Four lessons of this module have quietly assumed that coherent light simply arrives. Time to face the platform's original sin: silicon cannot lase. The material that guides, splits, phase-shifts, modulates and even detects light is congenitally incapable of producing it — and no amount of engineering cleverness has repealed the verdict, because it is written into silicon's band structure. Every silicon photonic chip in the world therefore begins with a diplomatic problem: the gain must come from somewhere else — a family of III-V semiconductors grown on a different planet, fab-wise — and getting that foreign light into a silicon waveguide cheaply, efficiently and by the million is one of the defining engineering battles of the field. This lesson is about the physics of the verdict, the laser physics that the III-Vs supply, and the three strategies for marrying the two.

The indirect-gap verdict

Recall from band theory that electrons in a crystal carry a crystal momentum \hbar k, and both energy and momentum must balance in any transition. A photon carries plenty of energy but essentially zero crystal momentum (its k is a thousandth of the Brillouin zone's width). In a direct-gap material like GaAs or InP, the conduction-band minimum sits at the same k as the valence-band maximum, so an electron can fall straight down and pay out its energy as a photon: radiative lifetimes are nanoseconds. In silicon the conduction minimum is displaced most of the way to the zone boundary. A recombining electron must shed a large chunk of crystal momentum that the photon cannot carry, so a phonon must participate — a three-body event whose rarity (a second-order process, in the language of Fermi's golden rule) stretches silicon's radiative lifetime to milliseconds. In a real, impure, carrier-crowded crystal, non-radiative channels (defects, Auger recombination) and free-carrier absorption recycle the energy into heat in microseconds or less. The radiative path loses the race by six orders of magnitude: pump silicon as hard as you like, it glows feebly and lases never.

Marrying III-V gain to silicon

The gain media that do work — InP- and GaAs-family quantum wells — are grown on their own expensive, small-diameter wafers in their own fabs. Three strategies bring their light to silicon, in ascending order of intimacy:

• Hybrid integration — build a complete III-V laser chip, then attach it: butt-couple it at the silicon chip's edge, flip-chip solder it into a recess, or leave it in a separate package feeding fibre. Each laser is individually made, tested and aligned — sub-micron alignment, one device at a time — which is honest, flexible and expensive per unit.
• Heterogeneous integration — bond unpatterned slivers of III-V material face-down onto the processed silicon wafer, then lithographically fabricate the lasers in place, aligned to the waveguides beneath by the stepper's precision rather than by a robot's. The laser mode lives partly in the III-V (for gain) and partly in the silicon (for everything else). Wafer-scale economics; this is how Intel has shipped millions of on-chip lasers.
• Monolithic growth — grow the III-V directly on silicon. The lattice constants disagree by ~4–8%, seeding threading dislocations that kill conventional quantum-well lasers. The rising answer is quantum-dot gain: carriers trapped in a dot cannot wander to a dislocation, so QD lasers tolerate defect densities that would be fatal otherwise — the long-game bet for true all-silicon-wafer laser economics.

Whichever way the gain arrives, single-wavelength operation comes from a DFB (distributed feedback) structure: a grating etched along the cavity reflects only at the Bragg wavelength \lambda_B = 2\, n_{\text{eff}}\, \Lambda, so of all the modes the cavity could support, exactly one is fed back and wins. Worked number: for \lambda_B = 1550\ \mathrm{nm} and n_{\text{eff}} = 3.2, the grating pitch is \Lambda = 1550/(2 \times 3.2) \approx 242\ \mathrm{nm} — comfortably printable, and the reason WDM systems can trust each laser to hold its channel.

The wall-plug ledger

For computing, the laser is an overhead — a standing power bill paid before a single bit moves — so its wall-plug efficiency (WPE), optical watts out per electrical watts in, haunts every system-level energy claim in this course:

A typical integrated DFB emitting 10 mW from 50 mA at 1.5 V manages \eta_{\text{wp}} = 10/(50 \times 1.5) \approx 13\%; the state of the art reaches 20–30%, and the balance leaves as heat on the chip, right next to temperature-twitchy rings. Two structural reasons the number is stubborn: below the threshold current the cavity's gain has not yet overcome its loss, so the first tens of mA buy no coherent light at all; and above threshold each photon costs the junction voltage while returning only the photon energy. Run the ledger:

Note the shape of the output: nothing, nothing, then a straight line — the threshold behaviour. And note the last line: at 15% WPE, every optical watt a photonic accelerator "computes with" costs nearly seven electrical watts before the electronics have drawn a thing. Wall-plug efficiency is where photonic-computing energy claims go to be audited.

Where this goes next

One more idea completes the sourcing story: some systems need not one wavelength but dozens — WDM links and photonic accelerators consume lasers by the comb-full. A single microresonator, pumped hard enough to go nonlinear, can shatter one pump into a ruler of equally spaced wavelengths — a frequency comb, a whole WDM source in one ring, whose computing applications get their own lesson in Module 10. With light finally sourced on-chip, the module has a full parts bin: splitters, phase shifters, rings, modulators, detectors, lasers. The closing lesson asks how anyone actually builds a chip out of them — the photonic design flow, from PDK to packaged part.