Why light?
Almost every byte that crosses an ocean, a continent or a data centre travels as light in glass. The reason is bandwidth. A radio link of Wireless systems works with carriers of a few gigahertz; light in the telecom band oscillates at about 193 THz, and the low-loss window of silica fibre is tens of terahertz wide — more spectrum than all radio services combined, in a strand thinner than a hair, immune to electrical interference. In 1966 Charles Kao and George Hockham, at Standard Telecommunication Laboratories in England, argued that glass could be made pure enough to carry light for kilometres (Kao and Hockham); Corning made such fibre in 1970, and Kao shared the 2009 Nobel Prize in Physics.
Photonics is now changing again, because of AI. The clusters of Hardware for AI join tens of thousands of accelerators with millions of optical links, whose power and cost have become a design constraint, and optics is moving from the front panel of a switch into the package beside the silicon. This lesson goes from the devices that make and detect light, through fibre and its limits, to the optical networks of 2026, lidar, and the long-promised idea of computing with light.
Making and detecting light
Definition— Semiconductor laser
A laser diode is a forward-biased p–n junction, in a material with a direct band gap, built so that electrons and holes recombine by emitting photons, and placed inside an optical resonator so that stimulated emission builds a coherent beam once the injected current exceeds a threshold. The photon energy is close to the band gap: GaAs-based lasers emit near 850–980 nm, InP-based lasers at 1,300–1,600 nm, the telecom bands.
In Semiconductors and diodes you met the band gap. Silicon's gap is indirect: an electron at the bottom of its conduction band cannot drop straight into a hole without also giving momentum to the crystal, so silicon emits light very weakly. That is why lasers are made of III–V compounds such as GaAs and InP, and why "silicon photonics" borrows its light from them. Modern lasers use thin quantum-well layers that confine electrons and holes together, an idea built on the heterostructures for which Zhores Alferov and Herbert Kroemer shared the 2000 Nobel Prize. Three kinds dominate communications:
- Distributed-feedback (DFB) lasers: an edge-emitting InP laser with a grating along its length that selects a single, stable wavelength — the workhorse of single-mode fibre links.
- Vertical-cavity surface-emitting lasers (VCSELs): tiny lasers that emit from the wafer surface at around 850 nm, cheap to make and test in arrays, used with multimode fibre over tens of metres (and in phone face-recognition sensors).
- Tunable lasers with narrow linewidth, for coherent systems, where the laser's phase carries information.
A photodiode works the other way: a reverse-biased junction (often p–i–n, with a wide intrinsic region) absorbs photons and each creates an electron–hole pair, so the current is proportional to optical power. Its responsivity is amperes per watt, where is the quantum efficiency — about 1.25 A/W at 1,550 nm for a perfect detector. InGaAs and germanium absorb at telecom wavelengths; avalanche photodiodes add internal gain for weak signals, such as lidar echoes.
Note
Communication and lidar lasers are mostly infrared and invisible, and the eye focuses them onto the retina, so never look into a laser, a transceiver port or the end of an optical fibre. Fibre ends are inspected with a fibre-inspection camera after confirming the light source is off. Products for the public are classified under the IEC 60825-1 standard; only Class 1 (safe under normal use) and Class 2 (low-power visible) lasers are suitable for experiments outside a laser laboratory.
Fibre: attenuation and dispersion
A single-mode fibre is a glass core about 9 µm across, with a slightly higher refractive index than the 125 µm cladding around it, so light is guided by total internal reflection (more exactly, as one guided mode of a dielectric waveguide, a cousin of the transmission lines of Transmission lines). Two properties limit it.
Attenuation. Silica's loss is lowest near 1,550 nm, about 0.2 dB/km in standard fibre (ITU-T G.652): after 100 km, 1% of the light is left. At shorter wavelengths, Rayleigh scattering from frozen-in density fluctuations, which grows as , dominates; at longer ones, infrared absorption by the glass itself. The best silica fibres have reached about 0.14 dB/km, a figure stuck for decades. In 2025 a team from the University of Southampton and Microsoft reported a hollow-core fibre, guiding light in air through a structure of nested glass membranes, with loss below 0.1 dB/km (Petrovich et al.); light also travels about 1.5 times faster in air than in glass, cutting latency, and Microsoft has begun deploying such fibre in its network.
