes to the broadcast channel and control information; DVB-S2 and DVB-S2X, which brought LDPC to satellite television and carrier-grade satellite links; Wi-Fi 6 and Wi-Fi 7, which added 1024QAM and 4096QAM; the CCSDS turbo and LDPC codes NASA uses for deep space, where the link is power-limited and coding gain substitutes for antenna size.

Coding gain is the cheapest decibel in the whole system, which is why every standards generation spends silicon on it. A decoder is a fixed block of chip area amortized across every unit shipped, while the alternative ways to find 3 dB are doubling transmit power, which costs energy on every site forever, or building 41% more reach into the antenna, which costs steel. That trade is why deep-space missions spend enormous decoding effort to avoid a larger dish, and why a cellular modem devotes a large share of its area to LDPC. The other economic fact is patents. Channel coding is where the essential-patent fights land, since a code is written into the specification and cannot be designed around, so modem vendors pay royalties on a per-device basis into pools that are a meaningful fraction of the chip's price. For a system designer, the practical consequence is that this layer is something you buy rather than build: the standard fixes the codes, the merchant modem implements them, and the differentiation left is in channel estimation and scheduling, which decide which modulation and coding scheme gets picked and therefore what fraction of the theoretical capacity a network actually delivers.

6 minThe Shannon Limit - Bell Labs - Future ImpossibleNokia Bell Labs · 50k+ views

7 minWhat is QAM modulation?Wireless Explained · 100k+ views

A Mathematical Theory of Communication (Bell System Technical Journal) · ETSI TS 138 212: NR; Multiplexing and channel coding (ETSI / 3GPP)

An antenna's gain is set by how large it is compared to the wavelength: gain equals 4π times effective aperture divided by wavelength squared. A 1 m dish at 12 GHz has 0.785 m² of physical area, about 0.47 m² of it effective at 60% efficiency, and a wavelength of 25 mm, which works out to roughly 40 dBi with a beam about 1.75° wide (beamwidth in degrees is roughly 70 times wavelength over diameter). The same dish at 30 GHz gives 8 dB more gain and a beam under a degree, because gain rises with the square of frequency for a fixed aperture. A phased array gets the same aperture a different way, replacing the reflector with many small elements spaced about half a wavelength apart, each fed through its own phase shifter and amplifier, so the beam points wherever the element phases add in step. Array gain over a single element is 10 log10 of the element count, so 1,024 elements give 30 dB. Half-wave spacing means one element per quarter of a wavelength squared, which is about 6,400 elements per square meter at 12 GHz and about 35,000 at 28 GHz.

Electronic steering is the whole reason arrays displaced mechanical mounts. A beam repoints in microseconds instead of seconds, there is nothing to wear out, one aperture can form several beams at once, and losing a few percent of the elements costs a fraction of a decibel rather than the whole antenna. That combination is what makes tracking a low-orbit satellite crossing the sky in minutes practical, and what lets a radar interleave search and track. The weaknesses are scan loss, grating lobes, and cost. Gain falls with the projected aperture as the beam steers off boresight, so about 3 dB is lost at 60° and the beam widens by roughly a factor of two, which means the array has to be sized for its worst scan angle rather than its best. If the elements are spaced further apart than about half a wavelength the array radiates a second unwanted beam, so element count cannot be traded away. And because both cost and DC power scale with element count, a large array at high frequency is expensive to build and hot to run, with the electronics drawing power continuously whether traffic is flowing or not.

If the thing you are pointing at does not move and you need only one beam, use a reflector. A 1 m dish is stamped metal and a feed, and no array will beat it on dollars per decibel. Choose an electronically steered array when the target moves, when you need several beams from one aperture, when repointing has to happen faster than a motor can manage, or when a mechanism is unacceptable on a vehicle or an aircraft. Below about 6 GHz the elements are large and cheap and arrays are routine on base stations; above about 20 GHz element count per unit area is four times higher for each doubling of frequency, so packaging and beamformer silicon dominate the bill and the aperture you can afford shrinks. If you need a very large aperture and can accept slow steering, a reflector on a mount, or a reflector with a small phased array as its feed, gets most of the benefit for far less money. For the terminal as a product, with its power, thermal and subsidy economics, see Flat-panel user terminal on the space-launch-and-satellites sheet; the question here is the antenna itself.

Gain = 4π × effective aperture ÷ wavelength², and beamwidth in degrees is roughly 70 × wavelength ÷ diameter · a 1 m dish at 12 GHz gives roughly 40 dBi and a 1.75° beam, and 8 dB more gain at 30 GHz (both derived here) · array gain = 10 log10(element count), so 30 dB at 1,024 elements · half-wave spacing gives about 6,400 elements per m² at 12 GHz and about 35,000 at 28 GHz · scan loss roughly 3 dB at 60° off boresight, with the beam about twice as wide · commercial silicon beamformer ICs serve 4 to 16 elements each · Starlink terminal build cost fell from about $3,000 to under $1,500 in roughly six months.

The Starlink and Amazon Leo consumer terminals, which are Ku-band arrays of thousands of elements; commercial Ku- and Ka-band beamformer chipsets from Anokiwave and Analog Devices, which put 4 to 16 transmit and receive channels on one silicon die; military active electronically scanned array radars such as the AN/APG-81; ThinKom's mechanically scanned variable-inclination arrays, which remain the aviation workhorse because they avoid scan loss; research arrays such as the 256-element Ku- and Ka-band SATCOM panels from UCSD.

An array's cost is per element, and element count scales with aperture area times the square of frequency, while a reflector's cost scales with area alone. That is what splits the market: reflectors dominate wherever a big aperture is needed and steering is slow, arrays dominate wherever steering matters, and the crossover moves toward reflectors as frequency rises. The historical cost driver was the transmit and receive module, which in military radars was a discrete gallium arsenide assembly priced in the hundreds to thousands of dollars per element, which is why arrays of that era appeared only where nothing else would do. Silicon beamformers changed the slope by integrating 4 to 16 channels per die in a commercial CMOS or SiGe process, so cost per channel now follows semiconductor volume rather than microwave module assembly. SpaceX cutting terminal build cost from about $3,000 to under $1,500 within roughly six months of starting deliveries is the clearest public demonstration of that curve. The margin sits with the beamformer chip vendors and with whoever can yield a multilayer board carrying thousands of controlled-impedance feeds, not with the antenna designer, so the practical read on where array prices go next is the chip roadmap rather than the antenna literature.

20 minPhased Arrays - Steering and the Antenna Pattern | An Animated Intro to Phased ArraysMarshall Bruner · 100k+ views

14 minWhat are Antenna Gain, EIRP, and Friis Equation?Iain Explains Signals, Systems, and Digital Comms · 10k+ views

Web-based Course - Adaptive Antennas and Phased Arrays (MIT Lincoln Laboratory) · Development of Flat Panel Active Phased Array Antennas Using 5G Silicon RFICs at Ku- and Ka-Bands (IEEE Access)

MIMO uses several antennas at each end of a link to do one of two different things with the same hardware. Spatial multiplexing sends independent data streams on the same frequency at the same time, and the number of streams is limited by the smaller of the transmit and receive antenna counts and by how much the environment scatters, so capacity multiplies rather than adding. Beamforming puts one stream through all the antennas with the phases aligned toward a user, which buys array gain of 10 log10 of the antenna count, 18 dB at 64 antennas, plus the ability to steer nulls at interferers. Massive MIMO does both at once with 32 to 64 antennas at the base station, and its real trick is multi-user operation: instead of trying to give one handset sixteen streams, the base station serves eight or more separate users on the same time and frequency resource by giving each one a beam the others cannot hear. 5G NR defines the channel state information codebooks for up to 32 antenna ports and up to eight layers per user. All of it depends on knowing the channel, because the precoding weights are computed from a channel estimate and are wrong the moment that estimate goes stale.

The strength is that spectral efficiency goes up without buying spectrum or raising transmit power, and the beamforming half also improves cell-edge coverage, which is often the benefit an operator notices first. The weakness starts with channel rank: multiplexing needs rich scattering, so a clean line-of-sight rural link supports one stream no matter how many antennas are bolted to the tower. The marketing number is the port count, and a 64-port radio delivers nothing like 64 times the capacity, because the real limit is how many users the scheduler can find with separable channels in the same slot; sector throughput in loaded urban cells is usually reported at two to four times a four-antenna baseline. Channel state information is the second constraint, and it is why massive MIMO is a mid-band TDD technology: in time-division duplex the base station measures the uplink and reuses it for the downlink by reciprocity, while in frequency-division duplex the handset has to feed back a codebook index whose size grows with antenna count until it eats the gain. The estimate also ages, since at 3.5 GHz and 100 km/h the Doppler shift is about 324 Hz and the channel stays coherent for roughly 1.3 ms, which is a slot or two. Handsets carry two or four receive antennas, so per-user downlink is capped at four layers however many the base station has.

Use massive MIMO where the traffic is dense and the environment scatters, which means urban and suburban mid-band TDD, and where the alternative is buying spectrum or building sites. Do not buy it for a rural macro cell with few users and a clear view, where the channel has rank one and a smaller radio with the same beamforming gain does the same job for less money and less weight. If you are in FDD low band, size the benefit on beamforming alone and treat multiplexing as a bonus, because the feedback overhead scales badly. If your users move fast, shorten the sounding period or fall back to wider beams, since a precoder computed from a 1.3 ms-old channel estimate points at where the user was. Before committing, check the tower: a 64-antenna mid-band radio unit is several times heavier than a 4-antenna one, and wind loading has forced more structural upgrades and lease renegotiations than most capacity plans budget for.

Spatial multiplexing gain = the smaller of transmit antennas, receive antennas and channel rank · beamforming array gain = 10 log10(antenna count), 18 dB at 64 antennas · 5G NR defines codebooks for up to 32 antenna ports and up to 8 layers per user · handsets carry 2 or 4 receive antennas, capping per-user downlink at 4 layers · at 3.5 GHz and 100 km/h, Doppler is about 324 Hz and channel coherence time about 1.3 ms (derived here) · sector throughput in loaded urban cells is usually reported at 2 to 4 times a 4-antenna baseline, not the 16 the port count suggests · 100 MHz of mid-band bought at $0.68 per MHz-POP costs $68 per person covered.

64T64R active antenna units on 3.5 GHz TDD from Ericsson, Nokia, Samsung and Huawei, which are the volume massive MIMO deployment worldwide; multi-user MIMO in Wi-Fi 5 onward, which brought the same idea to access points with four to eight antennas; 4x4 MIMO in LTE-Advanced, the first widely deployed spatial multiplexing; the LuMaMi testbed at Lund University and Rice University's Argos, the research systems that produced most of the early measured results.

Massive MIMO is bought as an alternative to spectrum and sites, and the comparison is worth doing in dollars. At the $0.68 per MHz-POP that 100 MHz of 3.45 GHz cleared for, spectrum costs about $68 per person covered, so doubling the capacity of one sector serving 2,000 people by buying another 100 MHz would cost roughly $136,000 of spectrum, while the radio unit that does at least as much costs tens of thousands. That gap is why every dense network upgraded the radio before buying more spectrum. What the arithmetic leaves out is the site. The heavier unit can trigger a structural analysis, a lease amendment, and a crew with a crane, and those costs land on the operator whether the capacity gets used or not. Power is the recurring item: a mid-band massive MIMO radio draws on the order of a kilowatt at full load, several times a conventional unit, and radio access equipment is a large share of a mobile operator's energy bill, which is why vendors now ship features that switch antenna elements and whole carriers off at night. The margin here sits with the handful of radio vendors who can build these units at scale, since the beamforming happens in vendor-specific silicon and firmware rather than in anything a third party supplies.

11 minWhat are Spatial Diversity and Spatial Multiplexing in MIMO?Iain Explains Signals, Systems, and Digital Comms · 10k+ views

13 minWhat is Beamforming (Massive MIMO)? Find Out With MpiricalMpirical · 50k+ views

Massive MIMO: Ten Myths and One Critical Question (arXiv / IEEE Communications Magazine) · ETSI TS 138 214: NR; Physical layer procedures for data (ETSI / 3GPP)

A 5G radio access network is the set of base stations, called gNBs, that connect handsets to a core network over the 3GPP New Radio air interface first specified in Release 15 in 2018. NR works in two frequency ranges, FR1 from 410 MHz to 7.125 GHz and FR2 from 24.25 to 52.6 GHz, with a maximum carrier bandwidth of 100 MHz in FR1 and 400 MHz in FR2, and carriers are aggregated to go beyond that. Its main departure from LTE is scalable numerology: subcarrier spacing is 15 kHz multiplied by a power of two, so 15, 30, 60 or 120 kHz, and because a slot is always 14 symbols, the slot shrinks from 1 ms at 15 kHz spacing to 0.125 ms at 120 kHz. Wider spacing therefore buys shorter transmission opportunities, which is where most of the latency improvement comes from, and it is also what makes millimeter wave workable, since a wide subcarrier tolerates more phase noise. The coding changed too, from turbo codes to LDPC for data and polar codes for control. The base station itself is usually split into a radio unit at the antenna, a distributed unit nearby that runs the time-critical scheduling, and a centralized unit that can sit tens of kilometers away, which is what lets an operator pool baseband across many sites.

The real gains over LTE come from three things at once: more spectrum in a single carrier, massive MIMO in mid-band TDD, and a scheduler that can address a user in a fraction of a slot. The catch is that the headline numbers belong to the standard, not to a deployed network. ITU's IMT-2020 requirements ask for 20 Gbps peak downlink, 100 Mbps user-experienced downlink, and 1 ms user-plane latency for ultra-reliable service, while commercial networks typically deliver 100 to 300 Mbps median download and 15 to 40 ms round-trip latency, because peak rate assumes one user with all the spectrum and the latency figure describes the air interface alone. It also helps to keep per-sector and per-user numbers apart: 100 MHz of mid-band with 64T64R radios produces 400 Mbps to 1 Gbps per sector, and that is shared across everyone in it. The second weakness is uplink. Mid-band is time-division duplex with a downlink-heavy frame pattern, so roughly a fifth to a quarter of the airtime is uplink, and uplink throughput lands well under a tenth of downlink at the cell edge. The third is energy: a mid-band 5G macro site with 64-antenna radios typically draws two to three times what an LTE-only site drew, and the radio access network is where most of an operator's network electricity goes.

This is not really a choice for a mobile operator, since 5G is where the device ecosystem and the new spectrum are. The decision that matters is standalone versus non-standalone. If you are adding capacity to an existing LTE network and want coverage on day one, non-standalone is the cheaper path, because the 5G carrier rides as a secondary cell on an LTE anchor and reuses the LTE core, and most of the world's commercial 5G still works this way. If you need network slicing, guaranteed low latency, voice over NR, or a private network with its own core, you need standalone, and you need to budget for a 5G core, a new signaling architecture, and device support that is still uneven. If your problem is indoor coverage or rural reach, do not buy standalone to fix it; buy low-band spectrum or more sites, because the physics is unchanged. And if you are sizing an application around 1 ms latency, measure the deployed network first, since the number in the specification is the air interface and yours will include transport, core and the internet.

FR1 410 MHz–7.125 GHz with a 100 MHz maximum carrier, FR2 24.25–52.6 GHz with 400 MHz · subcarrier spacing 15, 30, 60 or 120 kHz, so a 14-symbol slot runs 1 ms down to 0.125 ms · IMT-2020 asks for 20 Gbps peak downlink, 100 Mbps user-experienced downlink and 1 ms user-plane latency for ultra-reliable service · deployed networks typically deliver 100–300 Mbps median download and 15–40 ms round trip · 400 Mbps–1 Gbps per sector on 100 MHz of mid-band with 64T64R · roughly a fifth to a quarter of TDD airtime is uplink · a mid-band 5G macro site draws 2–3x an LTE-only site.

3GPP Release 15 (the first NR specification, 2018), Release 16 (2020, adding ultra-reliable low-latency and sidelink), Release 17 (2022, extending FR2 to 71 GHz) and Release 18, marketed as 5G-Advanced; T-Mobile US on 2.5 GHz, which launched a standalone core in 2020 and is the usual reference for mid-band coverage; Verizon and AT&T on C-band cleared in FCC Auction 107; China Mobile and Reliance Jio, both of which built standalone from the start; Ericsson, Nokia, Samsung, Huawei and ZTE, which between them supply nearly all the radios.

The radio access network is where most of a mobile operator's capital goes, typically 60 to 70% of network capital spending, and access networks as a whole absorb 15 to 20% of revenue indefinitely. Within a site the radios are the smaller part: roughly a quarter of what a new macro site costs to build, with land, steel, power, permits, backhaul and the installation crew taking the rest. Upgrading an existing site is much cheaper than building one, which is why nearly every 5G rollout started as a swap of radios on towers that already existed. What makes the arithmetic hard is that the revenue did not move. Average revenue per user is roughly flat in most markets that deployed 5G, so the capital was spent to hold share and to carry traffic that grows 20 to 30% a year at a falling price per gigabyte. Energy is the recurring line that got worse rather than better, since a mid-band massive MIMO site draws two to three times an LTE site, which is why vendors now sell sleep modes and carrier shutdown as features. Vendor margin is concentrated: after the restrictions on Huawei, the practical choices outside China are Ericsson, Nokia and Samsung, and the FCC's program to rip out and replace Chinese equipment was funded at $1.9 billion, priced by carriers near $5 billion, and topped up by Congress in 2023.

8 min5G SA vs 5G NSA: Difference between standalone & non standalone 5GCommsbrief Limited Official · 10k+ views

3 minPart 6: Standalone and Non-Standalone 5G - 5G for Absolute Beginners3G4G · 10k+ views

Minimum requirements related to technical performance for IMT-2020 radio interface(s) (International Telecommunication Union) · ETSI TS 138 300: 5G; NR; NR and NG-RAN Overall description (ETSI / 3GPP)

Open RAN is an attempt to turn a base station from one vendor's product into a set of parts that different vendors can supply. A conventional gNB is a single supplier's radio, baseband and software, connected by proprietary interfaces. The O-RAN Alliance, formed in 2018, publishes specifications that cut the same box at defined seams: a 7-2x split between the radio unit and the distributed unit over an open fronthaul interface, the 3GPP F1 interface between the distributed unit and the centralized unit, and new interfaces called E2, A1 and O1 that let a RAN intelligent controller and a management system reach into scheduling and configuration. The distributed and centralized units become software running on commodity servers rather than purpose-built baseband cards. The intended result is that an operator buys radios from one company, baseband software from another, servers from a third, and adds optimization applications from a fourth. Nothing in this changes the air interface, so an Open RAN network and a conventional one look identical to a handset.

What has been proven is that the interfaces work and that a network can be built on them: Rakuten Mobile in Japan, EchoStar's Dish network in the United States and 1&1 in Germany all launched on disaggregated architectures, and EchoStar reached the FCC's 70% population coverage milestone in June 2023. What has not been proven is the saving. Rakuten claimed roughly 40% lower capital cost and 30% lower operating cost than a conventional build, and its mobile segment still posted an operating loss near ¥500 billion in 2022 with cumulative losses since launch past ¥1 trillion, so the claim has never shown up anywhere a buyer can check it. The recurring problem in brownfield networks is that integration moves from the vendor to the operator. With one supplier, that supplier owns end-to-end performance and fault isolation; with four, somebody has to own the fronthaul timing budget, the interoperability testing and the finger-pointing, and that somebody is usually the operator or a paid integrator. Feature velocity has also been slower outside greenfield networks, because a new 3GPP feature has to land in several vendors' software before it works. Energy is a real weakness too: a distributed unit running on general-purpose servers draws more power than purpose-built baseband silicon, which is why Intel, Nvidia and Marvell all sell accelerators aimed at closing that gap.

If you are building a network from nothing, with no legacy to interwork with and no installed vendor to negotiate against, disaggregation is worth taking seriously, because that is the case where every deployment so far has actually shipped. If you are swapping radios on an existing national network, assume the bill will be integration-heavy and be explicit about who owns end-to-end performance before signing anything. If your goal is vendor diversity rather than a lower price, say so, and check whether the contract actually delivers it: AT&T's flagship open RAN agreement with Ericsson, worth up to $14 billion over five years, went to a single incumbent supplier. If you are an enterprise or a private network buyer, the disaggregation argument matters much less than it does for a carrier, since your site count is small and the integration overhead is the same. And if a vendor quotes a total cost of ownership saving, ask which parts are hardware, which are labor, and whether the comparison is against a greenfield build or against the network you already run.

O-RAN Alliance formed 2018; the 7-2x fronthaul split plus the E2, A1 and O1 interfaces are the defining seams · Rakuten claimed roughly 40% lower capital cost and 30% lower operating cost · Rakuten's mobile segment operating loss peaked near ¥500 billion in 2022, with cumulative losses past ¥1 trillion · EchoStar reached the FCC's 70% population coverage milestone in June 2023 · AT&T's open RAN contract with Ericsson is worth up to $14 billion over five years to one vendor · Vodafone has committed to open RAN at 30% of its European sites by 2030, starting with about 2,500 UK sites · NTIA's Public Wireless Supply Chain Innovation Fund was appropriated $1.5 billion by the CHIPS and Science Act.

Rakuten Mobile, which launched in April 2020 as the first national fully virtualized network and now sells its stack through Rakuten Symphony; EchoStar's Dish 5G network, built cloud-native on AWS, which hit its FCC coverage milestone in 2023 and was largely wound down in 2025 when EchoStar sold spectrum and moved subscribers onto a host network; 1&1 in Germany, which launched on Rakuten Symphony in 2023 and later added national roaming; Vodafone's UK deployment across Wales and South West England, about 2,500 sites using Samsung software on Dell servers, part of a commitment to open RAN at 30% of its European sites by 2030; Deutsche Telekom's O-RAN Town in Neubrandenburg; and the AT&T and Ericsson agreement of December 2023, the largest open RAN contract signed to date.

The global RAN equipment market is roughly $35–45 billion a year, and the case for Open RAN is that opening the interfaces lets new suppliers take a share of it and drives prices down for everyone. So far Open RAN sits somewhere near a tenth of RAN revenue, and most of that is single-vendor open RAN, meaning an incumbent's own equipment that happens to support the interfaces. That distinction is where the money went: the incumbents captured the demand for openness without giving up the account. The reason is that hardware is the low-margin part of a base station. Radios are increasingly a manufacturing business, while the margin sits in scheduler software, the integrated system and now the RAN intelligent controller, all of which the incumbents still own. Public money has been used to change this, with $1.5 billion appropriated to NTIA's Public Wireless Supply Chain Innovation Fund and a UK target of 35% of national mobile traffic over open RAN by 2030, and both are better read as industrial policy about supply chain concentration than as a cost argument. For a buyer, the honest summary is that disaggregation has demonstrably worked in greenfield networks with cloud-native operations teams, and has so far cost brownfield operators more in integration and slower features than it has returned in equipment discounts.

