An external laser source feeds continuous-wave light into a co-packaged optics engine from outside the switch package. Why the laser must leave, the RIN/linewidth/efficiency specs that decide winners, and the 80-90% margin at the laser die.
More Macroplane supply-chain guides · Baskets · Investment theses · Macro trends
An external laser source is a separate, continuously-on light source that sits outside a co-packaged optics module and feeds it laser light over fibre. It exists because a laser is the one part of an optical link that cannot survive next to a kilowatt-class switch chip. As co-packaged optics moves the optical engine onto the switch package, the laser moves the other way — out of the module entirely — and becomes a distinct product with its own supply chain, its own specifications, and margins that look nothing like a transceiver's.
This page explains what an ELS is, why the architecture demands one, which specifications separate a good one from a bad one, and who actually makes them.
👉 See it in context: the Photonics / CPO basket tracks the companies below, inside the Silicon Photonics & Optical Interconnects trend.
Strip an optical link down and it has four jobs: make light, encode data onto it, carry it, and detect it at the far end.
In a conventional pluggable transceiver, all four live inside one module you can pull out of a faceplate with your fingers. The laser is a millimetre-scale semiconductor chip a few centimetres from the modulator it feeds.
In a co-packaged optics design, jobs two and four move onto the switch package — the optical engine sits on the same substrate as the switch ASIC, millimetres from the silicon. Job one does not come with it. Instead:
You will see the same component called an external light source (ELS), a remote light source, or simply a CW laser source — the terms are used interchangeably, and NVIDIA's Quantum-X Photonics documentation refers to "external laser modules."
The ELS is, deliberately, a dumb component. It has no data path, no DSP, and no serialiser. It emits clean light at a precise wavelength, as efficiently as possible, for years without drifting. That apparent simplicity is misleading: it is one of the hardest parts of the system to build well.
Three physical constraints, and they compound.
Lasers hate heat. A semiconductor laser's output power falls as junction temperature rises, its threshold current climbs, and its wavelength drifts. Worse, degradation mechanisms accelerate roughly exponentially with temperature — a laser that lasts twenty years at 40 °C may not last two at 90 °C. A switch ASIC package running at kilowatt scale is one of the most hostile thermal environments in the rack, which is why CPO switches arrive liquid-cooled.
Silicon cannot make light. Silicon has an indirect bandgap, so it is a poor light emitter. Silicon photonics can guide, split, modulate and detect light extremely well on a CMOS-compatible process — but the light itself has to come from a III-V material, typically indium phosphide. Integrating InP onto silicon is possible via bonding or epitaxial growth, but doing it at yield and at temperature is exactly the problem everyone is trying to avoid.
Failure is not field-serviceable. A pluggable transceiver that fails gets swapped in minutes. An optical engine soldered onto a switch package does not — replacing it means replacing the switch. So the least reliable component gets pulled out of the sealed assembly and put somewhere replaceable. In most designs the ELS is itself a pluggable module in the faceplate, so a laser failure is a field swap rather than a dead switch.
Put together: the laser is the hottest-running, shortest-lived, hardest-to-integrate part of the link. So it leaves.
The first chart below is the whole argument in one line. Laser wear-out follows an Arrhenius relationship with junction temperature, so the difference between an external module at 40–55 °C and an on-package position at 80–100 °C is not a few percent of lifetime — it is roughly an order of magnitude.
The modelled numbers, since the shape matters more than any single point:
| Junction temperature | Relative lifetime | Versus 40 °C |
|---|---|---|
| 40 °C | 100% | baseline |
| 50 °C | 56% | 1.8× shorter |
| 60 °C | 33% | 3.0× shorter |
| 70 °C | 20% | 5.1× shorter |
| 80 °C | 12% | 8.2× shorter |
| 90 °C | 7.8% | 12.8× shorter |
| 100 °C | 5.1% | 19.7× shorter |
Arrhenius wear-out model, activation energy 0.5 eV — a mid-range value for InP devices — normalised to 40 °C. This is the model, not measured vendor data; real devices vary with design and drive current. The order of magnitude between an external module and an on-package position is the point.
They are both semiconductor lasers, and that is roughly where the similarity ends.
| Transceiver laser | External laser source | |
|---|---|---|
| Output | Modulated or CW, per-lane | Continuous-wave only |
| Power | ~10–20 mW typical | 100–400 mW+ per line |
| Wavelengths | One per lane | Often 8, 16 or 32 multiplexed |
| Noise budget | Shared with the whole module | Extremely tight — feeds many lanes |
| Field replaceable | Yes, whole module | Yes, as a separate module |
| Failure blast radius | One port | Many ports at once |
The blast-radius row is the one that changes the engineering. A transceiver laser feeds one link; if it dies, one port goes down. An external laser source can feed an entire optical engine's worth of lanes. Any noise, drift or failure it produces is multiplied across every lane it serves. This is why ELS specifications are so much tighter than a comparable transceiver laser's — the component is doing more work, and its mistakes propagate.
