Still required.
It routes fine-pitch electrical connections and supports the ASIC and optical engines. “CPO” describes placement, not the substrate material.
PM field guide · Updated 11 August 2026
When copper runs out of road, the package learns to speak light.
A physical, architectural and investable map of co-packaged optics—where the electrons stop, where photons begin, who supplies each layer, and which bottlenecks can actually move earnings.
Scale-out CPOEarly production
Optical scale-upEmerging / prove volume
Pluggable opticsStill dominant & coexisting
Core bottleneckYield + test + fiber attach
01 / Physical cutaway
Every part, in its actual place.
Read left to right for the signal path. Read vertically inside the package to see why the substrate and PCB do not disappear when optics moves closer to the ASIC.
CPO switch assembly · simplified, not to scaleElectrons → photons → electrons
electrical signalcontinuous laser lightmodulated optical data
Still underneath.
The switch board distributes power, management signals and lower-speed connectivity. CPO shortens only the highest-speed electrical reach.
The transmission medium.
The PIC modulates laser light; fiber carries those photons to the next receiver. Optics is the mechanism, CPO is the architecture.
02 / Signal flow
Where photons enter—and where data rides them.
A laser creates light, but light is not data until a modulator imprints the electrical signal. The receiving photodetector reverses the process.
Route bits
The host ASIC chooses where an electrical data stream should go.
Drive the PIC
The EIC conditions and amplifies the high-speed electrical signal.
Create light
An external InP CW laser supplies stable, unmodulated photons.
Modulate
MZI, MRR, EAM or another device imprints bits onto wavelengths.
Carry
FAU alignment couples light into fiber; WDM combines wavelengths.
Detect
A receiver PIC and TIA convert photons back into electrical bits.
03 / Architecture ladder
Same job, different conversion point.
Click a lens to highlight the central trade-off. “Closer” saves electrical power but makes packaging, thermal design, yield and field replacement harder.
Pluggable
ASIC ─ long electrical ─ [faceplate optic] ⇢ fiber
- Field replaceable and multi-vendor
- Long PCB trace needs retiming / DSP
- Best service boundary; higher power
LPO
ASIC ─ electrical ─ [linear pluggable] ⇢ fiber
- Removes or simplifies module DSP
- Still removable at faceplate
- Tighter host SerDes and channel margin
NPO / Open CPX
ASIC ─ short electrical ─ [socketed engine] ⇢ fiber
- Engine sits near but remains separate
- Known-good module can be tested first
- Connector loss buys reworkability
CPO
[ASIC + optical engines] common package ⇢ fiber
- Shortest electrical path
- Best density / power potential
- Hardest yield, thermal and service stack
XPO
ASIC ─ [12.8T liquid-cooled pluggable] ⇢ fiber
- Dense pluggable extension
- Targets 4× OSFP density
- Keeps replacement boundary outside package
04 / Network domains
Do not collapse scale-up, scale-out and DCI.
The same laser or optical-engine supplier can serve all three, but port count, reach, protocol, modulation and timing differ. That changes the revenue denominator.
Inside one compute domain
Scale-up
GPU/XPU→fabric switch→GPU/XPU
Copper is incumbent at short reach. OCI creates an optical PHY path, but broad optical volume remains a thesis to prove—not a shipped-port fact.
Between servers and racks
Scale-out
NIC→switch→rack
Ethernet/InfiniBand use pluggables today. Broadcom Bailly was delivered in 2024 and NVIDIA Spectrum-X Photonics entered production in 2026: CPO is no longer “zero shipments.”
Between sites
Scale-across / DCI
campus A→coherent optic→campus B
Power and land fragmentation increase coherent optics and fiber demand. This is usually a longer-reach transport architecture, not the same CPO package.
05 / Standards stack
Three MSAs solve three different problems.
Membership signals architectural influence—not guaranteed design share or revenue. The standards can coexist because they specify different layers.
OCI
Defines: an optical chiplet/PHY interface for AI scale-up.
Protocol-agnostic, NRZ + WDM. Supports pluggable, on-board and co-packaged implementations. Think signal language.