6 minVirtualized RAN, Cloud RAN, and Open RAN: Making Sense of the 5G RAN Alphabet SoupWoodrowWilsonCenter · 50k+ views

4 minWhat the @#$% is OpenRAN? And is it really open?Michael MacDonald · 50k+ views

Understanding O-RAN: Architecture, Interfaces, Algorithms, Security, and Research Challenges (arXiv / IEEE Communications Surveys & Tutorials) · Open Radio Access Network Security Considerations (NSA and CISA)

A private cellular network is an LTE or 5G network with its own core, its own SIM credentials and its own spectrum, serving one organization at one set of sites. 3GPP calls it a non-public network, and there are two shapes: a standalone network with its own identity and core, and a network that shares an operator's radios or spectrum but keeps its own slice and subscriber database. The spectrum comes from a regulator that has set aside a local-licensing route rather than from an auction. In the United States that route is CBRS, 150 MHz at 3550–3700 MHz shared in three tiers: naval radar and satellite earth stations are protected incumbents, 70 MHz is sold as ten-year county-level Priority Access Licenses, and the remaining 80 MHz plus any unused licensed channels are free General Authorized Access. Nobody transmits until a Spectrum Access System, a cloud database, grants a specific channel and power level, and coastal sensors tell the database to clear channels when a Navy radar appears. Germany, the UK and Japan run similar local-license schemes in nearby bands, generally at fees measured in hundreds or low thousands of currency units per year rather than the billions an exclusive national license costs.

The strengths over Wi-Fi are scheduling and mobility rather than raw speed. Airtime is assigned by the base station instead of contended for, so latency under heavy load has a bounded tail, and handover is decided by the network in tens of milliseconds where Wi-Fi roaming is decided by the client and can stall a session for 100 ms or more, which is what breaks automated guided vehicles and remote-controlled cranes. Coverage per radio is also much larger, so an outdoor yard, a port or a mine is a handful of radios instead of dozens of access points. The weakness is on the device side, and it is decisive in ordinary buildings. Every laptop, phone and tablet already has Wi-Fi, while a private cellular endpoint needs a module costing $50–200 or a gateway box, plus a SIM to issue and manage. The second weakness is operations: Wi-Fi is an IT function with a deep labor pool, while a private network needs core operations, subscriber management, SAS coordination and RF planning, which is why most enterprises buy it as a managed service instead of running it.

Compare the two on cost per square meter of coverage and cost per connected device, because those two numbers point in opposite directions. If the space is large, open and sparsely populated with devices, such as a port, a mine, a rail yard, a utility substation or a 50,000 m² warehouse with a few hundred vehicles, private cellular is usually cheaper, since one radio at $2,000–5,000 covering roughly 5,000–15,000 m² replaces several access points at $500–1,500 each covering roughly 1,000–2,000 m². If the space is an office full of laptops and phones, use Wi-Fi 6E, because the per-device cost of cellular dominates everything else and Wi-Fi's contention problem is not what limits an office. Pick private cellular whenever a mobile machine must not drop its session while moving, or when the application needs a latency ceiling rather than a latency average. If you only need a few hundred meters of outdoor coverage and have no radio staff, price a managed service before buying equipment, and check whether your integrator or your own team owns the core. In the United States, start on free GAA spectrum and buy a Priority Access License only if measurements show that contention is actually costing you throughput.

CBRS is 150 MHz at 3550–3700 MHz in three tiers: protected incumbents, 70 MHz of ten-year county-level Priority Access Licenses, and 80 MHz of free General Authorized Access · PALs cleared about $0.21 per MHz-POP · a CBRS small cell runs $2,000–5,000 and covers roughly 5,000–15,000 m² of open indoor space · an enterprise Wi-Fi 6E access point runs $500–1,500 and covers roughly 1,000–2,000 m² · a cellular module adds $50–200 per endpoint against zero for Wi-Fi · a private core is $50,000–200,000 bought outright, or a monthly subscription · network-controlled handover takes tens of milliseconds against 100 ms or more for client-decided Wi-Fi roaming.

FCC Part 96 CBRS with Spectrum Access Systems run by Google, Federated Wireless and others; Germany's 3.7–3.8 GHz local licenses from the Bundesnetzagentur, of which several hundred have been assigned to industrial sites; Ofcom Shared Access licenses in the UK and local 5G in Japan; the Port of Rotterdam and the Port of Hamburg; Mercedes-Benz Factory 56 in Sindelfingen and BMW's plants; mining networks at Rio Tinto and Boliden; and Nokia and Ericsson, which supply most of the equipment, alongside Celona, Betacom and the US carriers selling it as a managed service.

The fixed cost is the core and the integration, and the marginal cost is a radio plus a module per device, which is the opposite shape to a public network where spectrum dominates. Local licenses are cheap by design: CBRS Priority Access cleared at about $0.21 per MHz-POP against $0.68–0.94 for exclusive mid-band, and General Authorized Access costs nothing, so spectrum stops being the barrier to entry that it is for a carrier. A core bought outright runs roughly $50,000–200,000, or a monthly subscription if somebody else runs it, and for a large industrial site the first deployment is usually integration-heavy, with the equipment often the smaller half of the invoice once RF survey, cabling, device provisioning and application work are counted. The recurring money is in the managed service rather than the hardware, which is why carriers and system integrators have chased this market harder than the equipment vendors have. Scale is the honest caveat: trade bodies count on the order of one to two thousand organizations worldwide with a deployed private network, against millions of enterprise Wi-Fi installations, and most of the deployments are in a handful of verticals where a dropped session costs real money. If a business case rests on replacing Wi-Fi generally rather than on a specific machine that cannot tolerate a roaming gap, it usually does not survive the device-cost arithmetic.

3 minWhat is Private 5G? Private 5G vs. Wi-FiEye on Tech · 10k+ views

2 minWhat is CBRS?OnGo Alliance · 10k+ views

3.5 GHz Band Overview (Federal Communications Commission) · 47 CFR Part 96: Citizens Broadband Radio Service (Electronic Code of Federal Regulations)

Fixed wireless access sells home broadband over a mobile network. A customer gets a gateway, a small indoor or window-mounted box holding a 5G modem and a directional antenna, registers it against a street address, and the operator serves it from the same mid-band sector that serves phones. Nothing about the radio link is special: it is ordinary 5G NR, usually on 2.5 GHz or C-band with 64-antenna radios, and the gain comes from the gateway sitting still in a known direction, which lets the base station hold a narrow beam and a higher modulation than it could to a phone in a pocket. The large operators advertise plans mostly in the 100–400 Mbps range at $50–70 a month, the FCC's weighted mean maximum advertised download speed for residential fixed wireless was about 112 Mbps at the end of 2023, and the weighted mean price was $51.28 a month against $93.91 for cable. One-way latency runs 10–30 ms, so a round trip is typically 25–50 ms, which is worse than fiber and far better than a geostationary satellite.

The strength is that almost the entire cost was already paid for something else. The spectrum was bought at auction, the sites were leased and built, and the radios were installed to carry phones, so a new subscriber costs a $200–400 gateway that the customer plugs in themselves. That is why US residential terrestrial fixed wireless connections went from 1.96 million at the end of 2020 to 6.80 million at the end of 2023, roughly 247% growth, while cable's residential connections peaked at 74.6 million in 2021 and fell to 73.4 million by 2023. That took fixed wireless to about 6% of all US fixed residential connections, and 84% of the population could buy it by the end of 2023, 43% of them at 100 Mbps or better. The weakness is on the same page of the same report: the FCC notes plainly that fixed wireless providers do not have enough capacity to connect all their potential subscribers at once. A fixed wireless household uses on the order of 400–700 GB a month against 20–30 GB for a smartphone, so one line consumes what fifteen to thirty phones do, and operators sell it sector by sector only where a sector has headroom. Mobile traffic keeps growing on its own, so that headroom shrinks whether or not any new fixed wireless customers are added, and the fixed wireless traffic is usually deprioritized behind mobile, which means peak-hour speeds fall for those users first.

As a consumer or a small business, take it when cable or fiber is expensive or absent at your address, your household is not streaming several 4K feeds at once, and you can tolerate speeds that vary by hour. Avoid it if you need a consistent upload rate for video calls all day, because the time-division duplex frame gives uplink only about a fifth of the airtime. As an operator, sell it where sectors have spare capacity and stop when they do not, and treat it as a way to monetize idle inventory rather than as a broadband product with its own capacity plan. If your coverage plan needs fixed wireless to carry a large share of households in a dense market, price the extra spectrum, sectors and fiber offload that would take, because that is the same construction bill everyone else pays. For rural subsidy programs, fixed wireless is now competing directly with fiber and satellite, since the $42.45 billion BEAD program turned technology-neutral in 2025 and moved money away from fiber toward whatever covers a location cheapest.

US residential terrestrial fixed wireless connections 1.96 million in 2020 to 6.80 million in 2023, about 247% growth · cable residential connections peaked at 74.6 million in 2021 and fell to 73.4 million in 2023 · fixed wireless was about 6% of US fixed residential connections at the end of 2023 · plans mostly 100–400 Mbps at $50–70 a month, weighted mean maximum advertised download 112 Mbps, weighted mean price $51.28 a month against $93.91 for cable · gateway $200–400, self-installed, against $800–1,500 per home passed for fiber · a fixed wireless household uses 400–700 GB a month, fifteen to thirty times a smartphone · 84% of the US population had fixed wireless available at the end of 2023, and 43% at 100 Mbps or better.

T-Mobile 5G Home Internet and Verizon 5G Home, which between them reported roughly 11 million subscribers at the end of 2024 and have publicly targeted 20 million or so by 2028; AT&T Internet Air; the older unlicensed-band WISP industry serving rural America, which is what "terrestrial fixed wireless" mostly meant before 2021; Reliance Jio AirFiber in India, which added millions of connections within two years of launch; and the FCC Urban Rate Survey and Form 477 subscription data, which are the two public datasets the market share numbers come from.

This is the best unit economics in US broadband and the reason is that the fixed cost is somebody else's. Fiber costs $800–1,500 per home passed plus $150–250 of optics per subscriber and returns its capital over roughly 6–10 years at a 40% take rate. Fixed wireless costs a gateway and a self-install, so the payback on an incremental subscriber is a few months at $50–70 a month, and every dollar after that lands on spectrum and sites that were already going to be paid for. The catch is that this is inventory, not capacity. The operator is selling airtime that phones were not using, and each fixed wireless line displaces roughly twenty phone lines' worth of it, so once a sector fills the choice is to stop selling, buy more spectrum, split the sector, or push traffic onto fiber. That is why the same operators now forecast a ceiling in the low tens of millions of subscribers rather than a cable replacement, and why several of them have started buying or building fiber alongside it. Cable's response has been to sell mobile service back over an MVNO, so the two industries are now each reselling the other's spare capacity. For an investor the question to ask is not what fixed wireless costs today but how many quarters of sector headroom the operator has left in its densest markets.

6 minIs 5G Home Internet BETTER Than Fiber?Techquickie · 500k+ views

14 min5G Course - Fixed Wireless Access (FWA) Cases Advantages and Disadvantages5G Understanding · 5k+ views

2024 Communications Marketplace Report (Federal Communications Commission) · U.S. - The Rise of 5G FWA & The Battle for Fixed Broadband Customers (Ookla)

Unlicensed spectrum is spectrum anyone may transmit in provided the equipment obeys power and behavior rules, and Wi-Fi is what most of it carries. In the United States that means about 83.5 MHz at 2.4 GHz, roughly 500 MHz usable across the 5 GHz U-NII bands, and 1,200 MHz at 5.925–7.125 GHz that the FCC opened to unlicensed use in April 2020. The 6 GHz rules show how sharing is enforced without licenses: low-power indoor devices are capped at 5 dBm per megahertz and 30 dBm total effective radiated power with no external antenna, standard-power outdoor devices may reach 36 dBm but only after an Automated Frequency Coordination database checks them against the incumbent fixed microwave links in that location, and very-low-power devices are held to 14 dBm. Access to the air is by listen-before-talk with random backoff, so every device in range shares the same airtime and nobody has a reservation. The IEEE 802.11 standards set the rest: Wi-Fi 6 added OFDMA, 1024-QAM and 160 MHz channels for a 9.6 Gbps physical-layer maximum, and Wi-Fi 7 added 320 MHz channels, 4096-QAM and multi-link operation for roughly 46 Gbps.

The spectrum is free, the silicon ships in billions of units a year, and the result is the lowest cost per bit of anything in radio: a consumer router is $80–300 and an enterprise 6 GHz access point is $500–1,500 plus $300–800 for the cable drop and the install. Range is short, 10–50 m indoors and shortest at 6 GHz, which suits a building and rules out anything else. The weakness is that listen-before-talk gives no guarantee. Throughput per device falls as devices are added because they divide the same airtime, and latency under load has a long tail with no ceiling, so a link that measures 800 Mbps in an empty office measures something else at 9 am. Real rates also sit well below the headline: media access overhead takes 30–40% of the physical-layer rate, so a Wi-Fi 6 client in the same room typically gets 300–900 Mbps and a Wi-Fi 7 client on 320 MHz gets 2–4 Gbps. And the band is shared with everything else that uses it, from Bluetooth and Zigbee to microwave ovens at 2.45 GHz, so capacity depends on who else shows up and that changes without notice.

Use Wi-Fi as the default for anything inside a building you control, because the spectrum costs nothing, the client radio is already in every device, and the labor pool to run it is large. Move to 6 GHz when the 5 GHz band is congested and the clients support it, since 1,200 MHz of clean spectrum is the single biggest capacity upgrade available and it does not reach outside the building, which is a feature. If your application needs a latency ceiling rather than a latency average, or a machine must not lose its session while moving across a large site, use private cellular instead and accept the per-device cost. If you are covering an outdoor yard or a campus, do not try to solve it with more access points; the link budget and the contention both get worse with distance. And when you size a deployment, plan around the number of simultaneously active clients per access point rather than the coverage map, because contention rather than signal strength is what usually makes a network feel slow.

About 83.5 MHz at 2.4 GHz, roughly 500 MHz usable at 5 GHz, and 1,200 MHz at 5.925–7.125 GHz opened in April 2020 · 6 GHz power limits: 5 dBm/MHz with a 30 dBm cap indoors, up to 36 dBm outdoors under Automated Frequency Coordination, 14 dBm for very-low-power devices · Wi-Fi 6 physical-layer maximum 9.6 Gbps, Wi-Fi 7 about 46 Gbps on 320 MHz with 4096-QAM · media access overhead takes 30–40% of the physical-layer rate, so 300–900 Mbps per Wi-Fi 6 client and 2–4 Gbps per Wi-Fi 7 client in the same room · range 10–50 m indoors · consumer router $80–300, enterprise 6 GHz access point $500–1,500 plus $300–800 to install · spectrum cost zero.

FCC 20-51, the April 2020 Report and Order that opened 6 GHz, and the AFC systems now run by Broadcom, Federated Wireless, Google, Qualcomm and others; IEEE 802.11ax and 802.11be, certified by the Wi-Fi Alliance as Wi-Fi 6/6E and Wi-Fi 7, with 802.11bn aimed at reliability rather than peak rate; Europe's decision to open only the lower 500 MHz at 5945–6425 MHz, which splits the device market; and the unlicensed cellular variants LAA and NR-U, which put 3GPP scheduling into the same bands.

Wi-Fi is the case where the fixed cost of spectrum is exactly zero, and almost everything about the industry follows from that. Nobody bid for the band, so nobody has to earn a return on it, and the equipment competes on volume rather than on scarcity. Chipsets come from a handful of suppliers, Qualcomm, Broadcom and MediaTek among them, at thin margins on enormous unit counts, and the money in enterprise Wi-Fi has moved from the access point to the controller, the cloud management subscription and the analytics on top. For a building owner the installed cost is dominated by cabling and labor rather than by the radios, which is why access point count, not access point price, is the number that decides a budget. The strategic value is offload: the majority of smartphone data never touches a cellular network, and every gigabyte that lands on unlicensed spectrum is a gigabyte an operator did not have to buy spectrum and sites to carry. What the free band cannot buy is certainty, so anyone whose revenue depends on a latency guarantee ends up paying for licensed or shared spectrum instead, and that boundary is where buyers choose private cellular.

8 minWiFi 6 ExplainedPowerCert Animated Videos · 1m+ views

9 minWiFi 7 ExplainedPowerCert Animated Videos · 100k+ views

Unlicensed Use of the 6 GHz Band: Report and Order and Further Notice of Proposed Rulemaking, FCC 20-51 (Federal Communications Commission) · IEEE 802.11-2020: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications (IEEE Standards Association)

A microwave backhaul link is two dishes pointed at each other across clear line of sight, carrying Ethernet between a cell site and a fiber point of presence or between two buildings. Classic microwave uses licensed bands from 6 to 42 GHz in channels of 7 to 112 MHz, which delivers 100 Mbps to 1 Gbps per path in ordinary use and up to about 2.5 Gbps with 4096-QAM and cross-polarization interference cancellation. E-band, at 71–76 and 81–86 GHz, has 10 GHz of spectrum available in channels of 250 MHz to 2 GHz, which puts 1–10 Gbps on an ordinary link and up to 25 Gbps with channel aggregation and dual-polarization MIMO. Antennas run 0.3 to 1.8 m; a 0.3 m dish at 80 GHz has roughly 45 dBi of gain and a beam under one degree wide, so alignment and tower sway become design problems rather than installation details. Latency is under 100 microseconds per hop, because radio crosses air at essentially the speed of light instead of the two-thirds of it that light manages in glass.

The strength is schedule and cost against fiber: a hop costs $10,000–50,000 for both ends and goes in within 1–6 weeks including the license, against roughly $30,000 per mile for aerial fiber and past $200,000 per mile buried in a city, on a permitting schedule measured in quarters. Latency also beats fiber over the same geography, which is why trading firms pay for microwave chains. The weakness is rain, and it is what sets every link length. Rain attenuation rises steeply with frequency: at a 25 mm/h rain rate it costs roughly 4.6 dB/km at 28 GHz and about 10 dB/km at 80 GHz, so the fade margin needed depends on the rain rate exceeded at the availability target you promise, and those targets are strict: 99.9% availability allows 8.8 hours of outage a year, 99.99% allows 53 minutes, and 99.999% allows about 5. That gives a clear ladder. A 6–11 GHz link runs 30–80 km and is limited by multipath and terrain rather than rain; 18–23 GHz runs 5–15 km; E-band runs 3–5 km at 99.99% availability in a temperate climate and 1.5–3 km at 99.999%, and under 1 km in a tropical rain zone. Adaptive modulation softens this by dropping from 4096-QAM to QPSK during a fade, which keeps the link alive at roughly a sixth of the throughput, so the availability figure for the full rate and for any rate at all are two different numbers and vendors quote whichever is flattering.

Use microwave when fiber does not already pass the site and the traffic requirement is a gigabit or two for the length of the site lease. Use E-band when there is 1–5 km of clear line of sight to a fiber point of presence and the site needs multiple gigabits now rather than after a construction cycle; it is the cheapest multi-gigabit hop available and the registration is light-touch. Use 6–11 GHz when the span is tens of kilometers and a few hundred megabits is enough, which is still how most cell sites outside North America are fed. Do not plan an E-band link in a tropical rain zone at more than about a kilometer, and do not accept a link budget without seeing which rain rate and which availability target it was computed against. If the site will carry 10 Gbps or more for a decade and the right of way is obtainable, build fiber, because the radio's capacity ceiling arrives well before the lease does. Whatever the band, check the backhaul against what the radios on the tower can actually produce: a 1 Gbps hop behind a sector capable of 2 Gbps wastes the spectrum the operator paid for.

6–42 GHz on 7–112 MHz channels gives 100 Mbps–1 Gbps, or about 2.5 Gbps at best; E-band at 71–76 and 81–86 GHz gives 1–10 Gbps typical and up to 25 Gbps · rain at 25 mm/h costs roughly 4.6 dB/km at 28 GHz and about 10 dB/km at 80 GHz · 99.9% availability is 8.8 hours of outage a year, 99.99% is 53 minutes, 99.999% is about 5 minutes · link lengths 30–80 km at 6–11 GHz, 5–15 km at 18–23 GHz, 3–5 km at E-band for 99.99% in a temperate climate and 1.5–3 km for 99.999% · adaptive modulation holds the link through a fade at roughly a sixth of the throughput · latency under 100 microseconds per hop · $10,000–50,000 per hop, installed in 1–6 weeks.

Ericsson MINI-LINK, Nokia Wavence, Ceragon, Aviat, SIAE and Huawei, which supply nearly all of the market; FCC Part 101 fixed microwave licensing, which requires prior frequency coordination on each path, against the light-touch registration used for 70/80 GHz; ITU-R P.530, the recommendation every link budget in the industry is computed from; and the Chicago-to-New Jersey trading routes, where microwave chains carry a round trip in roughly 8 milliseconds against about 13 milliseconds on the fastest fiber.

Microwave is what an operator buys when the money is fine but the schedule is not. Almost all of the cost is up front and one-time: radios and antennas are usually about half the invoice, with the structural survey, the path survey, the frequency coordination fee and the crew making up the rest, and the only meaningful recurring items are the license and the rooftop or tower rent. That shape is why it still carries a large share of the world's cell sites, roughly 40–50% globally and well over half in much of Asia and Africa, while sitting under 10% in the United States, Japan and Korea where fiber is already dense. The comparison that decides most cases is against a leased circuit rather than against building fiber: a $30,000 E-band hop against a leased Ethernet circuit at $500–1,500 a month pays for itself in roughly two to five years, which is the arithmetic behind most tower-company backhaul decisions and is derived here rather than published. E-band prices fell when the radios moved from gallium arsenide onto silicon processes, and that is the main reason multi-gigabit hops stopped being a specialty product. The margin sits with a handful of vendors who can build a 1-degree-beamwidth antenna and a 4096-QAM modem cheaply, and increasingly with the tower companies who own the rooftops both ends have to sit on.

4 minBackhaul 101- Wireless Backhaul ExplainedCeragon Networks · 50k+ views

4 minBackhaul BasicsStreakwave · 100k+ views

Recommendation ITU-R P.530: Propagation data and prediction methods required for the design of terrestrial line-of-sight systems (International Telecommunication Union) · ETSI EN 302 217-2: Fixed Radio Systems; Characteristics and requirements for point-to-point equipment and antennas (ETSI)

A tactical radio network has to work with no towers, no fiber and no fixed plan, so every radio is both an endpoint and a router. That is what a mobile ad hoc network is: nodes discover their neighbors, build routes across each other, and repair those routes as everything moves. Routing splits into two families. Proactive link-state protocols such as OLSRv2 flood topology continuously so a route is ready the instant it is needed, using multipoint relays so that only a chosen subset of neighbors rebroadcasts each update. Reactive protocols such as AODV find a route only when traffic appears, which saves overhead and costs delay on the first packet. Military waveforms mostly use proprietary variants of these ideas, with spread spectrum and frequency hopping layered underneath for jam resistance. The hardware ranges from legacy VHF sets like SINCGARS at 30–88 MHz carrying 16 kbps, up to modern MIMO mesh radios running 5–20 MHz channels in the UHF, L and S bands that deliver tens of megabits on a single hop. Alongside the mesh sits a different design, the pre-scheduled data link, of which Link 16 is the one everything else is compared to: a 12-second frame divided into 1,536 time slots, with each participant given slots in advance, hopping across 51 frequencies between 969 and 1206 MHz at roughly 77,000 hops per second.