It is also why the power requirement is so much higher. One laser feeding sixteen lanes through splitters needs enough output that each lane still arrives above the modulator's operating point after every insertion loss along the way.
Four numbers do most of the work in separating a competitive ELS from an uncompetitive one.
RIN measures how much a laser's output power fluctuates around its average, in dBc/Hz — more negative is better. Because the modulator writes data onto whatever the laser hands it, any wobble in the CW carrier shows up directly in the received signal as jitter in the eye diagram and, eventually, bit errors.
RIN is the specification most often cited as a differentiator. Figures below −155 dBc/Hz are described by Irrational Analysis, a semiconductor analyst who covers this closely, as more than ten times better than the requirement — for Lumentum's part specifically. Treat vendor-specific comparisons with care, but the physics is not contested: RIN sets the noise floor for every lane the source feeds.
Wall-plug efficiency is optical power out divided by electrical power in. It matters more here than in a pluggable module for a structural reason: CPO's entire pitch is that it saves power, so any watts the laser burns come straight off the benefit.
Two numbers are worth separating. Power conversion efficiency at the laser die runs around 21–24% at 50 °C for a leading part. Module-level wall-plug efficiency, after drive electronics, thermal control and coupling losses, lands nearer 10–13% depending on cooling. Roughly nine-tenths of the electrical power going into an external laser source leaves as heat, which is why the gap between die and module efficiency is where a lot of engineering effort goes.
Linewidth is the spectral purity of the beam — how tightly the light clusters around its nominal wavelength, measured in megahertz or kilohertz. It matters for two reasons. In wavelength-multiplexed designs, wide linewidth causes channels to bleed into their neighbours. In microring-based optical engines, the ring is a sharply tuned resonator; a source that wanders relative to the ring's resonance loses power or falls out of lock.
The engineering tension is that the obvious way to narrow linewidth is a longer laser cavity, and longer cavities mean fewer die per wafer and worse yield. Shrinking cavity length while holding linewidth constant is therefore a direct manufacturing-cost win, and is one of the improvements Lumentum has publicly claimed.
More power means more lanes per laser, or more margin for losses. But power and lifetime trade against each other — driving a laser harder shortens its life and raises its operating temperature, which shortens it again. High-power, narrow-linewidth, low-RIN, long-lived, and efficient is a genuinely hard combination, and it is why the supplier list is short.
| Type | What it is | Fit for ELS |
|---|---|---|
| DFB (distributed feedback) | Single-wavelength laser with an internal grating | The mainstream choice. Narrow linewidth, high power, stable wavelength |
| EML (electro-absorption modulated) | DFB with an integrated modulator | Not for ELS — the modulator is redundant when the engine modulates |
| VCSEL | Surface-emitting, cheap, low power | No — insufficient power and linewidth for CPO |
| Quantum dot | Dot-based gain region, temperature-tolerant | Promising; better high-temperature behaviour, less mature |
For external laser sources the answer is overwhelmingly the high-power CW DFB laser, usually in the O-band around 1310 nm.
The EML row is worth dwelling on, because it explains a real revenue-mix risk. EMLs are today's high-margin datacom laser — the EML product category has been supply-constrained through the 800G ramp. But an EML integrates a modulator, and in co-packaged optics the optical engine already does the modulating. A vendor whose datacom laser business is mostly EML has to migrate that revenue to CW parts as CPO scales, rather than simply riding the same product into a new socket.
Quantum-dot lasers are the interesting long shot. Their carriers are confined in three dimensions, which makes them markedly less temperature-sensitive. If quantum-dot devices matured enough to run reliably inside the package, the entire argument for keeping the laser external would weaken. That is a real, if slow-moving, risk to the ELS thesis.
The CW-WDM MSA (Continuous-Wave Wavelength Division Multiplexing Multi-Source Agreement) is the industry agreement defining wavelength grids for exactly this application — 8, 16 and 32 wavelength variants in the O-band, with defined channel spacing and power levels.
Its purpose is interoperability. Without an agreed grid, every optical engine would need a laser source built to its own wavelength plan, and each pairing would be a bespoke engineering project. With one, a switch vendor can qualify multiple laser suppliers against a common specification.
Macroplane's graph records Lumentum ($LITE), Sivers Semiconductors ($SIVE.ST) and the private Ayar Labs among the members collaborating through the MSA. Sivers is the most directly exposed public name: its Photonics segment builds InP-based DFB lasers and gain chips as single emitters and arrays, explicitly targeting AI data centres, against a market capitalisation a fraction of Lumentum's.
