1. Introduction

A single-mode fiber pair carrying 32 carriers of 800 Gb/s across the extended C-band delivers 25.6 Tb/s, and every number in that sentence is a design decision rather than a property of the glass. The 32 comes from dividing 4.80 THz of usable spectrum by a 150 GHz slot. The 800 Gb/s comes from 118 GBd multiplied by eight bits per symbol and divided by the framing and forward error correction overhead. The slot comes from a symbol rate and a roll-off factor. Change the amplifier band coverage, the symbol rate or the modulation format and the whole figure moves, which is why two capacity numbers quoted without their conditions are rarely comparable.

This article covers the capacity of a wavelength-division-multiplexed (WDM) link on standard single-mode fiber: the quantities that define it, the arithmetic that produces it, the physical mechanisms that bound it, and the conditions attached to every published figure. Multi-core and mode-division fiber, passive optical network capacity, and packet-layer throughput sit outside that boundary and are not treated here.

One reference case threads the whole article: a 118 GBd dual-polarization 16-state quadrature amplitude modulation (DP-16QAM) carrier delivering 800 Gb/s in a 150 GHz slot, 32 of them across the extended C-band, over a 480 km regional route built from six spans of 80 km. Every worked value below belongs to that case, and the interactive sandboxes open on it, so the number in the prose is the number on screen.

2. Usable Spectrum, Occupied Bandwidth and Capacity: Definitions

Usable optical spectrum is the contiguous frequency range a line system delivers end to end at its specified gain, flatness and noise figure, measured in terahertz. Fiber capacity is the total information rate that spectrum carries, in bits per second, once every carrier placed inside it has been assigned a modulation format, a symbol rate and a forward error correction (FEC) overhead. Usable spectrum is a property of the line system; capacity is a property of what is provisioned into it.

Adjacent Quantities and Their Distinctions

Four pairs of adjacent quantities account for nearly all confusion about fiber capacity, and each pair separates on a single property.

Occupied bandwidth against slot width. Occupied bandwidth is the spectrum the carrier's own signal fills, equal to the symbol rate multiplied by one plus the roll-off factor. Slot width is the block of spectrum the network reserves for that carrier, which must be a whole number of 12.5 GHz steps on the flexible grid of ITU-T G.694.1 and is therefore always at least a little wider. The difference between them is the guard band.

Slot width against channel spacing. On a fixed grid the two are the same number because every channel gets the same block. On a flexible grid they separate: a route can carry a 137.5 GHz slot next to a 75 GHz slot, and the spacing between two carrier centres is then the mean of their two half-widths, not a grid constant.

Noise bandwidth against occupied bandwidth. The bandwidth that sets capacity and keys the OSNR conversion is the symbol rate, because a matched root-raised-cosine receive filter has a noise bandwidth of exactly the symbol rate whatever the roll-off. Roll-off widens the occupied bandwidth and therefore the slot; it does not enter the capacity term or the OSNR conversion. Mixing the two is the most common arithmetic error in a capacity estimate.

Band definition against usable spectrum. The ITU-T band letters name wavelength ranges. What a line system delivers is a narrower or wider frequency range set by the amplifiers, the gain-flattening filters and the wavelength selective switches along the route. Section 4 gives the two numbers that most often get swapped.

Units and the Conversion Arithmetic

Applying that to the reference carrier: 118 GBd at a roll-off of 0.10 occupies 129.8 GHz, so the smallest flexible-grid slot that holds it is 12.5 × 11 = 137.5 GHz. Route designs that cross several reconfigurable optical add-drop multiplexer (ROADM) nodes usually assign the next step up again, 150.0 GHz, because each wavelength selective switch narrows the passband a little and the extra 12.5 GHz absorbs that narrowing. The guard band is then 20.2 GHz, about 10.1 GHz either side of the carrier.

3. Capacity Arithmetic from One Symbol to One Fiber

Capacity on a channel limited by additive Gaussian noise follows Shannon's relation, and the form that applies to a coherent optical carrier carries every physical variable explicitly.

Dividing through by the symbol rate gives the per-symbol form, 2 · log₂(1 + SNR) bits per symbol, which is the number to compare against a modulation format. It is derived from the complete relation above and should never be presented on its own as if the bandwidth had disappeared, because bandwidth is exactly what a spectrum plan provides. The MapYourTech treatment of Shannon's limit for fiber optics works through the same relation with a different entry point.

Rate Arithmetic for One Carrier

Three quantities separate the number a modem advertises from the number the client sees. Gross line rate is the symbol rate multiplied by bits per symbol. Net client rate is the gross rate divided by one plus the framing and FEC overhead. Slot spectral efficiency is the net rate divided by the assigned slot width, in bits per second per hertz.