The strength is that the network exists wherever the radios are, survives losing any node, and needs nobody to have surveyed the site first. The weakness is what relaying costs. A single-radio node is half duplex, so it must receive a packet and then retransmit it, which roughly halves usable throughput per hop, and interference makes it worse: a transmitting node also silences neighbors two hops away, so end-to-end throughput in a linear chain falls to about a third of the single-hop rate by two hops and settles near a quarter beyond about four. Across a whole network the scaling is worse again, with per-node capacity falling roughly as one over the square root of the node count, so a 50 Mbps radio can look like a few megabits per user once fifty of them are sharing the same air. The second weakness is that a mesh which beacons constantly is easy to find with direction-finding equipment, so low probability of detection and network liveness pull against each other. Size, weight and power is the third: a manpack transmitting 5–20 W is limited by battery, not by protocol.

Use a MANET when the nodes move, the infrastructure is absent or expected to be destroyed, and the traffic is video, sensor feeds or file transfer that can tolerate variable rate. Size it by hop count rather than by radio specification, and assume a quarter of the single-hop rate once any path runs past three or four hops; if that is not enough, add a directional backhaul link or a dual-radio node on a separate channel rather than more mesh nodes. Use a slot-based network like Link 16 when many platforms must share one agreed picture with bounded, predictable latency and the message is small: it carries about 26.9 kbps in its standard format and up to roughly 107 kbps packed, over about 300 nautical miles of line of sight, but access is scheduled rather than contended, so it does not degrade when the network gets busy or when routing has to reconverge. Keep both. Nearly every unit that runs a wideband mesh also carries a narrowband link and a satellite or HF path behind it, because the failure modes are different and the cheapest insurance against a jammed band is a link on another one. If your application needs a guaranteed message within a known time, do not build it on a contention-based mesh.

SINCGARS at 30–88 MHz carries 16 kbps; modern MIMO mesh radios on 5–20 MHz channels deliver tens of megabits on one hop · each relay hop roughly halves throughput, and a linear chain settles near a quarter of the single-hop rate beyond about four hops · per-node capacity across a network falls roughly as one over the square root of the node count · Link 16 uses a 12-second frame of 1,536 time slots, carrying about 26.9 kbps in the standard format and up to roughly 107 kbps packed, over about 300 nautical miles line of sight · it hops across 51 frequencies between 969 and 1206 MHz at roughly 77,000 hops per second · a manpack transmits 5–20 W · commercial mesh nodes run $5,000–20,000 against $30,000–80,000 for a certified military manpack.

OLSRv2 (RFC 7181) and AODV, the two reference points every proprietary mesh protocol is described against; TrellisWare's TSM, Persistent Systems' Wave Relay, Silvus StreamCaster and Rajant InstaMesh, which are the mesh waveforms actually fielded; the US Army's Integrated Tactical Network and its two-channel manpack and leader radios from L3Harris and Collins Aerospace; Link 16 and its MIDS terminals, carried by thousands of aircraft, ships and ground stations across more than 30 nations, with TTNT and Link 22 as the higher-rate follow-ons.

Two very different cost structures sit inside this category. Commercial mesh radios are a volume business now driven by drones and robotics rather than by infantry, and a Silvus- or Doodle-class node runs $5,000–20,000. A military radio with the same silicon inside costs several times that, $30,000–80,000 for a certified two-channel manpack, and the difference is almost entirely certification, environmental qualification and low volume rather than performance. Link 16 sits in a third bracket, with terminal installations historically running into the hundreds of thousands of dollars per platform, and that is exactly why it persists: the cost of replacing it is the cost of touching every aircraft, ship and command post in the alliance at once, which no capacity argument has been able to justify. What a buyer is actually paying for at the top of the range is the shared message catalog, since a track picture that thirty nations interpret identically is worth more than bandwidth. The commercial pull-through runs the other way and is the interesting part for an investor: uncrewed systems buy mesh radios in quantities infantry never did, which is funding waveform and MIMO development that then gets certified back into military programs.

5 minHow Mesh Networks WorkThe Berkman Klein Center for Internet & Society · 100k+ views

8 minData Link Processing and ManagementNorthrop Grumman · 10k+ views

RFC 7181: The Optimized Link State Routing Protocol Version 2 (Internet Engineering Task Force) · Multifunctional Information Distribution System-Low Volume Terminal (MIDS-LVT) (Federation of American Scientists)

An orbital filing is a legal reservation of frequencies, and for a geostationary satellite a position on the arc, obtained through the International Telecommunication Union. The Radio Regulations are a treaty between governments, so the applicant is an administration rather than a company, and an operator has to find a government willing to file and to stay responsible for the network afterward. In the unplanned bands the rule is first come, first served: the administration sends advance publication information to the Radiocommunication Bureau, then a coordination request under Article 9, then a notification that records the assignment in the Master International Frequency Register, and priority runs from the date the Bureau received a complete filing. Appendix 4 lists the data that has to be supplied and Appendix 5 sets the criteria for identifying which other administrations you must coordinate with, which for a large constellation means dozens of bilateral negotiations running for years. A separate regime covers part of the spectrum by plan instead: Appendices 30 and 30A for broadcasting and 30B for the fixed-satellite service give every member country a predetermined geostationary position and a block of spectrum whether or not it operates a satellite, which is how the treaty balances efficient use against equitable access. A recorded assignment is international recognition and interference protection, not permission to sell service, because that requires a license and landing rights in each country separately.

The strength is that priority is real, enforceable, and cheap relative to the hardware it protects. Under the fee schedule in ITU Council Decision 482 that applies to filings received from 1 July 2017, an advance publication costs 570 CHF, a coordination request 20,560 to 33,467 CHF, and a notification 30,910 to 57,920 CHF. Those are the flat fees; a small filing of under 100 units is charged instead on a start fee plus units, which is why the space-launch sheet quotes a lower floor for the same schedule. Taking the flat column, the whole chain for one network runs roughly 52,000 to 92,000 CHF (adding those three flat fees, which is arithmetic done here rather than a published total). Against a constellation costing billions, that is close to free, and the result is the weakness: filings are submitted for systems nobody intends to build, and every one of them has to be coordinated around by everyone who files later. Rwanda's two filings of September 2021, covering 327,230 satellites between them, are the extreme case and cost the same as any other pair of filings. The deadlines are the counterweight and they bind. Under No. 11.44 an assignment has to be brought into use within seven years of the Bureau receiving the filing; under No. 11.44B a geostationary assignment counts as brought into use only when a satellite capable of that assignment has sat at the notified position for a continuous 90 days, with the Bureau told within 30 days of the end of it. For non-geostationary systems, Resolution 35 (Rev. WRC-23) adds milestones after that seven-year period: 10% of the notified satellite count within two more years, 50% within five, and 100% within seven, and missing one does not cancel the filing but shrinks it, capping the recorded number at ten times, then twice, then exactly the number actually deployed.

Treat the filing as a long-lead item and start it before the spacecraft design is fixed, because seven years sounds generous and coordination with a crowded band routinely eats several of them. If you are a national operator, file through your own administration and accept that you are also buying that administration's regulatory competence, since the Bureau deals with the government and not with you. If you are a startup without a cooperative home regulator, you will end up filing through someone else's administration, and you should price the loss of control: the assignment sits with that government, your claim on it is contractual, and a dispute with the host is a dispute you cannot appeal to the ITU. When you are valuing a company whose main asset is a filing, ask three questions before anything technical: what the priority date is, how much of the coordination is actually agreed rather than pending, and how many satellites the operator can fund by the next Resolution 35 milestone. And do not confuse the ITU layer with the market layer, because a recorded assignment sells nothing until each country's regulator grants access, which in the United States means an FCC Part 25 authorization with its own milestones and bond.

Seven years from receipt of the filing to bringing into use under No. 11.44 · geostationary bringing into use requires 90 continuous days at the notified position, reported within 30 days · Resolution 35 milestones of 10%, 50% and 100% at two, five and seven years past that · ITU fees of 570 CHF for advance publication, 20,560–33,467 CHF for coordination and 30,910–57,920 CHF for notification, roughly 52,000–92,000 CHF for the chain (summed here) · Rwanda's September 2021 filings covering 327,230 satellites.

Rwanda filed two constellations totaling 327,230 satellites in September 2021 on behalf of E-Space, having previously launched one three-unit cubesat, RwaSat-1, in 2019. Omnispace holds its 2 GHz mobile-satellite rights through a Papua New Guinea license, which SpaceX raised directly in a May 2026 letter asking whether a Papua New Guinea licensee considers itself bound to coordinate in good faith. Kepler filed the roughly 115,000-satellite Aether system through the German administration. Appendices 30, 30A and 30B, Resolution 35 (Rev. WRC-23), and 47 CFR Part 25 are the four documents that decide most of these questions.

Almost none of the money here is the fee. The cost of a filing is the interference analysis, the coordination staff, and the years of negotiation with every administration whose network overlaps yours, which for a large non-geostationary system is a standing team and outside consultants rather than a one-time submission. What that spending buys is a date, and a date is worth whatever the band is worth: an early priority in Ku or Ka band is the reason several satellite companies have been acquired at prices their in-orbit assets did not justify. The fees fall on the notifying administration, which normally passes them to the operator, so the government's exposure is administrative rather than financial, and that asymmetry is exactly what makes hosting other people's filings attractive to a small administration. The milestone rules changed the calculus for speculative filings without ending them, since a filing that misses its milestones is trimmed to what was built instead of being canceled, so the downside of filing big and building small is a smaller recorded system rather than nothing. For an operator the practical budget item is the gap between filing and revenue: seven years of holding costs before a single satellite has to exist, and then a deployment schedule that is now a regulatory obligation rather than a business plan. The spacecraft themselves, meaning the bus, payload and constellation design that the schedule commits you to, are on the space launch and satellites sheet.

11 minOutdated Rules, New Space Race: Modernizing LEO Spectrum PolicyCenter for Strategic & International Studies · 1k+ views

ITU Radio Regulatory Framework for Space Services (International Telecommunication Union) · 47 CFR Part 25 - Satellite Communications (Electronic Code of Federal Regulations)

A VSAT network is a hub at a teleport plus a large number of small remote terminals, all sharing one satellite's capacity in a star topology where every remote talks to the hub and never to another remote. The terminal is where the name comes from: a very small aperture antenna, typically 0.75 to 2.4 m, with a 0.75 to 1.2 m dish enough at Ka band and 2.4 m or more needed at C band. The outbound direction is one wide carrier, usually DVB-S2X, broadcast to every remote in the beam and time-shared among them, with adaptive coding and modulation switching format frame by frame so a site under rain drops to a more robust mode instead of dropping out. DVB-S2X carries constellations up to 256APSK and roll-off factors down to 0.05, which is what lets an operator sell most of a transponder's theoretical capacity rather than half of it. The return direction is the harder half and runs MF-TDMA under DVB-RCS2: each remote bursts in time and frequency slots the hub assigns it, so return capacity is scheduled rather than owned, and a remote that has nothing to send gives its slots back. Everything about the design follows from the delay, since a geostationary round trip cannot beat 477 ms and services measure 550 to 650 ms.

The strength is that coverage inside a beam is uniform and immediate. A site anywhere in the footprint gets the same service in days with no trenching, no right of way, and no cost penalty for being remote, which is the reverse of every terrestrial option and is the entire reason the category exists. The weaknesses are delay and sharing. TCP will not fill a long path on its own: with the classic 64 KB receive window and a 600 ms round trip, one connection carries 65,536 bytes every 0.6 seconds, which works out to about 0.87 Mbps no matter how much satellite capacity is available. Every VSAT network therefore runs performance-enhancing proxies that terminate TCP locally and acknowledge on the satellite's behalf, and that trick stops working when traffic is encrypted end to end, because TLS 1.3 and QUIC hide the transport the proxy needs to see. The other weakness is oversubscription: capacity is sold contended, often somewhere between 10:1 and 50:1, so the committed information rate is the only number in the contract that describes what the site will actually get in the busy hour.

Use VSAT when you have many fixed sites spread over a continent, you need them all on the same network with the same configuration, and terrestrial service is either unavailable or unreliable at enough of them to make one contract simpler than fifty. If the traffic is transactional and small, meaning point of sale, ATM authorizations, telemetry, SCADA polls, and store replication overnight, half a second of delay costs nothing and the economics work. If the traffic is interactive, meaning voice, video calls, remote desktop, or a cloud application that makes many small round trips, do not put it on a geostationary link, because 600 ms is a delay users notice on every keystroke; buy low-orbit service instead and accept a less mature enterprise feature set. Buy committed information rate for the traffic that matters and best effort for the rest, since paying for a committed rate across the whole site is the most common way these contracts get expensive. And if any site has a terrestrial option at all, use it and keep the VSAT terminal as the backup path, which at $1,000 to $5,000 of hardware per site is cheap insurance and is where most enterprise satcom spending now sits.

Antennas of roughly 0.75–2.4 m, 0.75–1.2 m at Ka band and 2.4 m or more at C band · 477 ms round-trip floor through geostationary orbit and 550–650 ms in service · DVB-S2X up to 256APSK with roll-off down to 0.05 · about 0.87 Mbps per TCP connection at a 64 KB window and 600 ms round trip, derived here from those two figures · contention typically 10:1 to 50:1 · terminals of $1,000–5,000 and capacity at $100–500 per Mbps per month.

Hughes JUPITER, ST Engineering iDirect, Gilat SkyEdge and Comtech are the hub and terminal platforms most enterprise networks run on, and Viasat and SES are the largest capacity sellers into them. The standards are ETSI EN 302 307-2 for the DVB-S2X outbound and ETSI EN 301 545-2 for the DVB-RCS2 return. Typical deployments are retail chains and bank branch networks, ATM and lottery terminals, oil and gas and pipeline SCADA, mining sites, and maritime service on Inmarsat Global Xpress and Fleet Xpress, now inside Viasat.

The fixed cost is split unusually: the customer pays for terminals and installation, and the service provider pays for the hub. A shared hub at a teleport is a seven-figure asset that only makes sense across hundreds or thousands of remotes, which is why almost nobody builds one and most enterprises buy a managed service from an integrator that already has one. The recurring cost is space segment, and its pricing model changed with high-throughput satellites: capacity used to be leased by the megahertz, a whole 36 MHz transponder at a time for a year or more, and is now sold by the megabit as managed capacity. At $100–500 per Mbps per month, a site with 5 Mbps committed runs $500–2,500 a month, which is arithmetic rather than a quoted price and is the number to compare against whatever a terrestrial circuit would cost at the same address. For the satellite operator the marginal cost of another bit is close to zero once the satellite is in orbit, so the price is set by how much paying demand sits under that particular beam, and capacity over empty ocean serves no customers at any price. That is why the spacecraft-side question of how many spot beams point where, covered on the space launch and satellites sheet, decides what this service can be sold for. The margin in the middle goes to the integrator that owns the hub, the field service, and the customer relationship, and low-orbit competition has been compressing it since 2022.

3 minWhat is VSAT?Global Invacom Group · 10k+ views

21 minVSAT BasicsYedaCenter · 50k+ views

RFC 3135 - Performance Enhancing Proxies Intended to Mitigate Link-Related Degradations (IETF) · Second Generation DVB Interactive Satellite System (DVB-RCS2); Part 2: Lower Layers for Satellite standard (ETSI)

Satellite backhaul puts the transport circuit behind a site on a satellite instead of fiber or microwave. The site itself is ordinary: a normal 4G or 5G radio, a normal core connection, and a satellite modem where the fiber handoff would be. What changes is the budget behind it, because the circuit is now metered and the capacity is shared with everyone else under the same beam. A rural cell carrying voice and light data usually needs 5 to 20 Mbps of backhaul, and a busy one 50 to 200 Mbps, so the satellite circuit and not the radio is what caps the site. Orbit decides whether the result is usable for interactive traffic: geostationary adds 550 to 650 ms to every session, medium orbit around 100 to 150 ms, and low orbit 20 to 50 ms round trip, which is why satellite backhaul went from a last resort to a mainstream option after low-orbit capacity arrived in volume. The same terminal is also sold without a cell site behind it, feeding a village Wi-Fi network, a school, or a clinic directly, and in that form it is the cheapest way to connect a place that no operator will build toward. That is a large market: the ITU counted 2.2 billion people still offline in 2025, and almost all of them live where terrestrial buildout has already been judged uneconomic.

The strength is schedule and indifference to distance. A satellite-backhauled site is live in days rather than the 6 to 24 months a fiber build takes, and the cost is the same whether the site is 5 km past the end of the fiber or 500, which is the opposite of every terrestrial option and the reason satellite keeps winning the hardest 2% of locations. The weakness is that the cost structure is inverted: almost nothing is fixed and almost everything is recurring, so a link that looks cheap to install is the most expensive megabit in the network for as long as the site exists. At $500 to $3,000 a month, ten years of service runs $60,000 to $360,000, which is more than a microwave hop at $10,000 to $50,000 costs to build and more than a fiber spur costs in many places. The other weaknesses are shared: the operator does not control who else is under the beam, so the busy-hour rate is a contract term rather than a physical property, and on a geostationary link the delay degrades voice quality and makes any chatty application feel broken. What has changed the arithmetic is price, since Euroconsult put satellite data capacity at about $260 per megabit per second per month in 2023, down 77% over five years.

Work out the ten-year cost before the install cost. If there is line of sight to a fiber point within about 30 km, a microwave or E-band hop almost always wins on a ten-year view, and if fiber can be built at all along an existing right of way it wins by more. Use satellite when the site is genuinely unreachable, meaning an island, a mountain valley, a jungle or desert route with no corridor, or a temporary site, and when restoring service after a cut. Pick the orbit by traffic: if the cell carries voice and interactive data, use low or medium orbit and accept the newer, thinner enterprise feature set; if it carries messaging, telemetry, and overnight bulk, geostationary is fine and usually cheaper per megabit. If you are serving households rather than a cell site, compare against the subsidy available, because at roughly $30,000 per location for rural fiber, a satellite terminal is the cheaper answer at the tail of almost every program, which is what drove BEAD's shift to technology neutrality in 2025.

5–20 Mbps of backhaul for a rural cell and 50–200 Mbps for a busy one · installed in days · $500–3,000 a month, so $60,000–360,000 over ten years (multiplied out here) against $10,000–50,000 to build a microwave hop · 550–650 ms round trip through geostationary orbit, 100–150 ms at medium orbit, 20–50 ms at low orbit · satellite data capacity at about $260 per Mbps per month in 2023, down 77% over five years · about $30,000 per location for rural fiber · 2.2 billion people offline in 2025.

SES sells medium-orbit backhaul over O3b and O3b mPOWER to island and Pacific operators, and Intelsat sells CellBackhaul over geostationary capacity. Gilat and ST Engineering iDirect build the small-cell-over-satellite platforms most rural sites run on, and Kacific serves the Pacific and Southeast Asia from Ka-band geostationary capacity. Starlink is now used for cell backhaul and community gateways in the same markets. On the funding side, the US universal service fund distributes about $8B a year and BEAD put $42.45B toward unserved locations before turning technology-neutral in 2025.

This entry is the clearest case on the sheet of a link with almost no fixed cost and a high marginal one, which is the inverse of fiber and the reason the two are complements rather than competitors. For the operator, the terminal and install are a few thousand dollars and the circuit is a monthly bill that never ends, so satellite backhaul is bought when the alternative is a capital project that will not be approved rather than when it is cheap. On the supply side the price has fallen fast: Euroconsult put average revenue at about $260 per megabit per second per month for data services in 2023, down 77% over five years, while the cost of supplying high-throughput capacity in North America fell from roughly $40 to about $12 per Mbps per month between 2019 and 2023. Both halves of that matter. Falling prices expanded the set of sites where satellite backhaul makes sense, and falling prices also compressed the margin of the operators selling it, which is why capacity sellers have been buying service providers and moving up into managed services. Most rural deployments are still funded by somebody other than the end customer, through a universal service levy, a national program, or a license condition, so when evaluating one of these businesses the first question is which subsidy it depends on and how long that program runs. The satellite capacity itself, meaning how many beams point where and what a spacecraft costs to build and launch, is on the space launch and satellites sheet.

23 minSatellite Cellular BackhaulYedaCenter · 5k+ views

Measuring digital development: Facts and Figures 2025 (International Telecommunication Union) · Broadband Equity Access and Deployment Program (BroadbandUSA)

Satellite IoT moves messages, not streams. A tracker on a shipping container, a well-head pressure sensor, a buoy, or a livestock collar sends tens to a few hundred bytes every few minutes to a few times a day, and the whole system is designed around that shape rather than around throughput. Two architectures deliver it. A crosslinked or geostationary system is always overhead, so the message goes out immediately: Iridium's 66 satellites in six planes of eleven at 783 km cover the poles as well as the equator, and its Short Burst Data service measures better than 0.44 kbit/s mobile-originated and 0.35 kbit/s mobile-terminated in practice, which is a modem speed from 1985 and entirely sufficient for a 200-byte position report. A store-and-forward system skips the crosslinks: a satellite passes over the device, collects the message, and drops it at a ground station on a later orbit, so hardware is much cheaper and delivery takes minutes to hours, because a low-orbit satellite is above a given point for roughly 8 to 12 minutes on 4 to 6 passes a day. Spectrum is mostly L-band around 1.6 GHz and S-band, with some systems in VHF, and 3GPP's narrowband non-terrestrial standards now let ordinary cellular chipsets do the same job, which is covered in the direct-to-cell entry.

The strength is coverage with no infrastructure at all: roughly 10% of the Earth's surface has terrestrial connectivity, and a satellite IoT device works on the other 90% with no site, no tower, and no roaming agreement. Power and cost follow from the small message: a modem transmitting a few seconds a day at 7 W peak runs for years on a battery, and modules have come down to consumer-electronics prices. The weakness is that airtime is priced like nothing else in communications. A representative Iridium Short Burst Data plan ran $19.50 a month with 17 kbytes included and $1.40 per kbyte beyond it, which is about $1,400 per megabyte, so the discipline of the whole design is sending fewer bytes rather than compressing them. The second weakness is that delivery is not guaranteed on any schedule: a store-and-forward system will deliver eventually, and eventually is not a specification an alarm can use. The third is commercial. This is a market of very low revenue per device, so it only works at very large device counts, and a service with 50,000 units on it will not cover its own ground segment.