Standardisation cuts both ways for suppliers. It grows the market and makes designing in a second source cheap — which is good for buyers and for challengers, and erosive to whoever currently holds the socket.
| Company | Role in the laser layer |
|---|---|
| Lumentum ($LITE) | High-power CW DFB lasers and ELS modules; CW-WDM MSA member |
| Coherent ($COHR) | Lasers, VCSELs, InP chips; also isolators and fibre-array units |
| Sivers Semiconductors ($SIVE.ST) | InP DFB lasers and gain chips, single emitters and arrays |
| POET Technologies ($POET) | Optical Interposer platform integrating CW lasers into engines |
| Applied Optoelectronics ($AAOI) | Lasers and transceivers, vertically integrated |
| MACOM ($MTSI) | Laser drivers, TIAs and photonic ICs around the source |
| Hamamatsu Photonics ($6965.T) | Laser diodes and precision photonic components |
| Ayar Labs, Lightmatter, Celestial AI | Private; optical I/O and light-source ecosystems |
Note that $LITE supplies CW lasers into $POET's interposer platform and high-volume laser sources into Ayar Labs' light-source ecosystem — the same component reaches the market through several different badges.
Upstream, the constraint is material rather than optical. InP substrates and epiwafers come from AXT ($AXTI) and IQE ($IQEPY); WIN Semiconductors ($3105.TWO) and Sanan ($600703.SS) provide compound-semiconductor foundry capacity; Amkor ($AMKR) handles assembly and test. The whole layer sits on the compound semiconductor epitaxy category, and the InP & Substrates basket tracks it.
Six-inch InP wafer capacity is the number worth watching. InP is more brittle and harder to process than silicon, wafers are smaller, and yields at larger diameters have been a recurring industry problem. Laser supply is ultimately gated by that, not by design capability.
This is the part that makes the ELS commercially interesting rather than merely technically interesting.
| Layer | Indicative gross margin |
|---|---|
| Ultra-high-power laser die | 80–90% |
| External laser source module | 50–60% |
| Isolators, fibre-array units | 30–40% |
| Contract module assembly | Low double digits |
*Figures as characterised by Irrational Analysis; treat as directional. The author discloses a long position in $LITE.*
The spread is the story. A laser die that is genuinely hard to make earns software-like margins. The module that packages it earns half that. The passive components around it earn a third. Assembling the whole thing earns least of all.
This inverts the usual reading of co-packaged optics as a threat to optics vendors. CPO does compress the value captured by module assembly — but it simultaneously increases the value of the laser, because the laser gets harder (more power, tighter noise, more lanes per source) exactly as the module gets simpler. A vendor's exposure to CPO depends entirely on which row of that table it sits in.
An external laser source (ELS) is a separate module that generates continuous-wave laser light and delivers it over fibre into a co-packaged optics engine. It sits outside the switch package because lasers degrade quickly at the temperatures a kilowatt-class switch ASIC produces. The optical engine modulates the light the ELS provides; the ELS itself carries no data.
Because a laser is the least heat-tolerant and least reliable part of an optical link, and a co-packaged optical engine cannot be replaced in the field without replacing the switch. Moving the laser outside keeps it cooler, makes it field-serviceable, and avoids having to integrate indium phosphide onto silicon inside the hottest part of the package.
Continuous wave (CW) means the laser emits steady, unmodulated light at constant power rather than pulsing to encode data. The data is added later, by modulators in the optical engine, which switch that steady beam on and off. The ELS supplies the carrier; the engine writes the signal onto it.
RIN — relative intensity noise — measures fluctuation in a laser's output power, in dBc/Hz, where more negative is better. It matters because the modulator encodes data onto whatever the laser provides, so carrier noise becomes signal noise. Since one external laser source can feed many lanes, its noise affects all of them at once. Leading CPO lasers are quoted below −155 dBc/Hz.
Almost always a high-power continuous-wave DFB (distributed feedback) laser, typically in the O-band around 1310 nm, built on indium phosphide. DFBs give the narrow linewidth, wavelength stability and output power the application needs. VCSELs lack the power and spectral purity; EMLs include a modulator that is redundant when the optical engine modulates.
The Continuous-Wave Wavelength Division Multiplexing Multi-Source Agreement is an industry specification defining standard wavelength grids — 8, 16 and 32 channels in the O-band — for continuous-wave laser sources feeding silicon-photonic engines. It lets switch vendors qualify multiple laser suppliers against one specification. Lumentum, Sivers Semiconductors and Ayar Labs are among the collaborating members.
Lumentum ($LITE) and Coherent ($COHR) are the largest public suppliers of high-power CW lasers. Sivers Semiconductors ($SIVE.ST) builds InP DFB lasers and gain chip arrays specifically for AI data centres. POET Technologies ($POET) integrates CW lasers into its Optical Interposer platform, and Applied Optoelectronics ($AAOI) is vertically integrated into lasers. Ayar Labs, Lightmatter and Celestial AI are private.
