Light Sources for Co-Packaged Optics

Scope

This article covers the light-source side of co-packaged optics: the reasons shipping designs exclude the laser from the package, the standardized interfaces that make an external source practical, the arithmetic that decides how many lasers a switch needs, the two multi-wavelength source technologies competing for the role, and the wall-plug-to-modulator power accounting. The electronics side of the same story, where the DSP went and what that buys, is covered separately in Pluggable, LRO, LPO, CPO. Background: co-packaged optics, external laser source.

Why the laser moved out

Co-packaged optics is an exercise in shortening the electrical path between switch ASIC and modulator, and every centimeter of that shortening drags the optics into a hotter, less serviceable, more expensive neighborhood. For modulators and photodiodes the trade is acceptable. For lasers it fails three separate ways.

The first failure is thermal. A switch package lid runs tens of degrees hotter than a faceplate cage, and of all optical components the laser responds worst: threshold rises, wall-plug efficiency falls, wavelength walks off its WDM grid, and the Arrhenius clock on wear-out runs faster. The second is statistical. In the failure budget of an optical engine the laser dominates, and the units are not symmetric: a dead laser inside a co-packaged switch condemns the most valuable board in the rack to rework, while a dead laser in a pluggable shelf is a field swap. Qualification regimes of the Telcordia GR-468 kind exist precisely because laser reliability is the long pole, and no operator wants that pole planted inside a sealed package. The third is economic: separating source from engine decouples two immature yields, and lets an engine take light from whichever source technology matures first.

So the shipping architecture keeps the gain medium at arm's length: continuous-wave light is generated in a replaceable module and piped in over polarization-maintaining fiber, and the engine modulates what arrives. The laser becomes a consumable, in the way line cards and fans already are.

The interfaces that make it a product

Two standards turn this from a proprietary arrangement into a market. OIF's External Laser Small Form-Factor Pluggable implementation agreement (ELSFP, published 2023) defines the blind-mate pluggable module, its PM-fiber output, and its management, which is what lets a laser shelf be serviced like any other pluggable. The CW-WDM MSA defines the wavelength grids, 8, 16, and 32 lines in the O-band, so that a multi-wavelength source and the engine consuming it can be designed by different companies. Connector and cage hardware for ELSFP is already in vendor catalogs.

Counting lasers

How many lasers a port needs depends on the modulator architecture, and the difference is the headline number in vendor announcements. A conventional pluggable uses one directly modulated or externally modulated laser per lane: eight lasers for an eight-lane 1.6T module, each modulated and none shared. A split-and-modulate engine instead takes one continuous-wave source per wavelength, divides it through an on-chip splitter tree, and lets a microring modulator imprint each lane's data downstream. Lanes then share sources, and one sufficiently powerful DFB can feed several ports' worth of rings; this is the arithmetic behind Nvidia's claim of several-fold fewer lasers per port in its co-packaged switches. Fewer, but individually more powerful and more critical: losing one shared source now drops multiple lanes, which is why external shelves carry N+1 spares.

A comb source pushes the sharing to its limit: one device emits every wavelength on the grid at once, and the splitter tree is replaced by a demultiplexer. Whether that consolidation is progress depends entirely on the per-line power and the reliability of the one device now carrying the whole port.

The comb contenders

Two technologies plausibly supply a full CW-WDM grid from a single unit, and they fail in opposite directions.

The quantum dot comb laser is a semiconductor laser whose inhomogeneously broadened gain lets a single-section device mode-lock into tens of lines with low per-line intensity noise; the quantum dot laser entry covers the physics. Its virtues are those of any diode laser: direct electrical pumping, familiar packaging, commercial availability today. Its constraint is total power divided among lines, and line counts in the tens rather than hundreds.

The Kerr microcomb generates its grid by four-wave mixing in a high-Q microresonator, pumped by a single strong laser. Line counts come easily; the historic objections have been conversion efficiency, per-line power, and the laboratory apparatus once needed to keep a soliton alive. That objection list has been shrinking: integrated designs that combine the pump laser and the resonator on one chip, with turnkey soliton start-up, were demonstrated at commercial-scalability level by university groups in 2026, and normal-dispersion comb designs trade bandwidth for materially better efficiency. As of late 2026 quantum dot combs are the conservative bet and microcombs the higher-ceiling one, and engines designed to CW-WDM grids can accept either, which was the point of standardizing the grid.

The power accounting

The external arrangement pays optical tolls, and they are best presented as a worked chain. Suppose an eight-wavelength engine needs 1 mW (0 dBm) of continuous-wave power at each microring modulator, a representative figure. The light must first cross the blind-mate connector (about 0.5 dB), a PM patch (0.2 dB), and a fiber-to-chip coupler (1.5 dB for a good edge coupler), then an on-chip 1 by 8 distribution tree: 9.0 dB of unavoidable splitting plus perhaps 0.5 dB excess. The path totals 11.7 dB, so each wavelength leaves the shelf at about 15 mW. At a 15% source wall-plug efficiency the eight wavelengths cost roughly 0.8 W of electrical power, which against the engine's 1.6 Tb/s is about 0.5 pJ/bit spent on illumination alone, before any electronics. The chain explains the two obsessions of CPO source engineering: coupler loss, because every dB at the chip edge is paid by every wavelength, and source efficiency, because the multiplier on the whole chain is 1/WPE. It also explains why per-line power is the microcomb's entry fee: a comb whose lines leave the chip at a fraction of a milliwatt starts the same chain 10 dB behind.

Numbers of this kind are representative rather than universal; a design with lower ring drive requirements, better couplers, or higher-efficiency sources moves every figure, and published engine budgets differ by factors of a few in both directions. The structure of the accounting does not change.

What remains unsettled

Three questions stay open as of late 2026. Whether the industry converges on shared DFBs, quantum dot combs, or microcombs is a live competition, and the standardized grid means it need not resolve soon. Whether external sourcing survives the move of optics into the scale-up domain, where per-GPU optical I/O multiplies the port count again, is an economics question about connectors and fiber management as much as photonics. And the efficiency frontier, sources whose wall-plug efficiency and coupling recover the external path's tolls, is where the laser-diode results feeding this architecture, high-power SOAs and uncooled O-band sources among them, will matter most.

The Datacenter Link Budget Explorer puts the module-level power figures in fleet context, and the wall-plug efficiency calculator computes source efficiency from LIV data.

References: OIF, External Laser Small Form-Factor Pluggable (ELSFP) Implementation Agreement (2023); CW-WDM MSA specifications (cw-wdm.org); IEEE Spectrum coverage of Nvidia Quantum-X and Spectrum-X co-packaged switches (2025-2026); Norman, Jung, Wan & Bowers, APL Photonics 3, 030901 (2018) on quantum dot lasers on silicon; UCSB integrated microcomb reports (2026).