Use satellite IoT when the asset moves outside cellular coverage and the data is small, periodic, and tolerant of delay: containers, rail cars, vessels, remote tanks and pipelines, environmental sensors, and anything on open water. If the asset spends most of its life in coverage and only occasionally leaves it, buy a hybrid module that prefers cellular and falls back to satellite, because you will pay satellite rates only for the gaps. If a message must arrive within seconds, use a crosslinked or geostationary system and pay for it; if hours are acceptable, store-and-forward cuts both hardware and airtime cost substantially. Size the byte budget before choosing anything, since at roughly $1,400 per megabyte the difference between a 20-byte report and a 200-byte one decides the business case. And if you need a live video feed, remote desktop, or anything measured in megabits from a remote site, this is the wrong category entirely and you want VSAT or low-orbit broadband.

Messages of tens to a few hundred bytes, with 200-byte position reports typical · Iridium at 66 satellites in six planes of eleven at 783 km · Short Burst Data measured above 0.44 kbit/s mobile-originated and 0.35 kbit/s mobile-terminated · a plan at $19.50 a month with 17 kbytes included and $1.40 per kbyte beyond, about $1,400 per megabyte (converted here) · 7 W peak transmit power · 8–12 minutes of visibility on 4–6 passes a day for a store-and-forward satellite · roughly 10% of the Earth's surface covered by terrestrial networks.

Iridium Short Burst Data is the reference real-time service, with modules sold to makers as well as to industry, and Orbcomm, Inmarsat and Globalstar are the other incumbents. Berg Insight counted 44 satellite IoT operators and a subscriber base above 3.9 million at the end of 2021, with Iridium at 1.3 million, Orbcomm at 1.1 million across its own and Inmarsat's networks, and Globalstar at 0.42 million, forecasting 21.2 million by 2026. The newer entrants split by architecture: Myriota, Astrocast, Kinéis, Fleet Space and Lacuna Space are store-and-forward, while Skylo and Sateliot sell standards-based narrowband service that ordinary cellular modules can use.

The fixed cost sits with the operator and the marginal cost sits with the customer, which is the reverse of most of this sheet. A narrowband constellation is cheap by satellite standards because the payload is small and the satellites can be, but the ground segment, the regulatory work, and the device ecosystem still cost the same as a big system, so the operator carries a large fixed base and needs millions of devices to cover it. On the customer side almost nothing is fixed: a module is tens to a few hundred dollars and the airtime is a monthly fee in the low tens of dollars, so the buying decision is per-device and easy to reverse, and switching costs are low. That combination keeps average revenue per device falling while device counts grow, which is exactly the pattern Berg Insight's 40.3% annual growth forecast describes. Two things change the picture from here. Standards-based narrowband service means the module is an ordinary cellular part rather than a proprietary one, which will push module prices toward mass-market silicon and take the hardware margin out. And the incumbents are exposed to that: Iridium reported $165.2M of service revenue in the third quarter of 2025 and withdrew its $1B 2030 revenue target, citing competition it expects to arrive late this decade. If you are evaluating one of these companies, the numbers that matter are devices in service, revenue per device per month, and what fraction of the fleet could be served by a standard cellular chipset instead.

5 minSparkFun Rock Block Satellite Communication ModuleSparkFun Electronics · 50k+ views

2 minIridium Satellite Communication for Arduinonootropic design · 10k+ views

Lessons Learned Using Iridium to Communicate with a CubeSat in Low Earth Orbit (Lawrence Livermore National Laboratory) · Space-Terrestrial Integrated Internet of Things: Challenges and Opportunities (arXiv)

Direct-to-cell works because of a standards change and a spectrum arrangement, not because of anything new inside the phone. 3GPP Release 17 was the first release to define non-terrestrial network operation, covering New Radio satellite access in Frequency Range 1 for handsets and small terminals plus narrowband IoT and LTE machine-type access over satellite. It started with two mobile-satellite bands, n255 at 1525–1559 and 1626.5–1660.5 MHz and n256 at 1980–2010 and 2170–2200 MHz, and the list has since grown to include n254 and, in Release 18, the Ka-band pairs n510 to n512 at 27.5–30 GHz up and 17.3–20.2 GHz down. The hard part it had to solve is that the cell is far away and moving. A terrestrial cell assumes a round trip well under a millisecond and Doppler of a few hundred hertz, while a low-orbit satellite closes at roughly 7 km/s, which shifts a 1.6 GHz carrier by about 37.5 kHz, and the far edge of a beam is hundreds of kilometers further from the satellite than the center. The answer is pre-compensation: the satellite broadcasts its ephemeris and a common timing offset, the handset uses its own GNSS fix to correct both timing and frequency before it transmits, and the specification's own FR1 test channels assume only 200 Hz of Doppler is left after that. The satellite itself is usually a transparent payload, meaning the base station stays on the ground and the spacecraft relays, so from the core network's point of view the satellite is one more cell in the operator's network. This entry covers the standard, the spectrum position, and the economics that follow from both. The spacecraft that fly these payloads, meaning AST's large unfolding arrays and SpaceX's dedicated direct-to-cell satellites, are on the space launch and satellites sheet.

The strength is distribution. Every phone already in a pocket is a potential terminal, there is no hardware to sell or subsidize, and the traffic lands in the carrier's existing core with the carrier's existing subscriber records. The weakness is capacity, because one beam covers a cell hundreds of kilometers across and shares a few megabits per second across everyone in it, which is why commercial service starts at messaging, emergency calling, and low-rate data rather than anything resembling broadband. The structural weakness is whose spectrum it is. The FCC's Supplemental Coverage from Space framework, adopted 14 March 2024, added a secondary, not co-primary, mobile-satellite allocation to a defined list of terrestrial bands, so the satellite must not interfere with primary services and cannot claim protection from them, and it may only transmit under a lease from terrestrial licensees who together hold every license on that channel across a defined geographically independent area. That makes the satellite operator a tenant of the carrier unless it buys spectrum of its own. Secondary status also has teeth in the other direction: Omnispace, whose 2 GHz mobile-satellite rights come through a Papua New Guinea ITU filing, reported its noise floor rising as Starlink direct-to-cell payloads came into service on nearby frequencies.

If you are a mobile operator, buy direct-to-cell as a coverage and safety feature and price it as one, because closing a dead zone this way costs far less than building a tower into terrain that has none. Check the band list before assuming it applies to you: the US framework covers 600 MHz, 700 MHz, 800 MHz, broadband PCS at 1850–1915 and 1930–1995 MHz, and the AWS-H block at 1915–1920 and 1995–2000 MHz, so an operator whose only holdings are mid-band has nothing to lease into it today. Do not sell it as a broadband substitute, since a beam shared across a region cannot carry what your terrestrial network carries and customers told otherwise will churn. If you are building devices rather than networks, use the narrowband non-terrestrial profile with a standard cellular chipset instead of a proprietary satellite module, because the module then costs what mass-market silicon costs. And if you are investing, the question that decides the returns is whether the satellite operator owns spectrum or leases it, since that alone determines whether it collects a wholesale fee or the subscriber's bill.

3GPP Release 17 as the first non-terrestrial release, Release 18 extending it above 24 GHz · satellite bands n255 at 1525–1559 and 1626.5–1660.5 MHz and n256 at 1980–2010 and 2170–2200 MHz, later joined by n254 and the FR2 bands n510 to n512 at 27.5–30 GHz up and 17.3–20.2 GHz down · about 37.5 kHz of Doppler at 1.6 GHz for a satellite closing at 7 km/s, and 200 Hz assumed after pre-compensation in the FR1 test channels · US supplemental coverage bands at 600, 700 and 800 MHz plus PCS 1850–1915 and 1930–1995 MHz and AWS-H 1915–1920 and 1995–2000 MHz, all secondary · framework adopted 14 March 2024 · a few megabits per second shared across a cell hundreds of kilometers wide.

Starlink Direct to Cell runs in T-Mobile's PCS spectrum in the US, and AST SpaceMobile operates under leases of AT&T, Verizon and FirstNet low-band. Skylo sells narrowband messaging over geostationary L-band to Android and iOS handsets using the standard's IoT profile. Iridium is making its 66-satellite constellation compatible with the 5G standards under the name NTN Direct, was in in-orbit testing as of late 2025, and has a roaming agreement with Deutsche Telekom. The governing documents are the FCC's Single Network Future Report and Order (FCC 24-28), 3GPP Release 17 and 18, and ETSI TS 138 108 for the satellite radio requirements.

The unusual thing about this category is that the customer-premises cost is zero. Every other satellite service on this sheet has to put a terminal at the user's end, and terminal cost is what has historically gated consumer satellite businesses; here the terminal is a phone somebody already bought. That moves the entire fixed cost into orbit and into spectrum, and the spectrum half is now the larger number in at least one case, which is the clearest statement anyone has made about where the value sits. Revenue looks like a feature rather than a service: the carrier sells it as an add-on or bundles it, and the satellite operator collects a wholesale fee per subscriber that is measured in single-digit dollars a month. Marginal cost per bit is high by satellite standards, because a beam covering a region hundreds of kilometers across cannot be reused nearly as aggressively as a spot beam pointed at a town, so capacity per paying user is thin and stays thin until satellites get much larger antennas. Two things would change the economics. If per-cell throughput rises enough to sell data rather than messages, the service competes with terrestrial coverage instead of supplementing it. And if an operator owns the spectrum rather than leasing it, it keeps the subscriber relationship instead of supplying the company that has one. Until then this is an infrastructure supply business selling into carriers, and the carriers set the price.

13 minIntroducing 5G NTN (Non Terrestrial Networks) | 5G Telecoms Training from MpiricalMpirical · 5k+ views

3 minNon-Terrestrial Networks (NTN) Basics - Part of NTN & 5G Courseitelcotech · 1k+ views

Single Network Future: Supplemental Coverage from Space, Report and Order (Federal Communications Commission) · 5G; NR; Satellite Access Node radio transmission and reception (3GPP TS 38.108 version 18.5.0 Release 18) (ETSI)

Almost all transmission fiber is single-mode: a doped silica core 8 to 9 microns across inside a 125 micron cladding, sized so that only one spatial mode propagates at 1310 and 1550 nm. Loss at 1550 nm is 0.18 to 0.20 dB/km in ordinary production fiber, and that number is what sets amplifier spacing at 80 to 100 km. Three ITU-T grades cover nearly everything in the ground. G.652 is the default, with about 17 ps/nm per km of chromatic dispersion at 1550 nm and roughly 80 square microns of effective area. G.654 pushes the cable cutoff wavelength up so the core can be pure silica with 110 to 150 square microns of effective area, which cuts loss to 0.15 to 0.17 dB/km and raises the launch power the fiber tolerates before nonlinearity distorts the signal. G.657 gives up a little effective area for bend tolerance, holding loss under control at a 7.5 mm bend radius in the A2 grade and 5 mm in B3. Multimode fiber with a 50 micron core survives inside buildings and data centers because a wide core lets a cheap VCSEL couple into it, but modal dispersion stops it after a few hundred meters. Around the glass sits the cable, which is where the variety is: loose-tube and ribbon designs from 12 to 864 fibers, rollable-ribbon micro-cables reaching 3,456, and steel armor wherever a backhoe or a rodent can get to it.

The strength that matters commercially is that the glass does not go obsolete. A G.652 fiber installed in 1995 carries today's 800G wavelengths, because every capacity upgrade since then happened in the terminal equipment rather than in the cable, and the cable is the expensive half to replace. Loss of 0.2 dB/km, immunity to electromagnetic interference, and a 25-year-plus service life follow from the material. The weaknesses are mechanical and dispersive. Standard fiber loses real power at a tight bend, which is why fiber to the home moved to G.657 for drops and risers. Splicing G.652 to G.654 joins fibers whose mode field diameters differ by roughly 2 microns, and each such splice costs about 0.1 to 0.3 dB, so operators do not mix grades along a route. Contaminated connector end faces remain the leading cause of link faults in installed plant. And the latency floor of about 5 microseconds per kilometer comes from the refractive index of silica, which is what puts one-way delay from New York to London at roughly 28 ms over a 6,000 km cable route; no cable design improves it.

Default to G.652.D for everything: metro rings, backbone, feeder, and the trunk side of an access network. If the route runs past roughly 1,000 km with long amplifier spans and you intend to run 400G or 800G wavelengths, specify G.654.E instead, because 0.03 to 0.05 dB/km of lower loss plus the larger effective area is worth 1 to 2 dB of system margin over that distance; below that, the premium buys nothing you can use. Use G.657.A2 for drop cable, in-building riser, and anything an installer will coil inside a wall box, which is now standard practice in fiber to the home. Use multimode only for links under a few hundred meters inside one building, where the cheaper transceiver is the whole argument. Consider hollow-core only where latency is worth money on its own, meaning exchange-to-exchange trading links and a few hyperscaler routes, and budget it by the meter rather than by the kilometer.

Single-mode core 8–9 microns in a 125 micron cladding · G.652 loss 0.18–0.20 dB/km at 1550 nm, about 17 ps/nm/km dispersion, roughly 80 square micron effective area · G.654 loss 0.15–0.17 dB/km with 110–150 square micron effective area · G.657 bend radius down to 7.5 mm (A2) and 5 mm (B3) · G.652-to-G.654 splice loss about 0.1–0.3 dB from the mode field mismatch · standard cables 12–864 fibers, rollable ribbon to 3,456 · latency floor about 5 microseconds per km in glass, roughly 28 ms one way New York to London, against about 3.4 microseconds per km in air-guided hollow core.

Corning SMF-28 Ultra as the reference G.652.D product and Corning TXF and OFS TeraWave as terrestrial G.654.E fibers; G.657.A2 drop cable in essentially every fiber-to-the-home build; 3,456-fiber rollable-ribbon micro-cable in hyperscale data center interconnect; Microsoft's purchase of the Southampton spinout Lumenisity in December 2022, after which it began installing hollow-core fiber in its own network.

The glass is cheap and the hole in the ground is not. Cable is usually well under a fifth of what it costs to build a route, against aerial construction near $30k per mile and buried urban construction past $200k per mile, which is why operators pull far more fibers than they need on the first install: adding strands to a cable already going into a duct costs very little, and going back later means paying the whole civil bill twice. The same arithmetic decides the grade question. G.654.E glass runs roughly two to three times the price of G.652, but on a long-haul route that difference is a rounding error next to trenching, so the choice is made on span loss budget rather than on fiber price. Bend-insensitive G.657 carries almost no premium now and pays for itself in avoided rework on drops. Hollow-core is the exception that proves the pattern: it comes off small specialty draws, is priced by the meter, and is bought only where a few milliseconds have a dollar value. Air guides light about a third faster than silica, so on a 6,000 km transatlantic route the saving works out near 10 ms one way (derived here from 5 versus 3.4 microseconds per km), and reported losses have come down into the range of ordinary single-mode fiber in the best laboratory results after a decade of losing several decibels per kilometer.

9 minFiber Cable Types Explained | Single-Mode (SMF) vs Multimode Fiber (MMF)CBT Nuggets · 10k+ views

2 minSeven Types of Optical Fiber! G.652 to G.657HYC · 1k+ views

The FOA Reference For Fiber Optics - Optical Fiber (Fiber Optic Association) · G.654: Characteristics of a cut-off shifted single-mode optical fibre and cable (ITU-T)

A coherent transponder mixes the received light with a local oscillator laser so the receiver recovers amplitude and phase in both polarizations, which gives four independent dimensions per symbol instead of the one that direct detection sees. Everything after that happens in a digital signal processor: chromatic dispersion, polarization-mode dispersion and laser phase noise are undone in the electrical domain, which is why coherent systems removed the dispersion-compensating fiber spools that used to sit in every amplifier hut. Rate comes from three knobs multiplied together, namely bits per symbol, symbols per second, and the two polarizations. Dual-polarization QPSK carries 4 bits per symbol, DP-16QAM carries 8, and DP-64QAM carries 12, so the first commercial coherent line at 100G ran DP-QPSK near 32 Gbaud, 400G runs DP-16QAM near 60 Gbaud, and OIF's 800ZR specifies DP-16QAM at a nominal 118.2 Gbaud. The current top of the range is 1.6 Tbps on one wavelength from a DSP on a leading process node. The other change is packaging: coherent used to mean a line card in a separate transport chassis, and it now also comes as a QSFP-DD or OSFP module that plugs straight into a router port.

Coherent detection is what let capacity per fiber keep rising after direct detection ran out of room, because it recovers the phase and so can use dense constellations that a photodiode cannot see. It also made the optical line simpler, since impairments the DSP can undo no longer need optical hardware to fix. The limits are now information-theoretic and thermal. Spectral efficiency for the pluggable classes is about 5.3 bits per hertz (400G in a 75 GHz slot, 800G in 150 GHz, both derived from the standardized channel plans), and line systems tuned for shorter spans reach 6 to 8, against a nonlinear Shannon limit for the fiber channel near 10 to 11, so the remaining headroom is a factor of two rather than a factor of ten. Denser constellations also need more signal-to-noise, so every step up in bits per symbol costs reach: 400ZR targets 120 km and 800ZR 80 to 120 km, while a line-side transponder with the same DSP family runs thousands of kilometers. And a pluggable coherent module dissipates roughly 20 to 30 W inside a faceplate designed for a client optic, which is what caps how many can go in one router.

If the span is a data center interconnect or a metro route under about 120 km and the traffic is Ethernet, put a 400ZR or 800ZR pluggable in the router and delete the transport shelf. If the route is regional, out to roughly 500 km, use OpenZR+ instead, which keeps the same form factor and buys reach with a stronger forward error correction code. If the route is long-haul, submarine, or needs OTN framing, performance monitoring and per-wavelength restoration, keep a line-side transponder, because that is where the extra reach and the operational tooling live. Check who owns the optical layer before you decide: a pluggable in a router means the IP team now owns the link budget, and that only works if there is an open line system underneath it and someone who can plan wavelengths. The laser, modulator and detector inside the module are covered on the optics and photonics sheet under optical transceivers; the question here is the transport system the module sits in.

DP-QPSK 4 bits per symbol, DP-16QAM 8, DP-64QAM 12 · 100G at about 32 Gbaud, 400G at about 60, 800ZR at a nominal 118.2 Gbaud DP-16QAM · per-wavelength rates of 400G, 800G and 1.6T shipping today · spectral efficiency about 5.3 bits per hertz for ZR-class pluggables and 6–8 for short-span line systems, against a nonlinear Shannon limit near 10–11 · 400ZR reach 120 km, 800ZR 80–120 km, OpenZR+ to roughly 500 km, line-side transponders thousands of km · pluggable coherent module power roughly 20–30 W.

OIF's 400ZR and 800ZR implementation agreements and the OpenZR+ MSA, whose original promoters were Acacia, Cisco, Fujitsu Optical Components, InnoLight, Juniper, Lumentum and NTT Innovative Devices; Ciena's WaveLogic, Nokia's PSE and Infinera's ICE DSP families on the line side; Cisco's purchase of Acacia, closed in 2021 at $4.5 billion, and Marvell's purchase of Inphi the same year at roughly $10 billion; Nokia's acquisition of Infinera, closed in 2025.

The pluggable shift moved money from chassis to silicon. A transponder used to arrive as a line card in a transport shelf with its own power, management and rack space, and the system vendor sold all of it; a coherent pluggable is a module in a router port that the operator already paid for, so the shelf, the client-side optics that fed it, and the floor space all come out of the bill. That is the reason hyperscalers pushed the standards, and it is why the transport equipment market stopped growing with traffic. What did not commoditize is the DSP. Designing one on a 5 or 3 nm node is expensive enough that only about half a dozen companies ship a competitive part, which is why Cisco paid $4.5 billion for Acacia and Marvell about $10 billion for Inphi rather than building one, and why Nokia bought Infinera in 2025. For the operator the unit economics are simple: on an installed fiber, going from 100G to 800G on a wavelength is roughly eight times the capacity for the price of two modules, and no civil work at all. The catch is that reach and spectral efficiency trade against each other, so a network that buys the cheapest module per bit and then finds it cannot reach the far end has bought nothing.

16 minWhat is Coherent Optical Technology?ZCorum · 10k+ views

10 min400G ZR, Coherent Transport Plug for IPoDWDM Applications in Switches and RoutersInphi · 1k+ views

About | OpenZR+ (OpenZR+ MSA) · Implementation Agreement for 800ZR Coherent Interfaces (Optical Internetworking Forum)

Dense wavelength-division multiplexing runs many transponders down one fiber pair at once, each on its own frequency from the ITU-T G.694.1 grid, and the line system is everything between them: multiplexers, amplifiers, and the switching that routes individual wavelengths. The grid is either fixed at 50 or 100 GHz or flexible in 12.5 GHz increments, which is what lets 400G on a 75 GHz slot sit next to 800G on 150 GHz. The C-band from 1530 to 1565 nm gives about 4.4 THz of usable spectrum, or roughly 6 THz on the extended plan most new systems use, and the L-band above it adds a similar amount, so a C+L line delivers roughly 10 THz in one pair. Across the C-band alone that is 96 channels at 50 GHz spacing, 64 at 75 GHz or 32 at 150 GHz, so a filled C-band carrying 400G on every 75 GHz slot comes to about 25 Tbps per fiber pair and a C+L line to roughly 50 (derived here). Erbium-doped fiber amplifiers every 80 to 100 km boost every channel at once rather than one at a time, which is the whole reason the economics work, and Raman pumping is added where a span is too long for that. At branch sites a reconfigurable optical add-drop multiplexer built from wavelength-selective switches drops some wavelengths to local transponders and passes the rest through, with 8 to 20 degrees at a large node.

Lighting another wavelength on an installed fiber costs a transponder and no construction, which makes it the cheapest capacity in telecommunications by a wide margin, and one amplifier serving 96 channels is why cost per channel falls as the line fills. The weaknesses come from the line being a shared analog medium that every channel rides at once. Total launch power is capped by fiber nonlinearity, so adding channels means lowering the power in each one, and a wavelength added without re-checking the power and dispersion budget can degrade every other channel on the fiber. Amplifier gain shifts when channels are added or removed, so the control system has to hold per-channel power through transients. Cascaded ROADMs narrow the passband a little each time, which limits how many nodes a wavelength can cross before the signal will not close. And the L-band needs its own parallel amplifier chain, so turning it on means new hardware in every hut along the route rather than a software change.

Use DWDM on any route where fiber is scarce or the civil work dominates, which is nearly every metro and long-haul route. If you own plenty of dark fiber over a short campus span, compare it against simply lighting parallel fibers, because at that distance pulling more glass is cheap and parallelism skips the wavelength planning. Use a fixed 50 GHz grid if every channel is the same rate, and flexible grid once you are mixing 400G, 800G and 1.6T, which occupy different widths. Fill the C-band before you add the L-band, since C+L roughly doubles spectrum and roughly doubles the line-system cost at every amplifier site. Choose an open line system, with third-party transponders riding it through a G.698.2-style single-channel interface, only if you have optical engineers who can own the power and OSNR budget; if you do not, a vendor-integrated system is cheaper once the integration labor is counted.