The laser die is the highest-margin layer in the optical stack — characterised at 80–90% gross margin for an ultra-high-power part, against 50–60% for the packaged ELS module and 30–40% for surrounding passives like isolators and fibre-array units. These figures come from third-party analysis rather than company disclosure, so treat them as directional.
Possibly. Quantum-dot lasers confine carriers in three dimensions and are markedly less temperature-sensitive than conventional quantum-well devices, which makes in-package operation more plausible over time. If they mature enough to run reliably beside a switch ASIC, the architectural case for an external source weakens. That is the main long-term risk to the ELS thesis.
A transceiver is a complete link in a module: laser, modulator, detector and electronics, handling data in both directions. An external laser source does one job — emit clean, constant light — and contains no data path at all. A transceiver typically outputs 10–20 mW per lane; an ELS often outputs hundreds of milliwatts and feeds many lanes at once through a multiplexed wavelength grid.
This is research and education, not financial advice. Specification and margin figures attributed to third-party analysis reflect that analyst's characterisation, not company disclosure, and that analyst discloses a long position in $LITE. Do your own work.
An external laser source (ELS) is a separate module that generates continuous-wave laser light and delivers it over fibre into a co-packaged optics engine. It sits outside the switch package because lasers degrade quickly at the temperatures a kilowatt-class switch ASIC produces. The optical engine modulates the light the ELS provides; the ELS itself carries no data.
Because a laser is the least heat-tolerant and least reliable part of an optical link, and a co-packaged optical engine cannot be replaced in the field without replacing the switch. Moving the laser outside keeps it cooler, makes it field-serviceable, and avoids having to integrate indium phosphide onto silicon inside the hottest part of the package.
Continuous wave (CW) means the laser emits steady, unmodulated light at constant power rather than pulsing to encode data. The data is added later, by modulators in the optical engine, which switch that steady beam on and off. The ELS supplies the carrier; the engine writes the signal onto it.
RIN — relative intensity noise — measures fluctuation in a laser's output power, in dBc/Hz, where more negative is better. It matters because the modulator encodes data onto whatever the laser provides, so carrier noise becomes signal noise. Since one external laser source can feed many lanes, its noise affects all of them at once. Leading CPO lasers are quoted below −155 dBc/Hz.
Almost always a high-power continuous-wave DFB (distributed feedback) laser, typically in the O-band around 1310 nm, built on indium phosphide. DFBs give the narrow linewidth, wavelength stability and output power the application needs. VCSELs lack the power and spectral purity; EMLs include a modulator that is redundant when the optical engine modulates.
The Continuous-Wave Wavelength Division Multiplexing Multi-Source Agreement is an industry specification defining standard wavelength grids — 8, 16 and 32 channels in the O-band — for continuous-wave laser sources feeding silicon-photonic engines. It lets switch vendors qualify multiple laser suppliers against one specification. Lumentum, Sivers Semiconductors and Ayar Labs are among the collaborating members.
Lumentum ($LITE) and Coherent ($COHR) are the largest public suppliers of high-power CW lasers. Sivers Semiconductors ($SIVE.ST) builds InP DFB lasers and gain chip arrays specifically for AI data centres. POET Technologies ($POET) integrates CW lasers into its Optical Interposer platform, and Applied Optoelectronics ($AAOI) is vertically integrated into lasers. Ayar Labs, Lightmatter and Celestial AI are private.
The laser die is the highest-margin layer in the optical stack — characterised at 80–90% gross margin for an ultra-high-power part, against 50–60% for the packaged ELS module and 30–40% for surrounding passives like isolators and fibre-array units. These figures come from third-party analysis rather than company disclosure, so treat them as directional. !A silicon wafer populated with photonic-IC dies, each carrying 32 integrated lasers — the wafer-scale alternative to a standalone external module. If integration like this reaches hyperscale qualification volume, the "external" in external laser source becomes optional (image: N. Nader / NIST, public domain).
Possibly. Quantum-dot lasers confine carriers in three dimensions and are markedly less temperature-sensitive than conventional quantum-well devices, which makes in-package operation more plausible over time. If they mature enough to run reliably beside a switch ASIC, the architectural case for an external source weakens. That is the main long-term risk to the ELS thesis.
A transceiver is a complete link in a module: laser, modulator, detector and electronics, handling data in both directions. An external laser source does one job — emit clean, constant light — and contains no data path at all. A transceiver typically outputs 10–20 mW per lane; an ELS often outputs hundreds of milliwatts and feeds many lanes at once through a multiplexed wavelength grid. This is research and education, not financial advice. Specification and margin figures attributed to third-party analysis reflect that analyst's characterisation, not company disclosure, and that analyst discloses a long position in $LITE. Do your own work.