For the reference carrier: 118 GBd × 8 bits per symbol gives 944 Gb/s gross, which carries 800 Gb/s of client traffic once framing and FEC take their 18%. In a 150 GHz slot that is 5.33 b/s/Hz. The Shannon bound at the SNR this carrier receives allows 10.4 bits per symbol, so DP-16QAM at 8 bits per symbol sits inside the bound with room, and the gap is what FEC and implementation cost rather than a physical shortfall. The FEC families used on coherent interfaces set that gap.

From One Carrier to One Fiber

Fill capacity is the arithmetic a planner runs first: divide the usable spectrum by the slot width, take the whole number of slots, and multiply by the net rate per carrier.

Fill capacity is an upper bound on a fully lit fiber and not a route capacity. The carriers at the band edges reach the receiver with a different OSNR from the ones at the centre, because amplifier gain is not perfectly flat and stimulated Raman scattering transfers power from the shorter wavelengths to the longer ones as the fiber fills. Section 8 separates fill figures from measured and trial figures for exactly this reason.

4. Optical Bands and Deliverable Line-System Spectrum

Silica fiber has low attenuation from roughly 1260 nm to 1675 nm, a window of about 60 THz, and the ITU-T divides it into six lettered bands. Attenuation is only part of what decides which of them carries traffic. The C-band dominates because erbium happens to provide flat, low-noise gain across it, and a band with an efficient amplifier carries traffic that a band with slightly lower loss and no amplifier cannot.

Amplifier Coverage as the Spectrum Constraint

Adding the L-band roughly doubles the usable spectrum, and it does so by adding a second amplifier stage with a longer erbium gain fiber, a band splitter and a band combiner at every site. Widening the C-band alone into a super C-band of about 6.1 THz keeps one amplifier stage and adds about 27% more spectrum. Widening both gives the super C plus super L plans that run between 11.6 and 12.2 THz depending on how far the L-band window is pushed. The MapYourTech analysis of C+L, extended C+L and super C+L spectrum carries the supplier-by-supplier position on each of those plans.

5. OSNR, GSNR and the Required-OSNR Threshold

Optical signal-to-noise ratio is the ratio of signal power to amplified spontaneous emission (ASE) power in a stated reference bandwidth, by convention 0.1 nm, which is 12.48 GHz at 1550 nm and is carried through the arithmetic as 12.5 GHz. A figure quoted without its reference bandwidth is not an OSNR. The OSNR fundamentals guide builds the quantity from the ASE physics; what matters for capacity is the relation between it and the SNR that appears in Shannon's expression.

At 118 GBd the conversion term is 10 · log₁₀(118 / 12.5) = 9.75 dB. A carrier that needs 14.8 dB of SNR therefore needs 24.5 dB of OSNR in 0.1 nm. Raising the symbol rate raises the required OSNR by the same 10 · log₁₀ factor for no change in the underlying receiver sensitivity, which is why a 200 GBd carrier and a 60 GBd carrier at the same modulation format quote required OSNR figures 5 dB apart.

Link OSNR from the Amplifier Chain

The cascade term is the part that binds long routes. Doubling the amplifier count costs 3 dB of receiver OSNR whatever else is done, so a route that runs 40 spans starts 16 dB behind a route that runs one. The full link-engineering formula reference sets out the multi-stage form where the amplifiers are not identical.

Required OSNR and the Implementation Gap

Required OSNR is the OSNR at which a given transceiver mode reaches its pre-FEC threshold. It has a theoretical floor and a practical value, and the two are separate quantities. The floor is the Shannon inverse: a format carrying m bits per symbol on two polarizations needs at least 10 · log₁₀(2^m/2 − 1) dB of SNR. Real receivers sit above that floor by an implementation gap that covers FEC shortfall, transmitter and receiver noise, and DSP approximation. The table below applies a 3 dB gap, which is a modelling assumption rather than a measured value, so a supplier datasheet remains the number to design against.

6. Physical Limits on Fiber Capacity

Shannon's relation bounds a linear channel. Optical fiber is not linear at the power levels a long route needs, and the departure from linearity is what turns an unbounded capacity expression into a finite one. Raising launch power raises the signal above the ASE floor one decibel for one, but the Kerr effect generates nonlinear interference (NLI) whose power rises as the cube of the signal power. Relative to the signal, NLI therefore grows two decibels for every decibel of launch power, and somewhere between the two slopes there is a maximum.

Figure 3 uses the reference route's noise budget: ten spans of 80 km at 0.22 dB/km, a 5 dB noise figure, and a nonlinear coefficient chosen so the optimum falls at 0 dBm. The peak reaches 13.9 dB of GSNR. Two decibels either side of the optimum costs about 0.75 dB on the low side and about 1.0 dB on the high side, so a power plan that lands within a decibel of the optimum has given up very little and a plan that overshoots by four decibels has given up 3.7 dB.