Dispersion. Different frequencies travel at slightly different speeds, so a short pulse, which contains a spread of frequencies, spreads out. In standard fibre the chromatic dispersion parameter is about ps per nanometre of spectral width per kilometre at 1,550 nm, and close to zero near 1,310 nm. Mathematically, each angular frequency of the pulse's envelope acquires an extra phase after a distance , where . That is a statement about the Fourier transform of the pulse, so the simulation is a Fourier transform, a multiplication and an inverse transform.
import numpy as np
# Chromatic dispersion of standard single-mode fibre at 1550 nm.
c = 2.99792458e8
lam = 1550e-9
D = 17e-6 # 17 ps/(nm km) in s/m^2
beta2 = -D * lam**2 / (2 * np.pi * c) # group-velocity dispersion, s^2/m
print(f"beta2 = {beta2 * 1e27:.1f} ps^2/km")
# A transform-limited Gaussian pulse, field exp(-t^2 / 2 T0^2), sampled on a fine time grid.
T0 = 20e-12 # 20 ps: roughly a symbol at 25 GBd
N, span = 2**14, 4e-9
t = (np.arange(N) - N // 2) * (span / N)
w = 2 * np.pi * np.fft.fftfreq(N, d=span / N)
A0 = np.exp(-t**2 / (2 * T0**2))
def propagate(A, z):
"""Linear propagation: each frequency picks up the phase beta2 * w^2 * z / 2."""
return np.fft.ifft(np.fft.fft(A) * np.exp(0.5j * beta2 * w**2 * z))
def rms_width(A):
p = np.abs(A)**2
p = p / p.sum()
mean = (t * p).sum()
return np.sqrt(((t - mean)**2 * p).sum())
LD = T0**2 / abs(beta2)
print(f"dispersion length L_D = T0^2/|beta2| = {LD / 1e3:.1f} km\n")
print(" distance simulated width theory sqrt(1+(z/L_D)^2) peak power")
w0 = rms_width(A0)
for km in (0, 10, 20, 50, 100):
A = propagate(A0, km * 1e3)
print(f" {km:5d} km x{rms_width(A) / w0:6.2f} x{np.sqrt(1 + (km * 1e3 / LD)**2):6.2f}"
f" {np.max(np.abs(A))**2:5.2f}")
# A coherent receiver undoes dispersion digitally: multiply by the opposite phase.
A100 = propagate(A0, 100e3)
recovered = propagate(A100, -100e3)
print(f"\nafter 100 km and digital compensation, max error = {np.max(np.abs(recovered - A0)):.1e}")A 20 ps pulse — about one symbol at 25 gigabaud — has a dispersion length of 18.4 km: after 50 km it is nearly three times wider and after 100 km five and a half times, smeared across its neighbours. The simulated widths match the analytic formula to the printed precision. The last line is the most important. Dispersion is linear and lossless — it only rotates phases — so it can be undone exactly by applying the opposite phase. Old systems did that with spools of special fibre of opposite dispersion; a coherent receiver does it in its digital signal processor, which is why modern long-haul links need no optical dispersion compensation at all.
Many colours, amplified: WDM and coherent optics
Wavelength-division multiplexing (WDM) sends many independent channels down one fibre, each on its own wavelength, like radio stations sharing the air. Dense WDM channels sit on a grid of frequencies defined by the ITU (G.694.1), traditionally 50 or 100 GHz apart, now in flexible slots of 12.5 GHz granularity. The C band (about 1,530–1,565 nm) holds about 4.8 THz; adding the L band roughly doubles it.
What made WDM practical was the erbium-doped fibre amplifier (EDFA), demonstrated in 1987 at the University of Southampton and at Bell Labs: a few metres of fibre doped with erbium ions, pumped by a 980 nm laser, amplify every channel across the C band at once, with no electronics. A long-haul link is a chain of spans of about 80–100 km, each followed by an amplifier.
Each amplifier adds amplified spontaneous emission noise, so the optical signal-to-noise ratio (OSNR) falls along the link, by 3 dB every time the number of spans doubles. Meanwhile, modulation moved far beyond switching the laser on and off. A coherent transmitter modulates the amplitude and phase of the light in two orthogonal polarisations — QAM constellations like those of Modulation, on an optical carrier — and a coherent receiver mixes the incoming light with a local laser to recover the full field, then uses digital signal processing to undo dispersion, track the polarisation as it wanders and correct the phase. As of 2026, single coherent wavelengths carry about 800 Gb/s to 1.6 Tb/s over metro distances, at symbol rates above 100 gigabaud, and pluggable modules built to the OIF's 400ZR and 800ZR agreements (OIF) link data centres tens of kilometres apart. The program builds a link budget and asks how much information Shannon's theorem from Information and coding allows.