Fixed grid 50 or 100 GHz, flexible grid in 12.5 GHz increments · C-band about 4.4 THz, roughly 6 THz extended, and about 10 THz across C+L · 96 channels at 50 GHz, 64 at 75 GHz, 32 at 150 GHz · roughly 25 Tbps per fiber pair with the C-band filled at 400G per 75 GHz slot, and 50 Tbps across C+L (derived here) · amplifier spacing 80–100 km · large ROADM nodes of 8–20 degrees.

ITU-T G.694.1 for the frequency grid and G.698.2 for the single-channel interfaces that let a third-party wavelength ride another operator's line; wavelength-selective switches from Lumentum and Coherent inside nearly every ROADM shipped; the Telecom Infra Project's Open Optical and Packet Transport group and its GNPy route-planning library, which exists so an operator can compute its own optical budget instead of asking the system vendor; C+L line systems on transatlantic and long-haul terrestrial routes.

The line system is the fixed cost of a lit route and the transponders are the variable cost, which is the split that decides how an operator spends. Amplifiers, ROADMs and the huts they live in are bought once for the whole route, and after that capacity arrives one module at a time with no civil work, so cost per lit wavelength falls steeply as fill rises and an operator running 8 of 96 channels is carrying nearly all of the fixed cost while collecting little of the benefit. The recurring cost is the sites: an amplifier hut every 80 to 100 km needs power, space, access and a lease, and that bill does not change with traffic. Adding the L-band is the one upgrade that behaves like a rebuild, because it needs a parallel amplifier chain at every one of those huts, so the rule of thumb is to fill C-band first and treat C+L as a decision about avoiding a new fiber build rather than about spectrum. Open line systems changed who captures the margin: the system vendor used to sell the line and the transponders together, and separating them moved the transponder purchase to a competitive module market while moving the optical engineering onto the operator's own staff. That trade is worth it at hyperscaler scale and usually is not at a regional carrier's.

5 minEDFA Amplifier principle in DWDM |Optical fiber| RAMAN Amplifier ROADM | OTN #roadm #otn #dwdmOpticsTrans · 10k+ views

19 minROADM DWDM WSS in optical network, Directional v/s Direction-less #telecom #optical #physicsOpticsTrans · 50k+ views

Optical Fiber Communications (RP Photonics Encyclopedia) · G.698.2: Amplified multichannel dense wavelength division multiplexing applications with single channel optical interfaces (ITU-T)

A passive optical network runs one fiber from a line terminal in the exchange to a splitter in the field, and from there to 32 or 64 homes, with nothing powered anywhere in between. Downstream traffic is broadcast to every terminal on the tree and each one keeps the frames addressed to it; upstream, the line terminal grants each subscriber terminal a time slot so their bursts arrive without colliding. GPON, in ITU-T G.984, carries 2.488 Gbps down and 1.244 Gbps up. XGS-PON, in G.9807.1, carries 9.953 Gbps in both directions on different wavelengths, 1577 nm down and 1270 nm up against GPON's 1490 and 1310, so the two run on the same fiber at the same time and an operator can migrate subscriber by subscriber. Above that sit the 25GS-PON multi-source agreement and ITU-T's 50G-PON. The link budget is what limits the tree: a 1:32 splitter takes about 17 dB and a 1:64 splitter about 20 dB out of a 28 dB Class B+ budget, which leaves roughly 20 km of reach.

The passive part is the point. Between the exchange and the home there is no power, no battery, no cabinet, no cooling and nothing to fail, so the outside plant has a 25-year-plus life and survives three generations of electronics without being touched. Sharing one line-terminal port and one laser across 32 homes is what brings the electronics cost per subscriber down to a fraction of the plant cost. The weaknesses follow from sharing. Capacity is shared, so a GPON tree with 32 active subscribers has about 78 Mbps each if they all pull at once, and the operator manages that with oversubscription assumptions rather than with dedicated bandwidth. The splitter is also a fault-isolation problem: a break in the feeder takes out every home behind it, and locating a fault through a splitter is harder than on a dedicated fiber. And a shared tree cannot be physically unbundled, which matters in regulatory regimes that require handing a competitor a dedicated fiber.

If you are building residential or small-business access at scale, use XGS-PON rather than GPON now, because the outside plant is identical and the electronics are the cheap half of the build. Use a 1:32 split as the default, go to 1:64 in dense areas where the loss budget still closes and take rate is low, and drop to 1:16 where you expect heavy sustained use. If subscribers need guaranteed symmetric multi-gigabit service, or the regulator requires physical unbundling, use point-to-point Ethernet instead and accept more fiber in the feeder and one line-terminal port per subscriber. Treat 50G-PON as a later card swap rather than a reason to delay, since the same fiber and splitters carry it. Whatever you choose, design the splitter locations and fiber counts for the highest split you will ever want, because moving a splitter later means going back into the ground.

GPON 2.488 Gbps down and 1.244 Gbps up; XGS-PON 9.953 Gbps symmetric · split ratios 1:32 typical, 1:64 common, 1:16 where utilization is high · 1:32 splitter about 17 dB and 1:64 about 20 dB of a 28 dB Class B+ budget, giving roughly 20 km reach · GPON at 1490/1310 nm and XGS-PON at 1577/1270 nm, so both run on one fiber · about 78 Mbps per subscriber if 32 GPON users pull at once (derived here) · outside plant life 25 years or more, across three generations of electronics.

ITU-T G.984 for GPON, G.9807.1 for XGS-PON and the G.9804 series for 50G-PON; IEEE 10G-EPON, which dominates in Japan and among US cable operators; the 25GS-PON multi-source agreement, created because 25G arrived faster in the market than in the ITU; XGS-PON overlays across AT&T, Frontier and most large European incumbents; the US BEAD program, which put $42.45 billion toward unserved locations and made access fiber a subsidized asset class.

The build cost is the trench and the electronics are an afterthought. A US fiber build runs roughly $800 to $1,500 per home passed, nearly all of it labor, permitting and restoration, against $150 to $250 of optics per subscriber, and connecting a subscriber later adds a drop, a terminal and an install visit for a few hundred dollars more. Splitting 32 ways is what produces that ratio, because one line-terminal port and one laser are amortized across 32 homes rather than bought per home. Passive plant then keeps the ratio from decaying: with nothing powered in the field there is no cabinet lease, no battery replacement cycle and no truck roll when the utility drops power, and the same fiber and splitters carry GPON, XGS-PON and 50G-PON in turn, so each capacity upgrade is a card and a customer-premises device rather than a construction project. Returns depend almost entirely on take rate. At 40% take and $60 to $70 a month against $800 to $1,500 per home passed, capital comes back over roughly 6 to 10 years, which is why fiber access is financed like infrastructure and why overbuilding a market someone else already passed rarely works: the second operator pays the same civil cost for half the subscribers.

7 minIntroduction to Passive Optical Networking (PON)Tellabs Optical LAN Training · 10k+ views

4 minGPON vs. EPON vs. XGS-PON - ProLabs Podcast ClipProLabs · 10k+ views

The FOA Reference For Fiber Optics - Fiber To The Home PON Types (Fiber Optic Association) · G.9807.1: 10-Gigabit-capable symmetric passive optical network (XGS-PON) (ITU-T)

A submarine cable is a small bundle of fiber pairs inside a steel and polyethylene structure wrapped around a copper conductor, and the conductor is there to carry DC power out to the amplifiers rather than to carry signal. In deep water the cable is about the diameter of a garden hose; near shore it gets steel armor and is plowed into the seabed, which is why nothing is visible at the beach. Erbium-doped repeaters sit every 50 to 100 km along the route and are fed a constant current from power equipment at both landing stations, so a single fault that breaks the conductor takes the whole span down rather than one channel. Older systems carry 4 to 8 fiber pairs; modern space-division designs carry 12 to 24, because the binding constraint on a long system is the electrical power that can be pushed out to the repeaters, so more pairs run at lower power each rather than fewer pairs run hard. Capacity per system runs from tens of terabits on older cables to a few hundred on new ones, with MAREA rated at 224 Tbps across 6,600 km. There are more than 600 active and planned systems and over 1.5 million km of cable in service, and the design life of a system is 25 years.

Nothing else moves intercontinental traffic at this price. Submarine cable carries about 99% of it, while the FCC's own filings put satellite at 0.37% of US international capacity, and the reason is capacity per dollar: a few hundred million dollars buys 200 to 400 Tbps on a transatlantic route, while the largest geostationary satellite ever built is designed for about 1 Tbps at a comparable price. Once the cable is wet the marginal cost of another terabit is terminal equipment and electricity. The weaknesses are that the asset is fixed and the wet plant cannot be changed. Fiber pairs are set at manufacture, so all later capacity growth comes from upgrading the terminal equipment, and coherent upgrades have raised installed cable capacity several times over across a system's life. Faults are routine rather than rare, at roughly 200 a year worldwide with about two-thirds caused by fishing gear and ship anchors, and a repair takes weeks because it needs a ship. Every route is also a single line on a map, so resilience comes from buying capacity on several systems rather than from anything inside one of them.

If you need intercontinental capacity measured in terabits, this is the only option, and the real decision is whether to build or to buy. Buy capacity on an existing system if you need it inside two years, if your volume is under a fiber pair, or if you are unwilling to hold a construction risk; building takes 2 to 3 years before the first bit moves. Build or join a consortium only when you have anchor traffic that fills at least a pair and a view of demand over a decade, which in practice means a hyperscaler or a group of carriers. For an island or a small market, take a branch off a passing trunk through a branching unit rather than commissioning a dedicated system, since the trunk's ship time is already paid for. Whatever you buy, buy diversity explicitly: two circuits on cables that share a landing station or a beach approach are one circuit in practice. Landing stations, repair economics and how capacity is priced are covered in the next entry.

Repeater spacing 50–100 km, powered by a constant-current DC feed from both ends · 4–8 fiber pairs on older systems and 12–24 on space-division designs · MAREA rated at 224 Tbps over 6,600 km · more than 600 active and planned systems and over 1.5 million km in service · design life 25 years · roughly 200 faults a year worldwide, about two-thirds from fishing gear and anchors · satellite carries 0.37% of US international capacity against about 99% on cable.

MAREA, the 6,600 km Microsoft, Meta and Telxius transatlantic system rated at 224 Tbps; Google's Dunant, the first space-division system with 12 pairs; 2Africa, a roughly 45,000 km system circling the continent; the four suppliers who build essentially all of it, SubCom, Alcatel Submarine Networks (which Nokia agreed in 2024 to sell to the French state), NEC and HMN Technologies; ITU-T G.977, which specifies what an optically amplified submarine system has to do.

Almost the entire cost is committed before the first bit moves, and the largest single item is ship time rather than glass. A transatlantic system runs a few hundred million dollars over 2 to 3 years covering marine survey, permits, the cable itself, the repeaters and the vessel days to lay it, which works out near $40,000 per km on a 6,600 km route (derived here from those two figures). After that the marginal cost of another terabit is a pair of transponders and the electricity to run them, so the price a buyer pays has far more to do with how much capacity the owner needs to sell than with what the bit costs to carry. Who buys has changed. New systems are increasingly funded by Google, Meta, Microsoft and Amazon rather than by carrier consortia, and the difference is structural: a consortium has to presell capacity to raise the money and therefore builds only what it can contract in advance, while a hyperscaler builds against its own traffic forecast and sells none of it. Because the wet plant is fixed for 25 years and only the terminal equipment improves, cost per delivered bit on a given cable falls steadily over its life without anyone spending capital on the ocean floor, which is also why old systems get retired for being uncompetitive rather than for wearing out.

6 minHow The Internet Travels Across OceansVISION · 5m+ views

11 minINSIDE the Largest Cable-Laying Ship Ever Built: How Undersea Cables Are LaidNauctis · 50k+ views

Frequently Asked Questions (International Cable Protection Committee) · G.977: Characteristics of optically amplified optical fibre submarine cable systems (ITU-T)

A cable landing station is a hardened building a few hundred meters to a few kilometers inland that holds three things: the power feed equipment driving up to roughly 15 kV of DC down the cable's copper conductor to the repeaters, the terminal equipment that lights each fiber pair, and the backhaul circuits to the nearest carrier-neutral facility. Between the station and the water sit a beach manhole, ducts run under the sand, and a sea earth electrode that closes the power circuit. Getting one built means a cable landing license from the national regulator, which in the United States comes from the FCC with an executive-branch national security review, plus coastal, seabed and environmental consents, and that process commonly runs 12 to 24 months and sets the schedule for the whole project. Landings concentrate heavily, so Bude and Porthcurno in Cornwall, Wall Township and Tuckerton in New Jersey, Marseille, Fortaleza and Singapore each host many systems. When a cable does break, the fault is located by measuring from the terminal, a ship mobilizes, grapples the cable, recovers both ends, splices in a new section and re-lays it, which takes roughly one to three weeks in the Atlantic and a month or more on remote routes where a repair permit has to be issued first.

The maintenance model is the part that works well. Owners join a zone agreement that keeps ships on standby across a region, so the cost of a fleet is spread over many systems and a several-hundred-million-dollar asset can be covered against faults for a low single-digit percentage of its capital cost per year. Concentrated landing points help too, since a beach with six systems already on it has ducts, permits, power and metro backhaul in place. The same concentration is the main weakness. Two circuits bought on different cables that share a landing station, a beach approach or a power feed are one circuit in practice, and a station is a building with a door. The fleet is the other constraint: the ICPC's own illustrated list of major cableships holds 63 vessels and says plainly that it is not complete, only a subset of those are on repair standby, a handful cover the whole Atlantic, most are over 20 years old, and nobody orders one speculatively. Against roughly 200 faults a year, a bad month means repairs queue.

If you are buying international capacity, ask which landing station and which beach each circuit uses rather than which cable, because that is where diversity is usually lost. If you are planning a new system, start permitting before you order cable, since consents run 12 to 24 months and the ship booking is the cheap part to move. Budget the backhaul from the landing station into the metro separately, because on short routes it often costs more than the subsea segment. Join a zone maintenance agreement rather than self-insuring, since a spot charter has to be arranged in the same week everyone else wants the same ship. If restoration time matters to your customers, buy pre-agreed restoration capacity on a physically separate system instead of assuming the network will reroute, and confirm that the alternative does not land on the same beach.

Power feed up to roughly 15 kV DC from each end · permitting and consents commonly 12–24 months · Atlantic repair typically 1–3 weeks, a month or more on remote routes · the ICPC lists 63 major cableships and calls the list incomplete · most of that fleet is over 20 years old · roughly 200 faults a year worldwide · annual maintenance cover typically a low single-digit percentage of system capital cost.

Bude and Porthcurno in Cornwall, Wall Township and Tuckerton in New Jersey, Marseille, Fortaleza and Singapore as the landing points where systems cluster; regional standby arrangements such as the Atlantic and Mediterranean cable maintenance agreements; ship operators including Global Marine, Orange Marine, SubCom, Alcatel Submarine Networks, IT International Telecom and E-Marine; carrier-neutral landing station operators who sell space to several systems at one beach rather than each owner building its own.

On a new build the money goes to marine survey, permits, the cable and repeaters, and vessel days, and the landing station itself is a small share of the total while sitting on the critical path. Once the system is in service the recurring bill has two parts: a standing charge under a zone maintenance agreement, usually a low single-digit percentage of capital cost per year, and day rates when a ship actually sails, which puts a single deepwater repair into the millions once mobilization, transit and vessel time are counted. The fleet explains why that price does not fall. A new cableship is a nine-figure order against charter revenue that arrives in unpredictable bursts, so the 63 vessels the ICPC lists keep aging and repair capacity does not grow with the number of cables. On the revenue side, capacity sells either as an indefeasible right of use paid up front over 15 to 25 years or as a monthly lease, and unit prices have fallen at a double-digit annual rate for years, so an owner does better selling early in a system's life than holding capacity back. The buyer mix has shifted with it: content providers now deploy more international capacity than internet backbone operators do, which means a growing share of new cable is built by companies that sell none of it and treat the whole system as a cost of running their own network.

9 minInside the Beach House Connecting the World’s Internetamaysim · 10k+ views

3 minSubCom - Subsea Fiber Optic Cable Recovery, Splice, & Release AnimationSubComChannel · 10k+ views

Submarine Cable FAQs (TeleGeography) · Cableships of the World (International Cable Protection Committee)

A terrestrial free-space optical link is a pair of small telescopes on rooftops or masts, each aiming a modulated infrared beam at the other through open air. Commercial units work at 850 nm or 1,550 nm and deliver 1–10 Gbps over spans from about 100 m to a few kilometers. Beam divergence is deliberately wide, usually 1–5 milliradians, which spreads the spot to 1–5 m at 1 km so a building swaying in the wind does not walk the beam off the receiver; higher-end products replace that margin with an active tracking gimbal and a narrower beam. Optical frequencies sit above the range the ITU allocates, so there is no license to win and no coordination to sit through, and a link goes in over days rather than months. Air also carries light faster than glass does, about 3.3 microseconds per kilometer against roughly 5, which is why latency-sensitive trading firms have bought open-air hops on routes where the fiber takes a longer path. Everything hard about the technology is the atmosphere in between.

The strengths are real: gigabit capacity with no spectrum fee and no trench, installation in days, a beam only a few meters across that an eavesdropper has to physically stand in, and no radio interference to coordinate. The weakness is fog: attenuation in dense fog runs 100–300 dB/km and higher, and it barely depends on wavelength because fog droplets are close to the size of the light, so moving from 850 to 1,550 nm buys almost nothing. A typical link holds 20–40 dB of margin, which means a 200 m span rides out fog that takes a 1 km span completely off the air. Rain is far easier at 10–25 dB/km even in a downpour, since raindrops are much larger than the wavelength and there are far fewer of them in a cubic meter. Turbulence adds scintillation, intensity swings of several decibels on millisecond timescales that are worst over hot roofs around midday, and the usual fix is several transmit apertures plus a receive aperture large enough to average the speckle out. The failure mode to plan for is not gradual degradation: the link runs at full rate or it is completely down, and the annual outage total is decided by the local fog statistics rather than by the equipment.

Use free-space optical when the span is short, the schedule is weeks rather than quarters, and either no spectrum is available or the right of way for fiber is not obtainable. Typical good fits are a campus crossing over a road you cannot trench, a temporary link for an event or a disaster restoration, and a hop where being hard to intercept is worth paying for. If the span is over about 1 km and you have promised 99.9% availability, do not use it alone, because that target allows 8.8 hours of outage a year and one bad fog season will spend all of it. In a coastal or valley climate with regular fog, treat FSO as the backup rather than the primary. The usual alternative is an E-band radio at 70/80 GHz, which carries 1–25 Gbps over 1–5 km for $10k–50k a hop with light-touch registration and keeps working in fog, and it wins most comparisons on those grounds. If you need both high capacity and high availability, run the laser and a millimeter wave radio on the same path, since fog takes out the laser, heavy rain takes out the radio, and the two rarely arrive together.

1–10 Gbps per link · spans of 100 m to a few km · 850 or 1,550 nm, beam divergence 1–5 mrad · fog 100–300+ dB/km against 10–25 dB/km for heavy rain · link margin 20–40 dB · $30k–100k per link pair installed · 3.3 microseconds per km of path

Taara, spun out of Alphabet in 2025 from the optical terminals built for Project Loon, advertises 20 Gbps at up to 20 km and reported carrying about 700 TB across the roughly 5 km Congo River crossing between Brazzaville and Kinshasa in 20 days at 99.9% availability; Canon's Canobeam, LightPointe, fSONA and Mostcom have sold enterprise-scale links for two decades. The 2000–2002 wave is the cautionary half of the record: Terabeam and AirFiber raised and spent hundreds of millions of dollars on metro optical access and neither survived the telecom bust, while Anova Technologies later ran laser and millimeter wave links between the Chicago and New Jersey exchanges for latency-sensitive trading, where the shorter air path is the whole product.

Almost all of the cost is the terminal pair, roughly $30k–100k installed, plus rooftop rights at both ends. There is no spectrum fee, no permit for a trench, and effectively no marginal cost per bit once the link is up. The comparison that decides most deals is against E-band radio rather than against fiber: $10k–50k a hop, more capacity, and no sensitivity to fog, which is most of why terrestrial optical has stayed a niche. Recurring cost is small in equipment terms, a window cleaning and an occasional realignment, but an operator who needs high availability has to keep a backup circuit running for the hours the laser is down, and that standby circuit can cost more per year than the link it protects. Terrestrial FSO has been described as five years from mass adoption for about thirty years, and the reason has not changed: the hardware works, and the weather record on a particular route decides whether the route does. The money that has been made here came from bounded niches — campus crossings, event and restoration work, trading paths, and rural spans where the alternative is no service at all — rather than from metro access, which is what each funding wave has promised.

5 minLaser Internet: Connecting Rural Areas - with TaaraDW Shift · 50k+ views

6 minFree Space Optics: A Critical Technology for Space and Terrestrial CommunicationsFiberGuide - Fiber Optics | Optical Networking · 1k+ views

Recommendation ITU-R P.1817: Propagation data required for the design of terrestrial free-space optical links (International Telecommunication Union) · A Contemporary Survey on Free Space Optical Communication: Potential, Technical Challenges, Recent Advances and Research Direction (arXiv)

An optical ground station is the earth end of a space laser link: a telescope of roughly 0.4–1.5 m aperture on a fast mount inside a dome, a beacon laser that gives the spacecraft something to point at, a wavefront correction stage, and a receiver that is either photon-counting or coherent. The spacecraft terminal is a different problem and is covered on the space launch and satellites sheet under optical crosslinks and laser downlinks; the ground station's job is to undo the atmosphere. That atmosphere causes two separate failures with two separate fixes. Turbulence scrambles the wavefront, so the focused spot is a moving speckle pattern several times wider than the diffraction limit and only a few percent of the light couples into the single-mode fiber a coherent receiver needs. Cloud is worse and simpler: water is opaque at 1,550 nm, so a covered sky gives no link at all rather than a degraded one, and no amount of transmit power changes that. Everything else about station design follows from those two facts. The other reason operators are building them is feeder links: replacing the farm of Ka-band gateway antennas behind a very-high-throughput satellite with optical uplinks is how the industry expects to reach terabit-scale gateway capacity, and that is a ground-segment project rather than a spacecraft one.