Effective Length and the Location of Nonlinear Generation

Nonlinear interference is generated where the power is high, which is the first part of each span. Effective length quantifies it as (1 − e^−αL) / α, which for an 80 km span at 0.2 dB/km is about 21 km. The remaining 59 km contributes attenuation and almost no nonlinearity. Two consequences follow: shorter spans with lower per-span loss improve OSNR without proportionally increasing nonlinear interference, and distributed Raman gain, which lifts the signal where it is weakest rather than at the span input, improves the noise figure without adding much to the nonlinear term.

7. Modulation Order, Symbol Rate and Reach

Capacity per carrier rises along two axes and each has a different cost. Adding bits per symbol costs OSNR: every step from QPSK to 8QAM to 16QAM to 64QAM adds two bits and roughly 3 to 3.5 dB of required OSNR, which on a long route translates into a shorter unregenerated distance. Adding symbol rate costs spectrum: the occupied bandwidth and the slot grow in proportion, and the required OSNR rises by 10 · log₁₀ of the ratio, so a higher-baud carrier needs more OSNR for the same modulation format even though its receiver sensitivity in SNR terms is unchanged.

The two axes behave differently in a capacity plan. Doubling the symbol rate doubles the carrier rate and doubles the slot, leaving spectral efficiency unchanged and reducing the transponder count, which is why the current generation of coherent modems reached for baud rate rather than constellation order. Doubling the bits per symbol doubles the carrier rate in the same slot, raising spectral efficiency, and costs reach. The guide to optical modulation and constellation diagrams works through the constellation side of that exchange.

Probabilistic Constellation Shaping Gain

Probabilistic constellation shaping (PCS) transmits low-amplitude constellation points more often than high-amplitude ones, which moves the transmitted distribution closer to the Gaussian one Shannon's bound assumes. The theoretical gain is 1.53 dB for an ideal shaping of a uniform constellation; deployed implementations recover a part of that. Shaping also makes the carrier rate continuous rather than stepped, so a route can be provisioned at 650 Gb/s rather than being forced to choose between 600 and 800. That granularity is what makes a modern modem's rate ladder look continuous.

8. Published Capacity Figures and Their Conditions

Capacity figures circulate without their conditions more often than any other number in optical networking, and the range between an arithmetic fill figure and a laboratory record is more than an order of magnitude. Five natures of figure appear in trade coverage and supplier material, and each answers a different question.

Reading the table by nature rather than by size: the first two rows answer “how much fits in the spectrum”, the third answers “what does the physics allow on a long route”, the fourth and fifth answer “what will a supplier commit to in a contract”, the sixth answers “what does a multi-supplier interface guarantee”, and the last two answer “what has been shown once, under conditions that are stated in the paper”. A research result on 39 km of metropolitan fiber with a laboratory transmitter is a real measurement and tells a planner nothing about a 1,500 km route.

Commitment Status by Spectrum Plan

Comparisons between spectrum options are where status classes matter most, because an option that exists in a datasheet and an option carrying live traffic are separated by several years of amplifier and equalization engineering.

9. Model Boundaries and the Direction of Each Error

Three models produce the capacity figures above — the fill model, the 58 relation and the required-OSNR model — and each carries assumptions that push its result in a known direction.

The fill model divides usable spectrum by slot width. It assumes every slot is identical, every carrier reaches its threshold, and no spectrum is lost to an inter-band gap or to ASE loading of unfilled slots. All three assumptions are optimistic, so a fill figure is an upper bound. On a C+L plan the inter-band gap alone removes a few hundred gigahertz, and on a route where the edge channels fall short of their required OSNR the deliverable capacity is lower again.

The 58 relation assumes identical spans, identical amplifiers and no nonlinear interference. Real routes have unequal spans, and because the cascade term averages while the worst span sets the binding value, the uniform-span result is optimistic for a route with one long span and pessimistic for a route whose spans are shorter than the average used. It also reports OSNR rather than GSNR, so it overstates the margin available on any route running near its optimum launch power.

The required-OSNR model in Section 5 adds a flat 3 dB implementation gap to the Shannon floor. Real gaps vary with the format, the FEC and the silicon generation, and they are usually larger for high-order formats than for QPSK. The model is therefore optimistic at 64QAM and close to right at QPSK. It is a teaching model that shows the shape of the exchange between format and OSNR; a supplier's route model, run on the measured plant, is the number to contract against.

10. Capacity Sizing Procedure and a Worked Route

Sizing a route's capacity runs in a fixed order, because each step consumes the output of the one before it and reversing two of them produces a plan that does not close.