import math
h, c = 6.62607015e-34, 2.99792458e8
nu = c / 1550e-9 # optical frequency, about 193 THz
def dbm(watts):
return 10 * math.log10(watts / 1e-3)
# 1. A photodiode turns photons into electrons: responsivity R = eta * q * lambda / (h c).
q = 1.602176634e-19
print(f"ideal responsivity at 1550 nm: {q / (h * nu):.2f} A/W (a real Ge or InGaAs diode: 0.8-1.0 A/W)")
# 2. Amplified spontaneous emission noise of an EDFA in the standard 0.1 nm (12.5 GHz) bandwidth.
ref_bw = 12.5e9
const = -dbm(h * nu * ref_bw)
print(f"h*nu*12.5 GHz = {dbm(h * nu * ref_bw):.1f} dBm, so OSNR = {const:.1f} + P_ch - L_span - NF - 10 log10(N_spans)\n")
def osnr_db(p_ch_dbm, span_km, n_spans, loss_db_km=0.2, nf_db=5.0, extra_db=1.0):
span_loss = span_km * loss_db_km + extra_db # fibre plus connectors and splices
return const + p_ch_dbm - span_loss - nf_db - 10 * math.log10(n_spans)
# 3. A dense WDM system: 64 channels at 75 GHz spacing filling the C band, 80 km spans.
baud, spacing, channels, p_ch = 64e9, 75e9, 64, 1.0 # 1 dBm per channel
print(f"C band: {channels} channels x {spacing / 1e9:.0f} GHz = {channels * spacing / 1e12:.1f} THz")
print(" distance spans OSNR (0.1 nm) SNR in signal band Shannon limit per channel whole fibre")
for spans in (1, 5, 25, 75):
o = osnr_db(p_ch, 80, spans)
snr = 10 ** (o / 10) * ref_bw / baud # noise counted in the signal's own bandwidth
cap = 2 * baud * math.log2(1 + snr) # two polarisations
print(f" {spans * 80:6d} km {spans:4d} {o:5.1f} dB {10 * math.log10(snr):5.1f} dB "
f"{cap / 1e9:5.0f} Gb/s {channels * cap / 1e12:5.1f} Tb/s")The constant 58 dB is just the noise of one photon per second per hertz, , in the 0.1 nm reference bandwidth, expressed in dBm. With 1 dBm per channel and 80 km spans, the OSNR falls from 37 dB after one span to about 18 dB after 75 spans (6,000 km, an ocean crossing), and the Shannon limit per 64-gigabaud channel falls from about 1.3 Tb/s to under 0.5 Tb/s. Over the whole C band that is tens of terabits per second per fibre, which is why submarine cables carry many fibre pairs and why research has turned to space-division multiplexing: fibres with several cores, or several modes, in one strand.
Optics in the data centre: silicon photonics and co-packaged optics
Inside a data centre the problem is different: distances of metres to a couple of kilometres, millions of links, and cost and power per bit above all. Most links use intensity modulation with direct detection, now with four-level pulse-amplitude modulation (PAM4) at 100 or 200 Gb/s per lane. A pluggable transceiver at the front of a switch or network card carries 8 lanes: 800 Gb/s modules were mainstream in AI clusters in 2024–25, and 1.6 Tb/s modules began shipping in volume in 2025–26. Each such module draws of order 15–30 W, much of it in the electrical circuits that drive signals across the circuit board between the switch chip and the module, and a large AI cluster contains hundreds of thousands of them. Google's data centres also use optical circuit switches, with arrays of steerable micro-mirrors, to reconfigure the network itself (Urata et al.).
Silicon photonics builds the optical parts of a transceiver with the tools of CMOS manufacturing: silicon waveguides (silicon is transparent at 1,310 and 1,550 nm), modulators that shift the phase of light by moving charge in silicon (in Mach–Zehnder interferometers or tiny ring resonators), germanium photodetectors grown on silicon, and lasers of InP either bonded onto the wafer or supplied from a separate module through fibre. Foundries including TSMC, GlobalFoundries, Tower Semiconductor and Intel offer silicon-photonics processes.
Co-packaged optics (CPO) moves the optical engines into the switch package, removing the long electrical path and most of the power it costs. Broadcom showed CPO Ethernet switches in 2024 and NVIDIA announced CPO InfiniBand and Ethernet switches in 2025, shipping from late 2025 into 2026 (Broadcom; NVIDIA). The arguments against are serviceability (a failed laser can no longer be swapped at the front panel, which is why external laser modules are favoured) and the need to change how switches are built and tested. A middle path, linear-drive pluggable optics, removes the module's own signal processing and relies on the switch chip. Beyond switches, several companies are developing optical I/O chiplets that would give accelerators optical links directly, for scale-up networks that span many racks.