Open CPX
Defines: the socket, form factor and mechanical/electrical/thermal/optical/management interfaces for optical engines.
Think replaceable near-package engine.
XPO
Defines: a dense, liquid-cooled 12.8T pluggable form factor.
Targets 204.8T per optical unit and 4× OSFP density. Think pluggable density extension.
06 / Investable supply chain
Follow both matter and yield.
The apparent “optics” revenue pool is split across compound-semiconductor materials, foundry wafers, EIC/PIC design, packaging, alignment, test, fibers and systems. The highest torque may sit where qualification is slowest.
01 · Starting materials
InP wafersAXTI · Sumitomo Electric
Photonics SOISoitec
Package materialsABF / laminate / substrate
02 · Light source
Epitaxy + CW lasersCOHR · LITE · AAOI
ELS moduleredundancy · cooling · control
Alternative sourcesVCSEL / EML by architecture
03 · Silicon
Switch / host ASICAVGO · NVDA · MRVL
EICdriver · TIA · control
PIC / SiPh foundryTSM · GFS · TSEM
04 · Optical engine
ModulatorMZI · MRR · EAM · TFLN
2.5D / 3D assemblyPIC + EIC integration
FAU / fiber attachsub-micron alignment
05 · Test & package
Wafer electro-optical testTER · FORM · 6857 JP
Known-good dieprevent bad die in costly package
Advanced assemblyTSM · ASE · AMKR
06 · Architecture
Pluggable / LPOserviceable faceplate
NPO / CPXsocketed near package
CPOcommon package assembly
XPOdense liquid-cooled pluggable
07 · Fabric
Fiber + connectorsGLW · APH · Molex · Samtec
Switch systemsANET · CSCO · NVDA
AI operatorhyperscaler · neocloud
07 / Bottleneck scoreboard
Six places the CPO ramp can break.
Scores are a qualitative PM prioritization, not measured probabilities. They rank how directly a failure can gate volume and how under-modeled the layer appears.
5/5
Electro-optical KGD test
Electrical pass is insufficient: the PIC must also meet wavelength, power, coupling and thermal specs before expensive co-assembly.
Watch: test time · parallelism · wafer probing · yield escape
5/5
Fiber attach & alignment
Sub-micron placement must survive thermal cycling at high volume. Passive alignment and automation determine throughput.
Watch: coupling loss · units/hour · rework · connector standard
4/5
High-power CW laser
CPO can use fewer laser dies but require much more optical output, efficiency, lifetime and redundancy per source.
Watch: qualified watts · $/watt · InP yield · new capacity
4/5
PIC process control
Modulator and waveguide performance varies with dimensions, temperature and wavelength. Foundry repeatability drives good optical engines.
Watch: wafer uniformity · insertion loss · wavelength control
4/5
Thermal separation
The ASIC is hot; lasers and resonant photonics are temperature-sensitive. Cooling must protect both without excessive complexity.
Watch: junction temperature · liquid loop · thermal drift
3/5
Serviceability / standards
A failed co-packaged engine can strand a valuable switch. Socketed engines, external lasers and interoperability reduce the blast radius.
Watch: field replaceability · MSA adoption · multi-source proof
08 / Public-equity map
Exposure is not the same as purity.
Representative, not exhaustive. “Direct” means the product touches the CPO bill of materials; it does not mean the revenue is already material. Filter by layer.
NVIDIA
NVDASpectrum-X Photonics CPO switches, AI scale-up demand pull, external-laser ecosystem and optical supplier capacity commitments.
Confirm: CPO units · attach rate · Rubin/Feynman topologyRisk: CPO is a small input to a vast platform.Broadcom
AVGOBailly CPO switch platform, Tomahawk switching and silicon-photonics integration; also custom ASIC exposure.
Confirm: customer ramps · optical-engine yield · revenue disclosureRisk: limited product-level revenue visibility.Marvell
MRVLOptical DSPs, switches, custom silicon, SiPh and Open CPX participation connect multiple layers.