The reason to build one is throughput per contact. NASA's Laser Communications Relay Demonstration runs each of its two optical space terminals at up to 1.2 Gbps, and the TBIRD experiment downlinked at 200 Gbps from a 6U CubeSat, rates no X-band or Ka-band pass approaches, and none of it needs a spectrum license or an ITU filing. Adaptive optics handles the turbulence well: a wavefront sensor and deformable mirror running at kilohertz rates raise single-mode fiber coupling from a few percent toward 30–50% and cut the depth of the fades. Clouds have no equivalent fix, and one site at a good high dry location gets only about 60–80% cloud-free line of sight over a year. Turbulence also has a schedule, strongest around local noon and weakest near sunrise and sunset, and it grows at low elevation angles, so an optical station has a higher minimum usable elevation than an RF antenna and gets shorter contact windows with a low-orbit satellite as a result. Several failures are purely operational: the dome stays shut in fog, high wind and after snow, the sun sweeps out seasonal keep-out zones the station cannot point into, wildfire smoke and volcanic particulate degrade links and dirty the optics, and in the US an uplink beam has to be coordinated so it is shuttered when a satellite crosses it.

Build optical ground capability when a mission is contact-limited rather than power-limited, meaning the spacecraft collects more data than its RF passes can carry down. If you can only fund one station, you have a demonstration rather than a service, so budget for a network from the start: as a rough guide, three well-separated sites reach roughly 90–95% availability and five to ten are needed to get past 99%. Pick sites for uncorrelated weather rather than for convenience, because two stations inside the same frontal system count as one. Keep an RF path for command and telemetry regardless, since a station that cannot open its dome still has to be told what to do. Check that the site has fiber before you fall in love with its seeing, because mountaintops chosen for clear air often have nothing behind them that can absorb 100 Gbps. And plan the handover: LCRD operators found that repointing the telescopes and reacquiring the link was fast, while rerouting the terrestrial network to a different station was the slow part, which is why delay-tolerant networking is used to buffer through the switch.

Telescope aperture roughly 0.4–1.5 m · adaptive optics lifts single-mode coupling from a few percent to 30–50% · one good site gives 60–80% cloud-free line of sight annually · three sites roughly 90–95% availability, five to ten needed past 99% · LCRD at 1.2 Gbps per optical terminal across two stations · TBIRD downlink at 200 Gbps from a 6U CubeSat · turbulence strongest at local noon, weakest near sunrise and sunset

NASA's LCRD ground segment is two stations, Optical Ground Station 1 at Table Mountain in California and Optical Ground Station 2 at Haleakalā in Hawaii, plus an RF station in New Mexico for the control path; MIT Lincoln Laboratory's TBIRD used Table Mountain as well. ESA operates a 1 m optical ground station at the Teide Observatory on Tenerife, and DLR runs a fixed station at Oberpfaffenhofen and a transportable one. On the commercial side, Kongsberg Satellite Services, Viasat and Airbus are building optical sites into existing teleport networks, and Cailabs sells a receiver that uses multi-plane light conversion to couple turbulent light into fiber without a deformable mirror.

The money goes into the telescope, mount and dome, the adaptive optics bench, and the site itself, and published prices are rare enough that the working figure people use is a few million dollars per station. The unusual part is the shape of the spending. Availability is bought by building more stations rather than by making one station better, so the second and third sites add the most availability per dollar and the eighth adds very little, which is the opposite of RF where a larger dish and more margin keep paying. Marginal cost per bit is close to zero once a station exists, so the whole economic question is how many sites the availability target forces. Site work is where budgets break: a mountaintop chosen for thin dry air usually has no fiber, and building a 100 Gbps path up to it can cost more than the telescope. Against that sits a Ka-band station at roughly $1–3M that works through rain with margin, which is why an operator with modest data volume has no reason to move. The buyers are relay operators and imaging constellations whose spacecraft collect far more than RF contacts can drain, and the likely structure is ground-station-as-a-service from a shared network rather than each operator building its own, for the same reason RF teleports consolidated: no single customer's traffic justifies ten sites.

8 minOptical Ground Station Oberpfaffenhofen OGSOPDLR Institut für Kommunikation und Navigation · 1k+ views

4 minWhat is Adaptive Optics?UCLA Galactic Center Group · 10k+ views

Insights and Observations from Operating a Geostationary Laser Communication Relay Mission: Operational Lessons from NASA's Laser Communications Relay Demonstration (LCRD) and Associated Optical Ground Stations (OGSs) (NASA) · Site Diversity in Downlink Optical Satellite Networks Through Ground Station Selection (arXiv)

These are the two ways to get a radio signal hundreds or thousands of kilometers past the horizon without a satellite, a tower chain or a cable. HF, 3 to 30 MHz, refracts off the ionosphere: a wave launched at a low angle bends back to earth 1,000–3,000 km away in one hop, and successive hops reach intercontinental distances. The frequency that works changes through the day, the season and the eleven-year solar cycle, roughly 3–8 MHz at night and 10–25 MHz in daylight, so an HF station chooses its channel continuously rather than once. Near-vertical incidence skywave is the same physics with the antenna aimed almost straight up: at 2–10 MHz the wave comes back over a circle out to about 300 km with no skip zone in the middle, which is how a unit covers a valley or the gap between ground-wave range and ordinary skywave range. Troposcatter works on a different mechanism, aiming a 1–5 GHz beam just above the horizon so that a small fraction of the power scatters off refractive-index irregularities in the troposphere and lands 100–300 km away. Loss on that path is 200–250 dB, which is why troposcatter has always meant high transmit power and large antennas. Neither path adds much propagation delay, since a 3,000 km ionospheric hop takes about 10 ms, but HF modems interleave over seconds to ride out fading, so the delay a user sees is set by the waveform rather than by the distance.

HF's advantage is that it needs no infrastructure whatsoever. A transceiver, a wire antenna and a battery reach across a continent with nothing in between that an adversary can jam from orbit, cut with an anchor, or refuse to sell you. The price is bandwidth: a standard HF channel is 3 kHz wide, which caps throughput in the kilobits and always will, and on top of that the path varies hour to hour, solar flares black the band out for minutes to hours (days at high latitudes during a polar cap absorption event), the band is crowded, and a transmitting station is easy to intercept and direction-find. Troposcatter trades that around. It delivers 10–200 Mbps over 100–300 km with nothing between the two ends and nothing in orbit, and because the useful scatter volume is narrow, a receiver off the great-circle path hears very little. Its weaknesses are power and mass: the scattered signal fades deeply and at random, so quadruple diversity — two antennas and two frequencies, or angle-diversity feeds — is standard equipment rather than an upgrade, amplifiers run from hundreds of watts into the kilowatts, and the terminal is vehicle-sized even now that antennas have come down from the 18 m billboards of the 1950s to 1.2–2.4 m dishes.

Use HF when the requirement is a message rather than a file and the path has to work with zero infrastructure behind it: command and control when satcom is denied, maritime and aeronautical long-haul, disaster response after the towers are down. Use near-vertical incidence skywave specifically when the users are within about 300 km of each other in terrain that blocks line of sight, since that is the one case ordinary skywave cannot cover. Do not plan a data service on HF, because 1.2–4.8 kbps in a 3 kHz channel is an order of magnitude short of what applications now assume; wideband HF widens the channel to 24 kHz and reaches roughly 120 kbps in good conditions, which carries email and compressed imagery but not video. Whatever the waveform, use automatic link establishment rather than manual frequency selection: 2G ALE under MIL-STD-188-141 sounds and scans and links in about ten seconds, 3G ALE under STANAG 4538 does it in one or two, and that is the difference between a network and an operator with a propagation chart. Reach for troposcatter when you need tens of megabits over 100–300 km, satellite capacity is jammed or unavailable, and a vehicle-sized terminal can be placed at each end. If line of sight exists at all, use a microwave hop instead, since it costs roughly a tenth as much and needs a tenth of the power.

HF 3–30 MHz in 3 kHz channels, 1.2–4.8 kbps typical and 9.6 kbps at best · wideband HF in 24 kHz reaching about 120 kbps · single-hop skywave 1,000–3,000 km, usable frequency roughly 3–8 MHz at night and 10–25 MHz by day · NVIS at 2–10 MHz covering out to about 300 km · ALE link setup about 10 s for 2G, 1–2 s for 3G · troposcatter 1–5 GHz over 100–300 km against 200–250 dB of path loss · modern troposcatter 10–200 Mbps on 1.2–2.4 m dishes

On the HF side, MIL-STD-188-141 for 2G ALE, STANAG 4538 for 3G, and MIL-STD-188-110 for the modem waveforms including the wideband appendix; the US federal SHARES network run by CISA as an interagency fallback; oceanic air traffic control, which still runs HF voice and HF data link where there is no VHF coverage. On the troposcatter side, the White Alice system across Alaska and NATO's ACE High chain across southern Europe were the Cold War build-outs, both retired once satellite capacity got cheap; the current revival is driven by contested environments, with Comtech supplying the US Army's troposcatter modernization and several NATO members restarting programs of their own.

Both of these trade a fixed cost for the absence of a recurring one, which is the whole reason they survive. An HF station is $5k–30k for a manpack or vehicle set with an antenna, and after that there is no airtime bill, no lease and, for a government user, no spectrum fee, against $500–3,000 a month for a satellite terminal doing a similar job. That arithmetic is why militaries and emergency services keep HF alongside far better links: the operating cost of a capability you use twice a year is what decides whether you still have it. A troposcatter terminal pair runs from several hundred thousand dollars to a couple of million, expensive next to a satellite terminal, but it buys a link with no space segment to lease and nothing an adversary can deny from orbit, and on a route carrying daily traffic the ten-year total usually comes out ahead. The cost that does not appear in either budget is people. HF networks need operators who understand propagation, and that skill base thinned badly between about 1995 and 2015 while satcom took over, so standing a network back up now takes longer to staff than to equip. Supplier counts are small in both categories, so prices are negotiated rather than competed down.

20 minUnderstanding NVISRohde & Schwarz · 50k+ views

3 minThe Cold War-era communications system that overcomes satellite jammingBFBS Forces News · 50k+ views

Recommendation ITU-R P.533: Method for the prediction of the performance of HF circuits (International Telecommunication Union) · Tropospheric Scatter: Theory vs. Predictive Models, NTIA TR-22-557 (NTIA Institute for Telecommunication Sciences)

Seawater is conductive, so radio dies in it: attenuation at 1 MHz is tens of decibels per meter, and only extremely low frequencies get through at all. The US Navy's 76 Hz system, retired in 2004, reached a submerged submarine and carried a handful of characters in fifteen minutes. Blue-green light does much better but only over 10–100 m in clear water and less in anything turbid. That leaves sound, and an acoustic modem is a transducer, a power amplifier and a signal processor that modulates a carrier somewhere between about 10 and 40 kHz. Three numbers describe the channel, and none of them are close to what a radio engineer expects: sound travels at roughly 1,500 m/s, so one kilometer costs 0.67 seconds each way. Absorption climbs steeply with frequency, from about 1 dB/km at 10 kHz to around 30 dB/km at 100 kHz, so the usable band is kilohertz wide rather than megahertz and shrinks further as range grows. And surface and bottom reflections give delay spreads of 10–100 ms, which at these symbol rates is hundreds of symbols of intersymbol interference that the receiver has to equalize away.

The strength is that nothing else works: acoustics reach kilometers on battery power, they work in water too turbid to see through, and the same hardware gives range measurements for free, which is why acoustic communication and acoustic positioning are usually the same box. Everything else on the list is a cost. Latency is set by the speed of sound and cannot be engineered down, so a 3 km link has a four-second round trip and any protocol that acknowledges each packet spends most of its time waiting, while bandwidth is set by absorption, so more transmit power buys range and not rate. Doppler is roughly a thousand times worse than in radio, because a 2 m/s platform against 1,500 m/s is a fractional shift of 1.3 parts per thousand, and the receiver has to track and resample continuously rather than correct once. Ambient noise comes from shipping below 100 Hz, from wind and breaking waves across most of the band, and in warm shallow water from snapping shrimp. Shallow water is much harder than deep water on every one of these counts, and performance shifts with the sound speed profile through the day, so a link that closed in the morning may not close in the afternoon.

Use acoustics whenever a node is submerged, has no cable, and needs to move data more than a few tens of meters: AUV command and telemetry, subsea production monitoring, ocean observatory offload, diver communication, and firing a release on a bottom package. Size the link by the range-rate product rather than by the headline data rate, because commercial modems land at roughly 40–50 km·kbps and the vendor quotes the rate at its shortest range; wanting 5 kbps at 10 km is asking for 50, which is at the edge of what exists. If the requirement is megabits, stop looking at acoustics and either put the node on a tether or a seafloor cable, or bring a vehicle within 10–100 m and use a blue-green optical link. If the requirement is a few hundred bytes a day, an acoustic release plus a satellite-linked surface buoy is usually cheaper than any network. Design the protocol around the delay: batch and pipeline rather than stop-and-wait, and expect retransmission to be expensive in both time and battery. If equipment from different vendors has to interoperate, JANUS (STANAG 4748) is the only standard, and at 11.5 kHz and about 80 bps it is a hailing channel to negotiate with rather than a data link to use.

Sound at roughly 1,500 m/s, so 0.67 s per kilometer one way and a 4 s round trip at 3 km · carriers of about 10–40 kHz with kilohertz of usable bandwidth · absorption roughly 1 dB/km at 10 kHz and 30 dB/km at 100 kHz · commercial rates from 80 bps to a few tens of kbps · range-rate product roughly 40–50 km·kbps, so 5 kbps at 10 km is at the edge · multipath delay spread 10–100 ms · modems $5k–30k

The WHOI Micromodem, widely used in research vehicles, runs 80–5,400 bps in about 4–5 kHz of band and is the reference point most published work compares against; EvoLogics S2C, Teledyne Benthos ATM sets, Sonardyne Modem 6, Kongsberg cNODE and Subnero cover the commercial market. NATO's JANUS became STANAG 4748 in 2017 and is the first standardized digital underwater protocol. NOAA's DART tsunami buoys are the clearest illustration of the trade: a bottom pressure recorder sends short acoustic messages up to a surface buoy, which relays them by satellite, because the acoustic hop is the only part that has to work underwater and it is kept as small as possible.

A modem is $5k–30k, more with a deep-rated housing and a matched transducer, and that number is small enough to be irrelevant next to the vessel that installs it. A survey vessel costs $30k–150k a day and an ROV support vessel more, so the economic case for an acoustic link is almost always the ship trip it avoids: a bottom node that offloads to a passing AUV or a surface buoy replaces a recovery cruise, and one avoided cruise pays for a great many modems. At the other end of the range, a cabled observatory gives unlimited bandwidth and continuous power but needs a cable ship to install: Ocean Networks Canada's NEPTUNE loop is roughly 800 km of cable and cost in the C$100M range. There is very little in between those two, so the practical choice is tens of thousands of dollars per node with kilobits, or tens of millions once with fiber. Offshore energy is the largest paying customer and defense the second, and both buy on demonstrated reliability rather than on data rate, which keeps prices high and the supplier list short. Anyone forecasting a step change in this market should be asked which of absorption, sound speed or ambient noise they expect to change.

1 min"What Is An Acoustic Modem?" Popoto Modem explains.Popoto Modem · 100k+ views

4 minOcean Acoustics | Ocean Literacy | FuseSchoolFuseSchool - Global Education · 10k+ views

Micromodem Overview (WHOI Acoustic Communications Group) · A Survey on Underwater Acoustic Sensor Networks: Perspectives on Protocol Design for Signaling, MAC and Routing (arXiv)

In a communications context terahertz means roughly 100 GHz to 1 THz, and the part being standardized sits at the bottom of that: D-band at 110–170 GHz, plus the 137 GHz of spectrum between 275 and 450 GHz that WRC-19 identified for fixed and land mobile use. IEEE 802.15.3d-2017 defines a point-to-point link across 252–325 GHz with channels from 2.16 up to 69.12 GHz wide, aimed at 100 Gbps. The attraction is bandwidth arithmetic: one 69 GHz channel is more spectrum than all the licensed mobile spectrum below 6 GHz combined, and at 5 bits per hertz that single carrier would deliver about 345 Gbps. What stops it is the air. Atmospheric absorption rises with frequency and is punctuated by water-vapor lines, so 183 GHz and 325 GHz are effectively walls, and the usable windows near 140, 220 and 300 GHz run from a few dB/km up to roughly 10 dB/km at ordinary humidity before rain is added. Solid objects are worse: a human body takes more than 30 dB out at 300 GHz and foliage is opaque, so these links are pencil beams a fraction of a degree wide and beam alignment rather than the modem is the hard engineering. The sources and detectors themselves are covered on the optics and photonics sheet.

The bandwidth is genuinely there, it is largely unclaimed because nobody uses it yet, and the same narrow beam that makes pointing difficult also makes interception difficult and lets many links share the same space without coordinating. Against that sit three hardware problems that have not moved much in a decade. Transmit power is the worst: CMOS and SiGe transceivers run out of gain somewhere around 100–200 GHz and deliver roughly 0 to +10 dBm, InP HEMT and HBT parts go higher at much higher cost and without integrating with the digital baseband, and the photonic route that produces the cleanest high-rate signals in the laboratory works by beating two lasers in a photodiode and needs an optical bench to do it. Phase noise from the frequency multiplier chains limits how high the modulation order can go, which caps spectral efficiency well below what the same modem achieves at 30 GHz. And the data converters are their own wall, because turning 100 GHz of bandwidth into bits calls for hundreds of gigasamples per second at usable resolution, which today costs tens of watts. Even with all of that solved, gas absorption and rain hold outdoor spans to hundreds of meters and low single-digit kilometers.

For a system that has to work now, do not use it. If you need multi-gigabit wireless today, buy an E-band link at 70/80 GHz, which carries 1–25 Gbps over 1–5 km with commercial hardware and light-touch licensing, or 60 GHz unlicensed for short indoor hops. If you are writing a 6G plan, note that 3GPP's own focus for 6G is the upper mid-band at 7–15 GHz rather than terahertz, because site count and coverage drive network cost far more than peak rate does. Put terahertz in the plan where its weaknesses do not bite: fixed short links, fronthaul between closely spaced radios, rack-to-rack inside a data center, and kiosk-style bulk transfer over a meter. When evaluating any claimed result, ask three questions before believing it applies to a product — how far did the link actually run, was it indoors, and was the transmitter a chip or a laboratory photonic bench. If the number quoted is a terabit, ask how many carriers were aggregated and over how many centimeters.

Roughly 100 GHz–1 THz, with 137 GHz between 275 and 450 GHz identified at WRC-19 · IEEE 802.15.3d channels of 2.16–69.12 GHz across 252–325 GHz, targeting 100 Gbps · one 69 GHz channel at 5 bits/Hz works out to about 345 Gbps, computed here · gas absorption a few dB/km at 140 GHz and roughly 10 dB/km near 300 GHz, effectively opaque at 183 and 325 GHz · body blockage over 30 dB at 300 GHz · silicon transceiver output around 0 to +10 dBm · reported outdoor demonstrations of 100 Gbps over about 100 m and lower-rate links to 500 m

IEEE 802.15.3d-2017 is the first standard written for these bands, and WRC-19's identification of 275–450 GHz is what makes them usable outside a laboratory. The funding vehicles are the visible part: the EU's Smart Networks and Services Joint Undertaking put €900M of public money behind 2021–2027 work, matched by industry, through projects including Hexa-X and TERAPOD, and NSF's RINGS program plays a similar role in the US alongside national programs in Japan, Korea and China. On results, NTT DOCOMO with NEC, NTT and Fujitsu reported 100 Gbps links at 100 GHz and 300 GHz over about 100 m in 2024, and LG with Fraunhofer HHI reported outdoor 6G transmission over 320 m in 2022 and 500 m in 2023 in the 155–175 GHz range.

Nobody is selling terahertz links, so the money in this category is public research funding and vendor R&D budgets: €900M from the EU for 2021–2027 with matched industry spending, plus the US, Japanese, Korean and Chinese equivalents. The first revenue that actually changes hands is test and measurement, because a 300 GHz frequency extender for a vector network analyzer is a six-figure instrument and the instrument makers get paid years before any radio ships. When a product does arrive, its economics will look like E-band's rather than 5G's, since the value is a fixed link that avoids a fiber build and the buyer is comparing a $20k radio against past $200k a mile for trenching in a city. Two things gate that arrival: solid-state transmit power at a price a link can carry, and an application that an E-band radio does not already serve at a tenth of the difficulty. The second is the harder one, and it is the question to put to any business plan in this space. A plan that treats terahertz as the capacity layer of 6G should also explain why 3GPP is working at 7–15 GHz instead, because a band that stops at a leaf does not reduce the site count that sets what a network costs.

7 minWhich will be the new spectrum for 6G? A practical reviewRohde & Schwarz · 1k+ views

8 minTerahertz CommunicationsProceedings of the IEEE · under 1k views

Recommendation ITU-R P.676: Attenuation by atmospheric gases and related effects (International Telecommunication Union) · A Survey on Advancements in THz Technology for 6G: Systems, Circuits, Antennas, and Experiments (arXiv)

A modem turns bits into a signal and back, and a waveform is the whole recipe it follows: carrier frequency, modulation, coding, framing, access method, and any hopping or spreading on top. In a software-defined radio the antenna and the analog front end still exist, but the signal is digitized early and everything after that point is arithmetic on samples rather than a circuit. Sampling has to run at least twice the bandwidth, so a 100 MHz channel needs roughly 200 million complex samples per second, and at 12 to 14 bits per sample that is 3 to 6 Gbps of raw data arriving before any demodulation happens. Older designs mixed the signal down to an intermediate frequency first and sampled at tens of megasamples per second; current parts sample the radio frequency directly at 2 to 6 gigasamples per second, which removes a whole analog mixing stage. Once the samples are in, filtering, demodulation, error correction and framing are all code, running on an FPGA for the sample-rate work and a CPU or DSP for everything slower. That is what makes a waveform a software artifact you can load, which is the idea NASA's STRS architecture and the military's Software Communications Architecture both formalized.

The strength is that one box handles many waveforms and takes new ones by software load, so a base station gets a standards release without a truck roll and a military radio takes a new hopping pattern without new hardware. Prototyping got cheap in the same move: a $30 RTL-SDR dongle covering 24 MHz to 1.7 GHz at 2.4 megasamples per second is a real receiver, and a $1,000 to $10,000 USRP-class board gives 56 to 200 MHz of instantaneous bandwidth with transmit. The weakness is energy. A demodulator running as general-purpose code costs one to two orders of magnitude more energy per bit than the same function as fixed silicon, which is why a phone's physical layer is hard-wired and only the protocol stack above it is software. Four things also stayed analog: the antenna, the low-noise amplifier, the filters, and the power amplifier. Filters need physical resonators, so a phone still carries dozens of discrete acoustic filters and duplexers, one per band combination, and a power amplifier runs at 30 to 50% efficiency and gets worse the more linearity a dense constellation demands. A wideband front end with no preselector also desensitizes easily, because a strong nearby transmitter consumes the converter's dynamic range and nothing in software gets it back.