Lidar: ranging with light
Lidar measures distance by timing light. A pulsed time-of-flight lidar sends nanosecond pulses and times the echoes with fast detectors — avalanche photodiodes or arrays of single-photon avalanche diodes (SPADs) — scanning the beam with spinning mirrors, MEMS mirrors or electronically, or flashing a whole scene at once. An FMCW (frequency-modulated continuous-wave) lidar sweeps its laser's frequency and mixes the returning light with the outgoing beam, as a radar does: the beat frequency gives the range, and the Doppler shift gives the target's speed directly.
import numpy as np
c = 2.99792458e8
# Pulsed time of flight: distance = c * round-trip time / 2.
for r in (10, 100, 250):
print(f"target at {r:3d} m: echo after {2 * r / c * 1e9:7.1f} ns")
print(f"1 ns of timing error = {c * 1e-9 / 2 * 100:.0f} cm of range error\n")
# FMCW: sweep the laser frequency by B over T; the echo, mixed with the outgoing light,
# beats at f = 2 R B / (c T), plus a Doppler shift 2 v / lambda from the target's motion.
B, T, lam = 1e9, 10e-6, 1550e-9
fs, n = 500e6, 5000 # sample the beat note for T = 10 microseconds
t = np.arange(n) / fs
R_true, v_true = 150.0, 30.0 # metres, metres per second (approaching)
f_range = 2 * R_true * B / (c * T)
f_doppler = 2 * v_true / lam
rng = np.random.default_rng(1)
beat_up = np.cos(2 * np.pi * (f_range - f_doppler) * t) + 0.5 * rng.normal(size=n) # up-chirp
beat_dn = np.cos(2 * np.pi * (f_range + f_doppler) * t) + 0.5 * rng.normal(size=n) # down-chirp
def peak(x):
spec = np.abs(np.fft.rfft(x * np.hanning(n)))
return np.fft.rfftfreq(n, 1 / fs)[np.argmax(spec)]
fu, fd = peak(beat_up), peak(beat_dn)
R_est = (fu + fd) / 2 * c * T / (2 * B)
v_est = (fd - fu) / 2 * lam / 2
print(f"beat frequencies: up {fu / 1e6:.1f} MHz, down {fd / 1e6:.1f} MHz")
print(f"estimated range {R_est:.2f} m (true {R_true}), speed {v_est:.2f} m/s (true {v_true})")
print(f"range resolution c/(2B) = {c / (2 * B) * 100:.0f} cm")A metre of range is 6.7 ns of round trip, so every nanosecond of timing error is 15 cm. In the FMCW case, the up-chirp and down-chirp beat notes (61.4 and 138.8 MHz) give the range as their average and the speed from their difference, here 150.05 m and 29.99 m/s from noisy samples. With a 1 GHz sweep the range resolution is 15 cm.
Wavelength is a safety and cost decision. Lidars at 905 nm use cheap silicon detectors, but the eye focuses that light, so pulse energy is strictly limited. Light at 1,550 nm is absorbed in the front of the eye before it reaches the retina, which allows far more power — and longer range — within eye-safety limits, at the cost of InGaAs detectors and fibre lasers. As of 2025–26, the largest volumes of automotive lidar are shipped by Chinese makers such as Hesai and RoboSense, alongside companies in the United States, Europe and Israel, and lidar is standard on robotaxis and on many cars sold in China. Lumenary Robotics covers what is done with the resulting point clouds in depth and point clouds and autonomous vehicles.
Computing with light: an honest assessment
Light is tempting for computing because some operations are free in optics: a lens performs a Fourier transform as light passes through it, and a mesh of interferometers performs a matrix–vector multiplication as light propagates. In 2017 a group at MIT demonstrated a small neural network on a programmable mesh of Mach–Zehnder interferometers (Shen et al.), and several companies have since built larger systems. In 2025 Lightmatter reported in Nature a photonic processor that ran standard networks such as ResNet and BERT (Ahmed et al.), and other groups in Europe, China and the United States have published optical inference chips.
The obstacles are as well understood as the promise:
- Conversion. Data arrive as electrical signals and must go back to them. Every input needs a digital-to-analogue converter and modulator, every output a photodetector and analogue-to-digital converter, and these electronics, not the optics, dominate the energy unless the matrix is large enough to amortise them.