Practical Example — 480 km regional route, six spans, 800 Gb/s carriers

Margin classes matter in that arithmetic and are easy to leave out. Span margin covers measurement uncertainty in the as-built loss, repair margin covers the splices a future cable repair will add, and aging margin covers component drift over the system's life. The treatment of beginning-of-life and end-of-life margin design separates them and shows why a single lumped figure hides which one has been consumed. Mixed-rate routes add a further step, covered in the guide to mixed channel-rate planning on DWDM line systems.

11. Diagnosing a Route Below Its Capacity Target

A route that carries less than its design figure has usually lost the capacity in one of six ways, and the observation narrows the search before any instrument is connected.

Filter narrowing is commonly omitted at turn-up and accounts for several of these observations. Each wavelength selective switch a carrier passes through removes a little of the passband edge, and the effect accumulates: a carrier that closes across two ROADM nodes may not close across seven in the same slot. The ROADM architecture guide covers the node side; on the capacity side the response is either a wider slot for express carriers or a shorter node count in the path computation.

12. Capacity Growth Beyond the C+L Plan

Per-fiber capacity has three sources of growth and the remaining range of each differs by an order of magnitude. Bits per symbol is close to its limit: deployed systems operate within 1 to 2 dB of the linear Shannon bound, so the remaining gains arrive in fractions of a decibel from shaping, FEC and lower roll-off rather than from new constellation orders. Symbol rate has more range, and the current generation used it: commercial coherent modems moved from about 95 GBd to 118–140 GBd in pluggables and 200 GBd in embedded modems. Spectrum has the clearest arithmetic of the three, since 4.8 THz extended to 12 THz multiplies fill capacity by 2.5 with no change to the modem.

Deployment and Standards Status in 2026

1.6 Tb/s carriers. A 200 GBd modem carrying probabilistically shaped 64QAM reaches 1.6 Tb/s on a single carrier at metro ROADM distances (vendor claim). On the pluggable side, 1Finity, Ciena, Marvell and Nokia have announced 1.6 Tb/s ZR and ZR+ modules built on 2 nm coherent DSPs, with samples expected from late 2026 and general availability later (announced product, analyst reporting). The OIF 1600ZR work that would make those modules interoperable is expected to land after that.

800G interfaces in volume. The OIF 800ZR implementation agreement specifies an 800 Gb/s coherent line interface for single-span amplified DWDM links of 80 to 120 km, running DP-16QAM at about 118 GBd in a pluggable form factor (standard-specified, OIF 800ZR IA). Shipment volumes remain small against 400ZR, which has cumulative shipments in the millions of modules, so 800ZR is at the start of its ramp rather than in the middle of it (analyst reporting).

Ethernet client rates. IEEE P802.3dj adds 200 Gb/s per lane and 1.6 Tb/s media access control parameters, and sat at draft 2.4 in early 2026. Until it completes, 1.6 Tb/s client interfaces ship against pre-standard drafts and supplier interoperability testing.

Multiband and hollow-core fiber. Adding the S-band to C+L takes the usable window to about 15.6 THz, demonstrated at 202.3 Tb/s over 39 km of deployed metropolitan fiber using thulium-doped and erbium-doped amplifiers (research demonstration). Hollow-core fiber offers a much wider low-attenuation window — one supplier reports an 18 THz figure — along with far weaker nonlinearity and lower latency, and several groups ran 400G and 800G ZR modules over hollow-core cable at OFC 2026 (field trial). Neither is a line-system product today.

More fiber pairs. When one fiber is close to its bound, the next order of magnitude comes from lighting more of them. Multi-rail line systems that amplify several fiber pairs in one shelf are the response, and they are the subject of the MapYourTech analysis of how AI demand is reshaping transport hardware and of the practical treatment of spectral efficiency and where its remaining gains sit.

13. Conclusion

Fiber capacity resolves into four numbers and one exchange rate. The four numbers are the usable spectrum a line system delivers in terahertz, the slot width each carrier is assigned in gigahertz, the bits per symbol the received OSNR supports, and the symbol rate the modem runs. Multiply them correctly and a fiber pair's capacity falls out; the reference case in this article gives 25.6 Tb/s from 4.80 THz, 150 GHz slots, 8 bits per symbol and 118 GBd. The exchange rate is that about 3 dB of OSNR supports two more bits per symbol, and that each doubling of the amplifier chain returns 3 dB the other way.

Two habits carry more weight than any single formula. State frequency edges rather than band names, because the 9% difference between the ITU-T C-band and the extended C-band propagates into every downstream figure. And attach a status class to every capacity number encountered: arithmetic fill, modelled bound, vendor claim, standard specification, field trial or research result. A route plan built on figures of mixed status will close on paper and fail on the plant.