If the waveform is fixed by a standard and you are shipping over a million units, buy a hard-wired modem chip and put the software effort into the stack above it. If you need several waveforms in one box, or the standard is still moving, or the same platform ships to customers in different bands, use a software-defined radio and accept the power penalty. If you are prototyping, monitoring spectrum, or teaching, start with a $30 dongle and move to a USRP-class board only when you need transmit or more than a few megahertz. If the device runs on a battery for years, such as a narrowband Internet-of-things sensor, use a fixed-function part, because the flexibility buys nothing and the current draw is the whole design. Whatever you pick, budget separately for the RF front end per band: a radio that is wideband on paper still needs its own filter, its own amplifier and its own matching for each band it actually operates in, and that hardware is usually where the schedule and the bill of materials go.

Sampling at 2× bandwidth, so about 200 Msps complex for a 100 MHz channel, and 3–6 Gbps of raw samples at 12–14 bits (derived here) · direct RF sampling converters at 2–6 Gsps · RTL-SDR dongle about $30, 24 MHz–1.7 GHz, 2.4 Msps, 8-bit · USRP-class board $1,000–10,000 with 56–200 MHz instantaneous bandwidth · software demodulation costs 1–2 orders of magnitude more energy per bit than fixed silicon · power amplifiers run at 30–50% efficiency and lose more as constellation density rises · RF front-end content in a 5G phone is typically $20–35, above the modem die itself.

GNU Radio and the Ettus USRP family, which together became the default research and prototyping platform; the RTL-SDR, a repurposed DVB-T television tuner chip that turned into a $30 general-purpose receiver; AMD Zynq RFSoC parts with eight direct-sampling converters on the same die as the FPGA fabric and Arm cores; NASA's Space Telecommunications Radio System and the SCaN Testbed flown on the International Space Station; the US military's Joint Tactical Radio System and its Software Communications Architecture; and every O-RAN radio unit, which is an SDR with a standardized interface to the baseband.

The cost curve that made software-defined radio the default was in the data converters. A 14-bit gigasample converter was a five-figure defense part in the 2000s, and the same capability now ships eight channels at a time inside a $2,000 to $10,000 RFSoC, with a $30 dongle at the bottom of the market because the television industry paid for that silicon first. For a systems company the effect is to move money from per-waveform hardware development to a platform plus software: one board, then engineering effort per waveform, which is why defense radio programs stopped buying a new box for every network. The margin sits with the merchant silicon vendors rather than the box builders, since almost everyone integrates the same handful of transceiver and RFSoC parts and differentiates on algorithms, calibration and integration. The part software never absorbed is where the remaining hardware profit went: filters, switches and power amplifiers are a separate supply chain dominated by Qorvo, Skyworks, Broadcom and Murata, and the RF front-end content in a 5G phone, typically $20 to $35, now runs above the modem die itself. If you are evaluating an SDR product claim, the useful question is what fraction of the bill of materials is still per-band analog hardware, because that fraction sets how much of the flexibility is real.

19 min#170: Basics of IQ Signals and IQ modulation & demodulation - A tutorialw2aew · 500k+ views

15 minThe Beginner's Guide To Software Defined Radio RTL-SDRTech Minds · 500k+ views

IQ Sampling (PySDR: A Guide to SDR and DSP using Python) · Space Telecommunications Radio System (STRS) Architecture Standard, Release 1.02.1 (NASA/TM-2010-216809/REV1)

A network needs two different things from a clock, and they cost different amounts to deliver. Frequency synchronization means every transmitter runs at the same rate, and 3GPP asks for ±0.05 parts per million at a macro base station and ±0.1 at a micro or pico cell. Phase and time synchronization means every transmitter agrees on when a moment is, and that is the hard one: 5G time-division duplex needs neighboring cells aligned to within about 3 microseconds of each other, which operators build as ±1.5 microseconds against a common reference. GPS is where nearly all of that reference comes from, because a $50 to $300 receiver module recovers UTC to tens of nanoseconds anywhere with a view of the sky, and no other source is remotely that cheap. Two mechanisms distribute it once you have it. Synchronous Ethernet, defined in ITU-T G.8262, carries frequency in the physical line rate itself, so it does not degrade under traffic load but carries no notion of time of day. Precision Time Protocol, IEEE 1588 in the telecom profile ITU-T G.8275.1, carries phase and time in timestamped packets, and it works only because every switch and router in the path acts as a boundary clock that removes its own queuing delay; the partial-support profile G.8275.2, which tolerates nodes that do not participate, is far harder to hold inside budget.

GPS timing is close to free, globally available, and traceable to national timescales, which is why it displaced the older practice of distributing frequency down the transmission hierarchy from a cesium standard in a central office. Its weakness is the received signal, which arrives around −160 dBW and is trivially jammed or spoofed; the sensing and navigation sheet covers the receiver and its defenses in detail. When GPS goes away, a site runs on its own oscillator, and holdover is arithmetic: time error equals fractional frequency offset times elapsed time, so an oven-controlled crystal holding 1 part in 10¹⁰ drifts 1.5 microseconds in 15,000 seconds, about four hours, while a rubidium at 1 part in 10¹¹ lasts around 42 hours and a cesium at 1 part in 10¹³ lasts months (derived here). The second weakness is that phase synchronization is a property of the whole path, not of a box: one switch in the chain that is not a boundary clock can add hundreds of nanoseconds of variable delay and there is no way to calibrate it out. The third is that timing failures are quiet. A cell whose phase has drifted does not stop; it interferes with its neighbors, throughput falls, and the alarm that fires is usually about capacity.

If you run a frequency-division network or a transport network, Synchronous Ethernet plus a GPS reference at a few sites is usually enough, and you can leave phase alone. If you run 5G time-division duplex, you need phase, and the decision is where the reference lives: put a GPS receiver at every site if the antenna placement is easy and you accept a jamming exposure at every site, or run G.8275.1 from a small number of grandmasters if every node between them and the cells is already a boundary clock. If any node in the path is not, do not plan on the partial-support profile as a first choice; upgrade the path during the next hardware refresh instead, because retrofitting boundary clocks one node at a time is the expensive way to do it. Size holdover against the outage you actually expect to ride out: an oven-controlled oscillator for four hours of margin, rubidium if you need to survive a day or two of jamming or a fiber cut, and a cesium primary reference only where a site is meant to be autonomous. Outside telecom the questions are the same and only the tolerance changes, since MiFID II requires trading systems to hold 100 microseconds of UTC and synchrophasors in a substation need about a microsecond. And test the failure explicitly, because a network that has never lost GPS has never demonstrated holdover.

Frequency accuracy ±0.05 ppm at a macro base station, ±0.1 ppm at micro and pico (3GPP TS 38.104) · 5G TDD needs neighboring cells within about 3 microseconds, built as ±1.5 microseconds to a common reference · GPS recovers UTC to tens of nanoseconds from a $50–300 receiver module · holdover to 1.5 microseconds is about 4 hours at 1 part in 10¹⁰, 42 hours at 1 part in 10¹¹, months at 1 part in 10¹³ (derived here) · G.8275.1 assumes every node is a boundary clock; G.8275.2 tolerates nodes that are not · GPS signals arrive at roughly −160 dBW · MiFID II requires 100 microseconds of UTC for high-frequency trading and synchrophasors need about 1 microsecond.

ITU-T G.8275.1 and G.8262, the two recommendations nearly every mobile operator's synchronization plan is written against; IEEE 1588-2019 as the underlying protocol; the GPS receiver on almost every macro cell site, substation and trading venue; the UK Government Office for Science Blackett review, which cataloged national dependence on satellite time; MiFID II RTS 25, which forced timestamping to 100 microseconds of UTC across European trading; IEEE C37.118 synchrophasors, which need about a microsecond; and alternative sources sold against GPS outage, including eLoran, Satelles satellite time and location, and White Rabbit fiber time transfer.

Timing is a small line item that gates a large one, which is the whole reason it gets underfunded. A GPS timing module is $50 to $300, a site grandmaster with an oven-controlled oscillator runs a few thousand dollars, one with rubidium runs into the tens of thousands, and a cesium primary reference is a six-figure purchase before installation. Against that, the asset being protected is a macro site carrying revenue-generating traffic and the spectrum license behind it, so the sensible way to think about the spend is as insurance priced against hours of degraded capacity. The expensive part is never the clock itself; it is the distribution path, because G.8275.1 requires boundary-clock support in every switch and router between the grandmaster and the cell, and upgrading a transport network node by node costs far more than the timing equipment. That is why operators do phase synchronization as part of a hardware refresh rather than as a project. Governments pay for GPS and every network consumes it free, which has left almost no commercial market for alternatives: eLoran, satellite time services and fiber time transfer all sell at thousands of dollars per site per year against a receiver that costs a hundred dollars once, so buyers so far are the few operators of critical infrastructure with a regulator asking about the failure.

15 minHow a PTP slave syncs with a PTP masterDavid Gessner · 50k+ views

18 minIntroduction to Precision Time Protocol (PTP) for Network SynchronizationCisco Industrial IoT · 10k+ views

G.8275.1: Precision time protocol telecom profile for phase/time synchronization with full timing support from the network (ITU-T) · Satellite-derived time and position: Blackett review (UK Government Office for Science)

A base station is a baseband unit plus one radio per sector, and in 5G the radio is usually an active antenna unit with 32 or 64 transmit chains behind a panel about a meter tall. 3GPP separates base stations into three classes by how close a user is allowed to get: wide area for macro cells, specified against a minimum coupling loss of 70 dB, medium range for micro cells at 53 dB, and local area for pico cells at 45 dB. In practice that maps to power. A macro sector radiates a few hundred watts of RF and covers 1 to 3 km in mid band; a small cell radiates 1 to 10 W from a streetlight or a utility pole and covers 100 to 300 m; a pico or femto cell inside a building radiates under a watt. A distributed antenna system takes a different approach to the same coverage problem: one signal source feeds many small antennas spread through a building, either over coaxial cable and splitters in a passive design or over fiber to remote radio units in an active one. Active systems cost more and carry multiple operators and multiple bands; passive systems are cheaper and usually serve one operator, often fed by a repeater picking up the outdoor macro signal.

Macro cells are the cheapest capacity per bit anyone can build, because one site with 100 MHz of mid band delivers 400 Mbps to 1 Gbps per sector across three sectors and reuses a structure, a power feed and a backhaul circuit that are already paid for. Their weakness is that coverage and capacity are the same resource: when a sector fills, the only fixes are more spectrum, more antenna ports, or another site, and another macro site is the expensive one. Small cells fix capacity exactly where it is needed and nowhere else, which is their strength and their problem, since a network of small cells multiplies the number of leases, power drops, backhaul circuits and maintenance visits without multiplying coverage. Distributed antenna systems are the only clean answer for a large building, because mid-band and millimeter-wave signals lose 15 to 25 dB through modern coated glass and no amount of outdoor power fixes it. Their weakness is that they are construction projects inside somebody else's building, priced per square foot, with a cable plant that is hard to upgrade when a new band arrives.

Add capacity to the macro layer first: more spectrum on an existing site, then more antenna ports, then sector splits, because all three reuse a site you already pay for. If a specific place stays congested after that, such as a transit plaza or a stadium concourse, put small cells there and expect the recurring costs rather than the radio to dominate. If the problem is inside a building, size the answer to the building: under roughly 100,000 square feet a passive system fed by a repeater usually suffices at $1 to $2 per square foot, and above about 500,000 square feet, or wherever more than one operator has to be carried, use an active fiber-fed system at $2 to $4. If you own the building and the carriers will not fund it, look at a neutral-host system or a private network on shared spectrum before assuming the coverage problem is theirs to solve, because for anything short of an airport or an arena it will not be. And check the backhaul before buying any of it: a small cell or a DAS head end behind a 1 Gbps circuit delivers 1 Gbps no matter what the radios can do.

3GPP base station classes at 70 dB minimum coupling loss for wide area, 53 dB medium range, 45 dB local area · macro sector radiates a few hundred watts of RF and covers 1–3 km in mid band, 400 Mbps–1 Gbps per sector on 100 MHz · small cell 1–10 W over 100–300 m · in-building pico and femto cells under 1 W · coated glass costs 15–25 dB above about 3 GHz · in-building DAS typically $1–2 per square foot passive, $2–4 active and multi-operator · FCC caps of $500 for an application covering up to five small wireless facilities and $270 per facility per year for recurring fees.

Ericsson AIR, Nokia AirScale and Samsung massive MIMO radios as the volume macro products; the 3GPP wide area, medium range and local area classes in TS 38.104; Corning, CommScope and SOLiD in-building DAS installations in airports and stadiums; neutral-host operators such as Boingo and Freshwave selling in-building coverage back to carriers; CBRS-based private networks used as an alternative to DAS in warehouses and campuses; and the FCC's 2018 Declaratory Ruling and Third Report and Order, which capped small cell fees and set 60- and 90-day shot clocks.

The radios are the small part of everything on this list. A macro base station's radios and antennas are roughly a quarter of what a new site costs to build, and the rest is steel, foundation, power, permits, backhaul and the crew; upgrading radios on an existing site is therefore far cheaper per delivered bit than anything that needs a new location. Small cells invert the ratio in the worst way: the radio is a few thousand dollars, and the installed cost lands in the tens of thousands once pole attachment, a power drop, fiber and permitting are counted, with a recurring bill of a few thousand dollars a year per unit for site rent, power and backhaul forever. That recurring line is what the FCC's 2018 order attacked, capping application fees at $500 for up to five facilities and recurring right-of-way fees at $270 per facility per year, and setting 60-day and 90-day review deadlines, because municipal fees of $1,000 or more per pole per year were what made dense deployment uneconomic rather than any equipment price. In-building coverage is usually somebody else's problem to fund, and that is the single most useful fact in this entry: carriers pay for DAS in venues where the traffic and the brand exposure justify it, meaning stadiums, airports and convention centers, and decline nearly everywhere else, which leaves the building owner paying $1 to $2 per square foot for a passive system or $2 to $4 for an active multi-operator one. The neutral-host model exists to arbitrage that gap, building the system once and charging each carrier for access, and shared-spectrum private networks compete with it by letting the building owner skip the carriers entirely.

3 minMacrocell vs. Small Cell vs. Femtocell: 5G Base Stations ComparedEye on Tech · 10k+ views

5 minPassive Distributed Antenna System (DAS) – All You Need To KnowWilson Connectivity · 100k+ views

ETSI TS 138 104: 5G; NR; Base Station (BS) radio transmission and reception (ETSI / 3GPP) · Accelerating Wireless Broadband Deployment by Removing Barriers to Infrastructure Investment: Declaratory Ruling and Third Report and Order, FCC 18-133 (Federal Communications Commission)

A cell site is a piece of real estate with steel on it. The parts are a fenced compound with an access road, a foundation, a structure, cabinets or a shelter holding the baseband and the power plant, a backhaul circuit, grounding and lightning protection, and obstruction lighting if the structure is tall enough to need it. Structures come in three shapes: a guyed mast is the cheapest per foot and needs several acres for the anchors, a self-supporting lattice tower needs a compound rather than a field, and a monopole costs the most and is often the only thing a municipality will approve. Typical macro heights run 30 to 60 m, and the structure is engineered under TIA-222 for the wind load of the antennas hanging on it, which is why adding a third tenant frequently triggers a structural analysis and a reinforcement job before anything gets mounted. Power is a utility drop feeding rectifiers that produce −48 V DC, a battery string sized for 4 to 8 hours, and in many markets a generator behind that; ITU-T L.1210 is the recommendation that covers these arrangements for 5G-era sites. Ownership is split in a way that surprises people: in the United States the tower company owns the structure and holds the ground lease, and the carrier owns the radios, the antennas and the backhaul circuit.

A tower is a 30-year asset carrying equipment that turns over every 5 to 7 years, so the steel absorbs several technology generations without being touched, and that durability is why the owner can write leases longer than any equipment cycle. The weakness is that the binding constraints are legal and physical rather than technical. Structural capacity caps how many tenants a tower can hold, ground leases have to be renewed with a landowner who knows what the site is worth, and zoning is decided by whoever shows up at the hearing. FCC shot clocks presume 90 days for a collocation and 150 days for a new site, and the record in the Commission's 2018 infrastructure proceeding includes jurisdictions taking one to three years anyway. Power is the other soft spot: a site with a bad grid connection needs a generator and someone to keep fuel in it, and in off-grid markets fuel logistics and fuel theft become a running operating problem rather than a line item.

Collocate on an existing structure whenever the radio-frequency plan tolerates it, because the approval is faster, the shot clock is shorter, and you skip the foundation entirely. Build a new tower only when no existing structure covers the search ring and the traffic justifies a 20-year commitment, and start the site acquisition and zoning work 12 to 24 months before you need the site on air, since that is the schedule driver and the radios are not. Use a rooftop where you need coverage in a dense downtown and the height is available, and accept that rooftop landlords renegotiate more aggressively than ground-lease landowners. If you are the operator, keep leasing rather than owning towers unless you have a reason to hold the real estate, because the tower companies will build to suit and carry the capital. If the site is off grid or on a grid that fails daily, size solar plus batteries to cover the normal load and keep the generator for the exception, which typically cuts generator run time enough to pay back the hardware in a few years.

Typical macro structure 30–60 m, engineered under TIA-222 for antenna wind load · a new US macro site usually costs $250,000–350,000 all in, with radios and antennas roughly a quarter of it · tower rent about $2,000–5,000 per month per tenancy on initial non-cancellable terms of 5–10 years, escalating around 3% a year in the US · a macro site draws about 3–5 kW, so 26,000–44,000 kWh a year (derived here) · backup batteries sized for 4–8 hours · FCC shot clocks of 90 days for a collocation and 150 days for a new site · American Tower held 149,686 communications sites at the end of 2025, with about 80% of its towers on leased land.

American Tower, which reported 149,686 communications sites at the end of 2025 including 42,224 in the US and Canada, with roughly 80% of its 148,824 towers on land it leases; Crown Castle and SBA Communications, the other two large US owners; Cellnex in Europe, Indus Towers in India and IHS Towers in Africa and Latin America; TIA-222 as the structural standard; ITU-T L.1210 for site power feeding; and American Tower's African operations, where it powers tenant equipment through shared generators, solar panels and battery storage under contracts that reduce its revenue if power availability falls below the agreed level.

Tower ownership is a real-estate business, and one ratio decides how well it does: tenants per tower. The tower company pays once to build, then pays ground rent, property tax, power and maintenance whether one tenant is on the structure or three, and the incremental cost of adding a tenant is close to nothing, so almost all incremental rent falls through to operating profit. Work an example: a $275,000 site with roughly $20,000 a year of ground rent, tax and maintenance, renting at $2,500 a month per tenancy, produces $30,000 of revenue and $10,000 of operating profit with one tenant, which is 3.6% on cost, and $90,000 of revenue and $70,000 of profit with three, which is 25% (arithmetic derived here, using the middle of the $2,000–5,000 rent range). That is why tower companies buy portfolios with low tenancy and why carriers consolidating onto fewer networks is the risk their investors watch. The lease terms make the revenue look like a bond: initial non-cancellable terms of five to ten years with multiple renewal options that commonly carry a tenant past 20 or 30 years, escalators averaging about 3% a year in the United States and inflation-indexed in most other markets, and American Tower alone reports more than $54 billion of non-cancellable future tenant lease revenue already under contract. Power is the cost line that varies most by geography: an on-grid site drawing 3 to 5 kW spends a few thousand dollars a year on electricity, while an off-grid site running a generator most of the day burns on the order of 15,000 to 20,000 liters of diesel a year, which is why solar and battery hybrids get installed in Africa and South Asia and rarely in North America. Backhaul is the carrier's cost rather than the tower company's, at roughly $500 to $4,000 a month per site for a leased circuit, and the schedule is set at the far end of the process from the equipment: site acquisition and zoning decide when a site goes on air, and radios ship in weeks.

13 minHow is a Cell Tower Built? (Part 1: Foundation)Mom Sally Vlogs · 10k+ views

7 minA Climber's Perspective - Stacking of a 490' Self-Supporting Tower in Watts, Oklahoma via HelicopterJTS - Johnston Technical Services, Inc. Channel · 1m+ views

American Tower Corporation Annual Report on Form 10-K for the year ended December 31, 2025 (U.S. Securities and Exchange Commission) · L.1210: Sustainable power-feeding solutions for IMT-2020 networks (ITU-T)

Above the physical layer everything is packets, and the job of an IP network is to get each one to the right next hop. A router reads the destination address and looks it up in a forwarding table by longest prefix match, taking the most specific entry that matches rather than the first one. Inside a single network that table is built by OSPF or IS-IS, which know the whole topology; between networks it is built by BGP, a path-vector protocol in which each autonomous system tells its neighbors which prefixes it can reach and through which networks. About 77,900 autonomous systems were visible in the IPv4 routing system at the end of 2025, and the default-free table they collectively produce held roughly 1,040,000 to 1,050,000 IPv4 prefixes, growing by about 44,000 entries a year; the IPv6 table is near a quarter of a million entries and growing at about 27,000 a year. Those tables have to be searched once per packet, so a line card carrying 10 Tbps has tens of nanoseconds to make each decision, which is why forwarding tables live in content-addressable memory rather than ordinary RAM. Underneath all of it sits a separate optical layer that carries wavelengths between sites and knows nothing about packets, and whether the two layers should stay separate is the longest-running argument in transport network design.

The strength is that nobody has to agree on anything except reachability. Each network sets its own routing policy, prefers whichever neighbors it likes, and announces what it chooses, which is how roughly 78,000 independently owned networks interconnect with no central operator and no negotiated topology. The weakness is the same property, because BGP believes what it is told: a network that announces someone else's prefix, or passes a route learned from one provider on to another, pulls in traffic that was never meant for it. That has happened repeatedly at global scale, and Pakistan Telecom took YouTube off the internet for most of two hours in February 2008 by announcing a more specific prefix than YouTube's own. A leak through a small Pennsylvania network reached Verizon in June 2019 and took large parts of Cloudflare offline, and in October 2021 Facebook withdrew its own prefixes during a configuration change and was unreachable for about six hours. RPKI origin validation now lets a network reject an announcement from an origin not authorized to make it, which removes the simplest hijacks but says nothing about whether the rest of the path is real. The slower problem is table growth: a router bought with a million forwarding entries was comfortable in 2020 and is out of room now, and the fix is new hardware rather than a configuration change.

If you connect to exactly one upstream provider, do not run BGP. A default route does the same work, and pulling a million-entry table into a router with one exit costs memory and buys nothing. If you connect to two or more providers, or you hold address space that has to stay reachable through either of them, run BGP and take full tables, because choosing between paths is the whole point of multi-homing. Filter every announcement you accept from a customer against a prefix list, publish route origin authorizations for the address space you hold, and drop RPKI-invalid routes; those three steps are what the MANRS program asks for and they remove most of the accidental incidents. On the layering question, put coherent pluggables straight into the router when you own the fiber, the route needs only a handful of wavelengths, and one team runs both layers. Keep a separate optical line system when the route needs dozens of wavelengths, when you want the optical layer to restore a wavelength without the routers noticing, or when the fiber is leased from someone whose line system you do not control.