- Precision. Optical computing is analogue. Noise, fabrication variation and thermal drift limit effective precision to a handful of bits unless calibrated constantly, while digital hardware has meanwhile moved to 4- and 8-bit formats with block scaling, narrowing the advantage.
- Memory and nonlinearity. There is no practical optical memory, and nonlinear functions are hard to do optically, so data must return to electronics between layers.
- Size. Optical components are microns to millimetres across, far larger than transistors, which limits how large a matrix fits on a chip.
As of 2026, photonics has unambiguously won the job of moving data, from oceans down to the package, and that is where most of the industry's investment goes; several former photonic-computing companies have turned to interconnect. Whether optical computing will beat digital accelerators on energy per useful operation, at equal accuracy, for mainstream workloads remains unproven. Quantum computing with photons, a different idea, appears in Quantum hardware engineering.
Common misconceptions
Misconception
"Fibre is fast because light is fast." Light in glass travels at about two-thirds of its speed in vacuum, slower than radio through air, which is why some financial firms use microwave links between exchanges. Fibre's advantage is bandwidth and low loss, not speed; hollow-core fibre recovers some of the latency.
Misconception
"Dispersion destroys information." Chromatic dispersion only rearranges phases, so a coherent receiver can undo it exactly in software. What does destroy information is noise, and the nonlinear mixing of channels at high power.
Misconception
"Silicon photonics means silicon lasers." Silicon is a poor light emitter. Silicon photonics makes waveguides, modulators and (with germanium) detectors in silicon, and takes its light from III–V lasers bonded to the chip or fed in by fibre.
Exercises
1. Photonics
Warm-upHumanAn unamplified single-mode link launches 2 dBm. Its receiver needs at least −24 dBm. Connectors and splices lose 1.5 dB in total, the designer keeps a 3 dB safety margin, and the fibre loses 0.25 dB/km. What is the longest link, in whole kilometres, that meets the budget?
2. Photonics
Warm-upHumanA directly modulated laser has a spectral width of 0.1 nm. Its light travels 40 km of standard single-mode fibre with dispersion ps/(nm·km). By roughly how many picoseconds do the fastest and slowest parts of its spectrum separate?
3. Photonics
CoreHumanA photodiode with quantum efficiency 0.8 receives −10 dBm of light at 1,310 nm. What photocurrent does it produce, in microamperes, to the nearest whole number? Use with C, J s and m/s.
4. Photonics
CoreComputationalWrite osnr_db(p_ch_dbm, span_loss_db, nf_db, n_spans, wavelength_nm=1550) for a chain of identical amplified spans: , where with and GHz. Then write max_spans(required_db, p_ch_dbm, span_loss_db, nf_db), the largest number of spans whose OSNR is still at least required_db (0 if even one span fails). Use and .
5. Photonics
CoreComputationalAn FMCW lidar sweeps its frequency by B hertz in T seconds, up and then down. A target at range moving towards the lidar at speed produces beat frequencies and , where and . Write fmcw_range_speed(f_up, f_down, B, T, wavelength) returning (R, v) in metres and metres per second ( m/s).
6. Photonics
CoreHumanWhy do modern long-haul coherent systems not need spools of dispersion-compensating fibre?
7. Photonics
CoreHumanA start-up says its photonic chip multiplies a matrix by a vector "at the speed of light with almost no energy". Which response reflects the state of the field in 2026 most fairly?
Further reading and resources
- Dielectric-fibre surface waveguides for optical frequencies · K. C. Kao and G. A. Hockham · 1966The paper that proposed glass fibre for communication; Kao shared the 2009 Nobel Prize in Physics for it.
- The Nobel Prize in Physics 2009
- ITU-T G.652: Characteristics of a single-mode optical fibre and cable · International Telecommunication Union
- ITU-T G.694.1: Spectral grids for WDM applications: DWDM frequency grid · International Telecommunication Union
- Optical Internetworking Forum (OIF)Publishes implementation agreements such as 400ZR and 800ZR for interoperable coherent modules.
- Broadband optical fibre with an attenuation lower than 0.1 decibel per kilometre · M. Petrovich et al. · 2025
- Mission Apollo: Landing Optical Circuit Switching at Datacenter Scale · R. Urata et al. (Google) · 2022
- Co-packaged optics · Broadcom
- Silicon photonics networking · NVIDIA
- Deep learning with coherent nanophotonic circuits · Y. Shen et al. · 2017
- Universal photonic artificial intelligence acceleration · S. R. Ahmed et al. (Lightmatter) · 2025