Confirm: scale-up vs scale-out mix · SiPh design winsRisk: architecture can shift value among DSP and optical-engine layers.Arista
ANETAI Ethernet switches and XPO launch. A demand and architecture beneficiary rather than a component-purity play.
Confirm: XPO availability · customer adoption · port mixRisk: merchant optical value may accrue outside Arista.TSMC
TSMCOUPE silicon photonics plus leading-edge ASIC and advanced packaging. COUPE-on-substrate production targeted for 2026.
Confirm: customer mix · optical package capacity · pricingRisk: optics immaterial relative to core foundry earnings.GlobalFoundries
GFSSpecialty silicon photonics foundry; SCALE optical-module solution is OCI-capable and supports CWDM/DWDM.
Confirm: wafer volume · customer qualifications · utilizationRisk: announcement may precede material production.Tower Semiconductor
TSEMSpecialty silicon-photonics and analog foundry exposure; candidate beneficiary as PIC wafers scale.
Confirm: AI-design wins · qualified wafer starts · capacityRisk: sector relevance does not prove bottleneck economics.ASE / Amkor
3711 TW · AMKRAdvanced assembly and test capability. CPO requires heterogeneous optical/electrical packaging, but named program exposure must be proven.
Confirm: optical-engine programs · attach/test contentRisk: foundries or captive lines retain the highest-value integration.Coherent
COHRInP lasers, transceivers, optical components, CPO/NPO platforms and high-power CW sources.
Confirm: high-power laser capacity · order coverage · marginRisk: capacity catches demand and normalizes scarcity rents.Lumentum
LITEEMLs, high-power CW lasers and ELS modules; NVIDIA partnership includes investment and purchase commitments.
Confirm: qualified output · CPO mix · $/watt · customer concentrationRisk: commitments are not the same as recognized revenue.Applied Optoelectronics
AAOITransceivers and a 400mW-class hot-swappable ELSFP product for CPO/NPO architectures.
Confirm: qualification · yield · customer concentration · cash needsRisk: small-base execution and competitive intensity.AXT
AXTIInP wafer substrate exposure. Benefits only if qualification, export access and wafer economics translate through laser demand.
Confirm: InP mix · permits · customer inventory · wafer pricingRisk: substrate availability may not be the binding constraint.Teradyne
TERPhoton 100 and double-sided silicon-photonics wafer probe support electro-optical known-good-die screening.
Confirm: installed base · test-cell ASP · utilization · revenue materialityRisk: CPO test remains small within total tester revenue.Advantest
6857 JPWorks with OpenLight on silicon-photonics test, alongside broad AI semiconductor test exposure.
Confirm: optical configurations · throughput · customer adoptionRisk: optical test contribution is not separately disclosed.FormFactor
FORMPharos photonics wafer probing and precision test interfaces connect optical and electrical measurement.
Confirm: probe-card replacement content · adoption · marginsRisk: custom interface spend may stay lumpy.Keysight
KEYSIntegrated-photonics characterization software and instruments used in R&D and production workflows.
Confirm: production vs lab mix · automation contentRisk: R&D exposure may not scale linearly with unit volume.Corning
GLWOptical fiber and connectivity; NVIDIA partnership targets expanded U.S. manufacturing for AI infrastructure.
Confirm: fiber-km content · product mix · capacity utilizationRisk: broad company and potential fiber commoditization.Amphenol
APHHigh-speed connectors and cable assemblies across data-center infrastructure; CPO exposure is adjacency rather than a pure play.
Confirm: optical attach content · product-specific design winsRisk: CPO can alter connector value rather than simply add it.09 / Adoption clock
Production exists; ubiquity does not.
The March article series usefully mapped the stack, but its “zero shipments” frame has been overtaken by disclosed scale-out milestones. Optical scale-up remains the larger, less proven option value.
51.2T switch with eight 6.4T optical engines was delivered to customers—an early scale-out CPO milestone.
Spectrum-X and Quantum-X photonics bring silicon photonics into the switch roadmap.