About 77,900 autonomous systems visible in IPv4 at the end of 2025 · default-free table of roughly 1,040,000–1,050,000 IPv4 prefixes, growing about 44,000 a year · IPv6 table near 250,000 entries, growing about 27,000 a year · tens of nanoseconds per forwarding decision on a 10 Tbps line card · 400ZR carries 400 Gbps over roughly 80–120 km of DWDM in a pluggable drawing under about 25 W · IP transit fell from over $1,000 per Mbps per month in the late 1990s to roughly $0.05–0.15 today · a 100 Gbps commitment at $0.10 works out to about $10,000 a month (derived from those two figures).

Cisco 8000 with Silicon One, Juniper PTX with Express silicon, Arista 7800 and Nokia's FP5, which carry most core traffic between them; the OIF 400ZR implementation agreement and the OpenZR+ extension, which put coherent optics in a router faceplate; internet exchanges such as DE-CIX Frankfurt, AMS-IX and LINX; RPKI and the MANRS program, which between them define what filtering a well-run network is expected to do.

Almost all of the cost is the box and the circuit, and the marginal cost of a packet is close to zero once both exist. IP transit, meaning the service of reaching everyone else, fell from over $1,000 per Mbps per month in the late 1990s to roughly $0.05–0.15 today, so a 100 Gbps commitment at $0.10 costs about $10,000 a month (derived from those two figures). Most large content traffic never pays that: it arrives settlement-free over a peering session at an exchange, where a 100 Gbps port typically runs $1,000–2,000 a month, or from a cache the content provider installed inside the access network. That leaves routers as the capital line item, and the money there has moved from custom silicon toward merchant chips, with Broadcom's Jericho and Tomahawk families now carrying a large share of ports that once required a vendor's own ASIC. The layering argument is mostly economic. A 400ZR pluggable carries 400 Gbps over roughly 80–120 km of DWDM, costs a few thousand dollars, and draws under about 25 W, against a transponder port that costs several times more, occupies a shelf and needs its own management system. What the operator gives up is the optical layer's ability to restore a wavelength on its own, and that resilience then has to be bought back at the IP layer instead.

16 minBGP: Border Gateway Protocol - ComputerphileComputerphile · 100k+ views

6 minIP over DWDM: Advantages and challengesFiberGuide - Fiber Optics | Optical Networking · 1k+ views

BGP in 2025 (APNIC Blog) · A Border Gateway Protocol 4 (BGP-4) (RFC 4271, IETF)

Software-defined networking means separating the control plane, which decides where traffic goes, from the forwarding plane, which moves it, and putting the decision in software that can see the whole network at once. The first widely used version was OpenFlow, published in 2008 and standardized by the Open Networking Foundation from 2011, which let a controller write forwarding rules directly into a switch. Almost no production wide-area network works that way now. What shipped instead is a controller driving standard protocols, using BGP-LS to learn the topology, PCEP or NETCONF to install paths, and segment routing to carry them, which leaves the routers able to keep forwarding when the controller is unreachable. Network slicing applies the same idea to a mobile network: a slice is an end-to-end logical network spanning radio, transport and core, identified by an S-NSSAI, with its own scheduling treatment and its own service-level agreement, and 3GPP defines standard slice types for mobile broadband, low-latency traffic, machine-type devices and vehicles. A device can be attached to at most eight slices at once. Slicing needs a 5G standalone core, because the non-standalone architecture most 5G networks launched with anchors the control plane in the LTE core and has nowhere to put a slice identifier.

A controller with a global view can do things no per-hop protocol can. Google's B4 inter-datacenter network, described publicly in 2013, keeps its links close to fully loaded by scheduling large transfers centrally, against the 30–40% average utilization typical of a wide-area network sized for its worst hour. SD-WAN, the commercial product built on the same idea, is the part of this that clearly paid: it replaced branch MPLS circuits costing roughly $600–1,500 a month with business broadband at $100–300, and used per-packet path selection to hold the application quality the private circuit used to guarantee. Slicing has produced very little revenue, and the reasons are specific rather than a matter of it being early. It needs a standalone core, which fewer than a third of commercial 5G networks have deployed; it needs devices, ordering systems and sales compensation that can express a slice; and a capacity guarantee is only worth paying for when the network is congested, which is a few hours a day in a few places. Enterprises that genuinely need a guarantee mostly buy a private cellular network instead, because they can see the equipment and nobody else can reallocate it.

If you are buying wide-area connectivity for tens or hundreds of branch sites, use SD-WAN. It is the one part of this with an obvious payback, and the arithmetic is just the circuit bill. If your backbone carries more traffic than shortest-path routing spreads evenly, put a controller on top of segment routing and let it compute paths, but keep the routers able to run without it, because a controller outage should not be a network outage. If you are evaluating a plan whose revenue depends on selling slices, discount it heavily unless the operator already runs a standalone core, the customer's traffic really competes with consumer traffic in the same sectors, and the customer can verify what was delivered. If the requirement is guaranteed capacity at a fixed site, price a private cellular network first, since it removes the argument about who gets the spectrum when the cell fills. And if a proposed slice is a quality-of-service class under a new name, treat it as one: dedicated bearers and dedicated access point names have been available since LTE and cost less to operate.

B4 keeps links close to fully loaded against 30–40% typical wide-area utilization · branch MPLS circuits at roughly $600–1,500 a month against business broadband at $100–300 · fewer than a third of commercial 5G networks run the standalone core slicing requires · a device can be attached to at most 8 slices at once · four standard 3GPP slice types, covering mobile broadband, low latency, machine-type devices and vehicles · OpenFlow published in 2008 and standardized from 2011.

Google's B4 wide-area network, the first large software-defined backbone described in public; the Open Networking Foundation's OpenFlow specification and the ETSI NFV industry specification group, which set most of the vocabulary; SD-WAN products such as Cisco Viptela, VMware VeloCloud, Fortinet Secure SD-WAN and HPE Aruba EdgeConnect; T-Mobile US T-Priority, launched in September 2024 as a 5G slice for first responders; 3GPP TS 28.530, which defines what a slice is and what has to be managed.

The two halves of this have opposite economics. SD-WAN is a straight substitution: the customer buys an appliance and a subscription, usually a few hundred dollars a month per site, and stops paying $600–1,500 for a private circuit, so the payback shows on the first bill and the market grew to a few billion dollars a year on that basis. Slicing is a large fixed cost with no matching revenue line yet. An operator has to deploy a standalone core, an orchestration layer, and changes to ordering, assurance and billing systems before anything spanning radio, transport and core can be sold at all, and none of that produces an invoice until an enterprise buys a slice. The marginal cost of one more slice is close to zero once all of it exists, which is what makes the model attractive on paper, but the fixed cost lands years earlier and the willingness to pay has been small. Where money has appeared it is in public safety and large events, where the buyer values priority during exactly the congested hour a slice is good for, and T-Mobile's first responder service is the clearest example. The useful question for an investor is which customer has a verifiable problem that a dedicated bearer or a private cellular network does not already solve more cheaply; slicing works technically, and that has not been the constraint.

9 minSoftware Defined Networking - ComputerphileComputerphile · 100k+ views

19 min5G Network Slicing Defined | MpiricalMpirical · 50k+ views

Software-Defined Networking (SDN) Definition (Open Networking Foundation) · ETSI TS 128 530: Management and orchestration; Concepts, use cases and requirements (ETSI / 3GPP)

A network operator can encrypt traffic at three different layers, and the choice decides how much overhead it costs and how much of the network can still read the plaintext. At layer 1 the transponder encrypts the whole wavelength with AES-256 before it goes into the fiber, which adds no bytes at all and under a microsecond of latency, but protects only the span between two transponders. At layer 2, MACsec (IEEE 802.1AE) encrypts each Ethernet frame between two ports with AES-GCM, runs inside the MAC so it holds line rate at 100G and above, and adds a security tag of 8 or 16 bytes plus a 16-byte integrity check value, roughly 2% of a 1,500-byte frame. At layer 3, IPsec (RFC 4301) wraps the IP packet in ESP between two gateways, adds about 60 bytes and forces a lower MTU, and negotiates keys with IKEv2. Keys come from one of two places: a pre-shared secret configured per link, or certificates issued by an internal authority, with the private material held in a hardware security module. The operational content of this entry is mostly the second half of its title, because the algorithms are settled and the failures are almost always about which key was where and when it last changed.

Encrypting lower down costs less and sees less. Layer 1 encryption protects everything on the wavelength, including the control traffic, with no throughput penalty and no per-flow policy, which suits a leased fiber pair between two of your own buildings. MACsec is hop by hop, so every switch in the path decrypts and re-encrypts, and a compromised intermediate device sees plaintext; that is fine inside a building and wrong across a carrier's network. IPsec survives an untrusted transport network because the gateways are the only endpoints, but done in software it costs roughly a core per few gigabits of AES-GCM, so anything past about 10 Gbps needs hardware offload, and the MTU reduction breaks applications that set the do-not-fragment bit. None of the three hides traffic analysis: volumes, timing and endpoint pairs stay visible unless you pad, which almost nobody does. The most common real failure is not cryptographic at all but an expired certificate or a pre-shared key nobody rotated in six years, and the outage that follows is usually longer than any breach would have been.

If both ends of the span are yours and the fiber or wavelength is leased from someone else, use layer 1 encryption; it is a per-port option on the transponder you were buying anyway and costs no bandwidth. If the path is Ethernet across a small number of devices you own, turn on MACsec, since modern switch silicon includes it and the only cost is running the key agreement. If the path crosses networks you do not control, or you need different policies for different tenants over the same circuit, use IPsec and budget for offload hardware above 10 Gbps. If the data has to stay confidential into the late 2030s, assume a copy of today's ciphertext is already stored somewhere and move the key exchange to a post-quantum or hybrid one now, because AES-256 stays strong against a quantum attacker while the public-key exchange that established the key does not. In IKEv2 the two practical routes are a long post-quantum preshared key (RFC 8784) or multiple key exchanges combining a classical and a lattice algorithm (RFC 9370); the algorithms themselves, and quantum key distribution as an alternative, are covered on the quantum technologies sheet.

Layer 1 AES-256 encryption adds no bytes and under 1 microsecond of latency · MACsec adds an 8- or 16-byte security tag plus a 16-byte integrity check, roughly 2% of a 1,500-byte frame · MACsec runs at line rate at 100G and above because it sits in the MAC · IPsec adds about 60 bytes per packet and forces a lower MTU · software IPsec costs roughly one core per few gigabits, so offload is needed past about 10 Gbps · AES-256 keeps roughly 128-bit strength against a quantum attacker, while the public-key exchange does not.

IEEE 802.1AE MACsec with the 802.1X key agreement, which is what most campus and data-center switch silicon implements; optical-layer encryption in Ciena, Nokia and Infinera transponders, sold as a per-port option; IPsec as specified in RFC 4301 with IKEv2 from RFC 7296, plus RFC 8784 and RFC 9370 for the post-quantum transition; the NSA's Commercial Solutions for Classified program, which requires two independent layers of commercial encryption from different implementations rather than one.

The cryptography is free and the key management is not. MACsec is a feature of switch and NIC silicon that is already paid for, so turning it on costs nothing in hardware and shows up entirely as operations: a certificate authority or a key store, a rotation schedule, and someone who owns it. Layer 1 encryption is sold as a per-port license adder on a transponder, which is real money but small against the line system it rides on, and it is the cheapest way to cover a leased fiber pair because it consumes no capacity. IPsec is the expensive one, since above about 10 Gbps it means buying gateways whose only job is encryption, and the throughput number on the datasheet is measured with large packets and falls sharply with small ones. Against all of that, the dominant cost of the post-quantum transition is inventory rather than algorithms: finding every device, appliance and protocol in the estate that does a public-key exchange, which is why large migrations are planned over five to ten years. For a buyer the useful rule is that the encryption line item is small and the credential lifecycle is where the outages and the audit findings come from, so budget for the second even when the first is a checkbox.

16 minMACSec (Media Access Control Security) OverviewBalram Shekhawat · 10k+ views

6 minIPsec ExplainedPowerCert Animated Videos · 100k+ views

802.1AE: MAC Security (MACsec) (IEEE 802.1 Working Group) · SP 800-77 Rev. 1, Guide to IPsec VPNs (NIST Computer Security Resource Center)

Interference is anything that raises the noise floor at a receiver or occupies the channel it was going to use, and most of it is accidental. The routine sources are co-channel energy from a neighboring cell, adjacent-channel leakage from a transmitter that is less clean than its specification, passive intermodulation from a corroded connector or a rusted joint on a tower that mixes two strong transmit signals into a product landing in the receive band, and consumer signal boosters installed backwards. Deliberate jamming is the same physics with intent, and the link budget favors the jammer, whose energy travels one way into the victim receiver while the wanted signal has already spent its path loss getting there. What actually takes commercial networks down is a different list. Of 188 major telecom incidents reported to ENISA by European regulators for 2024, 60% were system failures, 19% human error and 13% natural phenomena, while 8%, or 15 incidents, were malicious. The single largest technical cause was cable cuts at 41 incidents, 23% of the total, followed by faulty software changes at 14% and software bugs at 13%.

The tool that works best is route diversity, and it is also the hardest to verify. Two circuits from two carriers routinely share a duct, a bridge crossing, or the single conduit entering the building, because carriers lease capacity from each other and nobody advertises it, so a conduit-level audit is the only real check and it usually finds at least one shared segment. Frequency and space diversity on a microwave hop are cheap and genuinely move availability from about 99.9% to 99.999% (8.8 hours of outage a year down to 5 minutes), but they address fading and do nothing about a backhoe or a substation. Power is the constraint that appears in every multi-day event: a typical cell site holds 4 to 8 hours of battery and only a minority of sites have a generator, so a regional power failure produces a coverage map that degrades hour by hour no matter how the radios were engineered. Anti-jam techniques do work, and adaptive nulling with an antenna array can put 20–40 dB of rejection on a small number of jammers, but they cost an array and a custom waveform and are the wrong purchase for a network whose incidents are 60% system failures. The other honest weakness is detection, since an operator cannot fix interference it cannot find, and direction-finding an intermittent source across a metro area takes days of a specialist's time.

If your availability target is 99.99% or better, buy path diversity before anything else, and make the vendor show you the conduit route rather than a logical diagram. If the link is a microwave hop, add space or frequency diversity, which is the cheapest availability on this sheet. If the site has to survive a multi-day power failure, size the battery for the local restoration time and add a generator at the sites carrying the most traffic, because that is where the outage hours accumulate. If a fixed link is being jammed deliberately, changing medium usually beats hardening the waveform: fiber cannot be jammed, and a narrow-beam link at 70/80 GHz is hard to sit in front of. Reserve anti-jam waveforms and nulling arrays for links that have to work while somebody is actively trying to stop them, which on a commercial network is almost never, and see the counter-UAS and electronic warfare sheet for that discipline. If GNSS is the timing source, the resilience question is holdover, which this sheet's timing entry covers.

188 major telecom incidents reported to ENISA for 2024: 60% system failures, 19% human error, 13% natural phenomena, 8% malicious · cable cuts the largest single technical cause at 41 incidents, 23% of the total · faulty software changes cost 515 million user-hours in 2024 against 331 million for cable cuts · 99.9% availability is 8.8 hours of outage a year, 99.999% is 5 minutes · a typical cell site holds 4–8 hours of battery · adaptive nulling gives roughly 20–40 dB of rejection against a small number of jammers.

ENISA's annual telecom incident reporting and the FCC's Network Outage Reporting System, the two public datasets on why networks actually fail; the US 811 call-before-you-dig system; the Balticconnector damage in October 2023 and the C-Lion1 cut in November 2024, both attributed to dragged anchors; the Rogers outage in July 2022, which removed service from about 12 million Canadian subscribers after a configuration change; FCC enforcement against sellers of consumer jammers, which are illegal to market, sell or operate in the United States.

Resilience is paid for continuously and used rarely, which is why it is chronically underbought. A second diverse path roughly doubles the circuit cost and delivers nothing on the days the first one works, and the contract does not make up the difference: a typical carrier service-level agreement pays credits as a percentage of one month's fee, so a full day of outage returns a few percent of a monthly bill against a loss that is often much larger. That asymmetry is why buyers who genuinely need availability buy a second path rather than negotiate a stronger agreement. On the operator's side the spending is physical, meaning batteries and generators that run into the tens of thousands of dollars per site installed and then need fuel and testing, plus route engineering that costs planning time rather than equipment. The cheapest intervention is administrative: locate services such as 811 are free to the excavator and funded by the facility owners, because a cut costs far more than a mark-out. ENISA's 2024 data puts 331 million user-hours on cable cuts and 515 million on faulty software changes, which makes change control the second cheapest thing to buy. Anti-jam hardware sits at the far end of that scale, and for a commercial operator the money goes further against the two causes behind four out of five incidents.

4 minNetwork Redundancy - CompTIA Network+ N10-009 - 3.3Professor Messer · 50k+ views

3 minWhat is RF or Wi-Fi Interference? With Bob Young, Founder of FIFO NetworksNuts About Nets · 10k+ views

Telecom Security Incidents 2024 (European Union Agency for Cybersecurity) · Jammer Enforcement (Federal Communications Commission)

Spectrum reaches a user through one of three regimes, and they are three different businesses rather than three prices for the same thing. An exclusive license is a transferable right to transmit in a band over a geography for a term, usually 10 to 15 years with a strong expectation of renewal, carrying interference protection with legal force and buildout conditions that typically require covering a set share of the population within six years and more by twelve. A shared regime keeps the band open to several classes of user at once and arbitrates between them with a database: in the US CBRS band at 3550–3700 MHz, federal radar and satellite earth stations sit in tier one, 70 MHz of auctioned Priority Access Licenses in ten 10 MHz channels sit in tier two, and free General Authorized Access users take the remaining 80 MHz plus whatever priority spectrum is unused, with a Spectrum Access System assigning channels and coastal sensors telling it to clear the band within roughly five minutes when naval radar appears. Unlicensed spectrum has no license at all, only power limits and coexistence rules, and it is where Wi-Fi lives; the FCC opened 1,200 MHz at 5.925–7.125 GHz on that basis in 2020, with an automated coordination system protecting the fixed microwave links already there. Sitting above all three, the ITU Radio Regulations are a treaty that divides the spectrum by service and Region, and they are revised at a World Radiocommunication Conference held every three to four years. WRC-23 ran in Dubai from 20 November to 15 December 2023, and a conference of that kind is where a band gets identified for a use, not where any operator gets a license.

Exclusive licensing produces the one thing a capital-intensive network needs, which is a noise floor an engineer can plan against and a bank can lend against. The cost is that the money is spent years before the first customer and the supply cannot be increased on any useful timescale, because creating a new band means moving whoever is in it. Shared licensing is far cheaper, with CBRS priority licenses clearing around $0.21 per MHz-POP against $0.68 to $2.72 for exclusive mid-band, and it let enterprises and small operators hold spectrum for the first time, but the right is conditional: a general-access user can be displaced with no compensation and no notice beyond the database message, and the whole regime depends on sensors and a database staying correct. Unlicensed spectrum costs nothing to acquire and carries the largest device ecosystem in the world, and in exchange the capacity available at any moment depends on who else shows up, which is a hope rather than a specification. Harmonization through the ITU is what makes any of it cheap, since a band allocated the same way in most countries gets one chipset instead of one per market, and a band identified for mobile in only one Region gets devices late and expensive. The weakness of the treaty process is speed: a WRC decision starts a national proceeding that then takes years on its own.

If your business needs a guaranteed noise floor and you can raise capital years ahead of revenue, buy an exclusive license and treat it as the largest line item in the plan. If you need spectrum at one site or across a few counties, buy a CBRS priority license, and use general authorized access only where being moved off a channel is an inconvenience rather than an outage. If you control the building and are willing to own the interference risk yourself, use unlicensed spectrum at 5 and 6 GHz, which is the cheapest capacity available anywhere. If the plan depends on a band being reallocated, assume a decade and assume the incumbents get paid, because that is what every recent clearing has cost. If you need spectrum in orbit rather than on the ground, the mechanism is an ITU filing through an administration rather than a national auction, and this sheet's satellite spectrum and orbital filings entry covers it.

CBRS spans 3550–3700 MHz in three tiers: 70 MHz of priority licenses in 10 MHz channels, 80 MHz of general access · CBRS priority licenses cleared about $0.21 per MHz-POP against $0.68–2.72 for exclusive mid-band · 1,200 MHz opened for unlicensed use at 5.925–7.125 GHz in 2020 · exclusive licenses run 10 to 15 years with buildout conditions at six and twelve years · a World Radiocommunication Conference is held every three to four years, WRC-23 in Dubai from 20 November to 15 December 2023 · US incumbents were paid $10.1 billion to leave 600 MHz and $9.7 billion to clear C-band early.

The CBRS three-tier band, with Spectrum Access Systems operated by Google, Federated Wireless, CommScope and Sony, and the 20,625 county-level priority licenses sold in FCC Auction 105; the FCC's 2020 order opening 1,200 MHz at 6 GHz for unlicensed use with automated frequency coordination; WRC-23, which identified parts of the 6 GHz band for mobile broadband in Region 1 covering Europe, Africa and the Middle East, the same spectrum the United States had already made unlicensed; the C-band and 600 MHz clearings in the United States, which are the two reference cases for what moving an incumbent costs.

Spectrum is an asset with no maintenance cost and no ordinary depreciation, which is why carriers borrow against it and why the licenses can be worth more than the network built on them: Verizon carries roughly $150 billion of licenses against about $110 billion of net plant. That value is created by exclusion, so the regulator's real product is the clearing, and clearing is expensive because the incumbent has to be paid to move and has every reason to be slow. US broadcasters received $10.1 billion to vacate 600 MHz and then a 39-month repack schedule to actually do it; satellite operators received $9.7 billion in accelerated relocation payments for C-band on top of the auction price, in a process that opened in 2018, auctioned in early 2021 and finished clearing at the end of 2023. Federal bands are harder still, because the incumbent is an agency with a mission rather than a balance sheet, and money is not what it wants. Against that, shared and unlicensed spectrum are nearly free to acquire, and the costs move to operations: a CBRS deployment pays a spectrum access system a small recurring fee per device, and an unlicensed deployment pays in congestion it cannot control. For an operator the practical consequence is that the three regimes are three different capital structures, and the choice usually follows how much certainty the revenue needs rather than how much bandwidth the radio can use.

7 minIntroduction to CBRS and Spectrum SharingGoogle's Spectrum Access System · 5k+ views

12 minSpectrum auctions, spectrum liberalization and the incomplete triumph of CoaseMarginal Revolution University · under 1k views

3.5 GHz Band Overview (Federal Communications Commission) · Provisional Final Acts WRC-23 (International Telecommunication Union)