NVIDIA says Spectrum-X Ethernet Photonics is in production. OCI, Open CPX and XPO formalize separate interface paths; TSMC COUPE and GF SCALE advance.
XPO targets availability and optical scale-up could expand ports per accelerator. Treat timing, attach rate and bill of materials as scenarios, not facts.
10 / Source reconciliation
What the five articles add—and what changed.
The articles are one author/corpus, so repeated claims are not independent consensus. They are most valuable as an architecture map and hypothesis generator.
Right: the electrical power wall is real.
Moving conversion closer to the ASIC shortens high-speed electrical traces and can reduce retimer/DSP power.
Updated: shipments are not zero.
Broadcom delivered Bailly in 2024; NVIDIA disclosed Spectrum-X Ethernet Photonics production in 2026.
Right: CPO is a packaging problem.
PIC/EIC integration, external lasers, fiber attach, thermal control and known-good-die test determine volume yield.
Qualified: “DSP elimination.”
CPO can remove or reduce a pluggable-module DSP, but retiming, FEC and signal processing vary by topology.
Right: external lasers create a separate lane.
They isolate heat and enable replacement, but high optical power, efficiency and lifetime become key supplier constraints.
Scenario, not fact: market sizes and dates.
Power savings, yields, TAMs, ASPs and scale-up adoption timing are author/vendor estimates until observed in product and earnings data.
11 / PM operating system
What confirms—or kills—the thesis.
Do not trade the word “CPO.” Trade the rate-limiting layer, after proving revenue materiality, qualification and who keeps the economics.
Confirmation triggers
- Production switch SKUs disclose customer deployment, ports and repeat orders.
- Optical scale-up moves from MSA/specification to named XPU product and qualified volume.
- Laser suppliers report high-power CW/ELS mix, purchase commitments converting to revenue and constrained qualified output.
- Test and fiber-attach suppliers show faster orders, utilization and recurring consumables.
- Multi-source standards prove interoperability without sacrificing link margin.
Falsifiers
- Copper, active electrical cable or improved SerDes extends economical reach.
- LPO/XPO preserves serviceability with acceptable power, slowing CPO adoption.
- PIC yield or fiber-attach throughput fails to reach high-volume economics.
- New InP/laser capacity or alternative VCSEL architectures erase scarcity premiums.
- System vendors internalize the value while component suppliers face ASP pressure.
12 / Evidence ledger
Source the architecture; challenge the forecast.
Five linked Damnang articles were read in full. Primary company and MSA sources update their March 2026 snapshot. Links below support the claims nearest them.
The War of Light Has BegunMSA landscape and architecture transition · secondary analysis · March 2026.
The War of Light, Part 2Membership map and value-chain positioning · secondary analysis.
Why CPO Is Becoming Inevitable? Part 1Electrical power wall, architecture ladder and adoption thesis.
CPO, Fully Dissected · Part 2CPO anatomy: EIC, PIC, external laser, packaging and fiber.
The Illusion of CPO · FinalBottlenecks, company map and timing framework · secondary analysis.
OCI MSA · officialOptical PHY interface for AI scale-up; founding members and specification.
Open CPX MSA · officialSocket and interfaces for replaceable near-package optical engines.
Arista XPO announcement12.8T dense liquid-cooled pluggable and target density.
Broadcom Bailly deliveryMarch 2024 customer delivery of 51.2T CPO switch platform.
NVIDIA 2026 production updateSpectrum-X Ethernet Photonics disclosed as in production.
NVIDIA CPO anatomyExternal lasers, silicon photonics, package and ecosystem diagram.
TSMC COUPE roadmapCOUPE-on-substrate production timing and vendor performance claims.
GlobalFoundries SCALEOCI-capable silicon-photonics optical-module platform.
Teradyne Photon 100Production silicon-photonics wafer test and known-good-die workflow.
Advantest + OpenLightSilicon-photonics test collaboration.
FormFactor PharosPhotonics wafer probing and fiber-array alignment.
NVIDIA + LumentumInvestment, purchase commitment and future laser capacity access.
NVIDIA + CorningLong-term AI fiber and connectivity manufacturing partnership.