Co-Packaged Optics (CPO): Mapping the 2026-2028 Investment Landscape

The Capital Repository // February 2, 2026

Executive Summary

Co-packaged optics is transitioning from engineering aspiration to commercial reality, but the investment case requires careful parsing. The consensus narrative—"2026 is the CPO inflection point"—obscures a critical distinction that separates real near-term revenue from story-stock extrapolation: scale-out versus scale-up.

NVIDIA's CPO adoption is real, but confined to Ethernet and InfiniBand switches for rack-to-rack (scale-out) connectivity. NVIDIA's own announcement states Quantum-X Photonics CPO switches ship "later this year" [last year] (2025) and Spectrum-X Photonics CPO switches are "coming in 2026." The much larger prize—GPU-to-GPU interconnects within NVLink domains (scale-up)—remains copper-dominant through Rubin Ultra (2H 2027), with optics becoming necessary only when NVLink extends across racks—likely 2028+ with Feynman.

This matters because the optical component suppliers being bid up on "NVIDIA CPO" narratives are largely exposed to scale-out switches—a meaningful but finite opportunity. Meanwhile, Linear Pluggable Optics (LPO) has emerged as a credible bridge technology, potentially delaying or compressing the CPO adoption window.

Investment Framework At current valuations (2-5x EV/Sales across most names), the sector already prices in aggressive adoption curves. Size positions accordingly.

The Thesis in Three Sentences

CPO is a necessary technology for AI infrastructure scaling, but commercial deployment remains early-stage with most near-term volume in scale-out switches, not GPU interconnects.
The supply chain is concentrated around chokepoints—EML lasers, fiber array units, precision assembly—creating pricing power for incumbents but also exposing them to foundry integration risk.
Current valuations embed adoption curves that may be delayed by LPO bridge solutions, NVIDIA's copper persistence in scale-up, and TSMC packaging ambitions.

Understanding the Scale-Out vs. Scale-Up Distinction

This is the single most important concept for CPO investing.

Scale-Out (Where CPO is happening now)

Rack-to-rack connectivity via Ethernet (Spectrum-X) or InfiniBand (Quantum-X) switches
NVIDIA Quantum-X Photonics CPO InfiniBand switches: "Later this year" (2025) per NVIDIA; independent reporting indicates first Quantum 3450-LD CPO switch in H2 2025
NVIDIA Spectrum-X Photonics Ethernet CPO switches: "Coming in 2026" per NVIDIA; industry sources indicate H2 2026
Broadcom Bailly CPO (via Micas Networks): Currently shipping, with 1M+ "flap-free" port-hours demonstrated at Meta
This is real revenue for optical component suppliers today

Scale-Up (Where CPO is NOT happening yet)

GPU-to-GPU interconnects within NVLink domains
Current NVLink domains rely heavily on copper backplanes and cabling at rack scale (NVL72/NVL144 configurations)
Rubin Ultra NVL576 (2H 2027): Copper midplane persists for in-rack scale-up
Optics becomes necessary when NVLink extends across racks—the multi-rack step expected with Feynman (2028+)

Quantifying the Opportunity Gap The scale-out switch market represents perhaps 50,000-100,000 CPO switch ports shipping in 2026-2027 at $3,000-5,000 optical content per port = $150M-$500M in near-term optical revenue, scaling to $2-3B by 2028 as volumes ramp.

The scale-up GPU interconnect market—if/when it transitions to optics—represents 10-20x more ports at similar or higher content value. But this remains copper-dominant through at least 2028.

Why this matters: Companies being pitched as "NVIDIA CPO beneficiaries" are mostly exposed to the smaller, nearer-term scale-out opportunity. The much larger scale-up prize remains 2+ years away.

The LPO Bridge: A Credible Competitive Threat

Linear Pluggable Optics (LPO) has emerged as an intermediate solution that may delay or compress the CPO adoption window.

What's Verified

LPO eliminates in-module DSP, moving signal processing to host ASICs—same power-reduction goal as CPO, different implementation
Industry analysts project LPO and CPO combined will capture a meaningful share of 800G/1.6T deployments in 2026-2028, with LPO well-suited for short-reach applications
Multiple hyperscalers are qualifying LPO for short-reach links

Why LPO Matters

Eliminates DSP power consumption (the primary driver of pluggable inefficiency)
Works with existing faceplate infrastructure (no data center redesign)
Significantly cheaper than CPO (no advanced packaging)
Adequate for many short-reach applications where CPO's density advantage is overkill

Investment Implication LPO and CPO are likely complementary rather than purely competitive—combined adoption accelerates overall optics migration. But LPO's viability for short-reach applications compresses CPO's addressable market in the near term. If reliability/yield hurdles limit LPO at 1.6T+, CPO could capture share faster than projected.

TSMC COUPE: Value Chain Power Shift

The most significant structural risk to optical component suppliers comes from foundry integration. TSMC's Compact Universal Photonic Engine (COUPE) represents a potential power shift in the CPO value chain.

What's Reported

Industry packaging coverage indicates TSMC has presented advanced CPO integration with CoWoS/SoIC platforms
NVIDIA and Broadcom are among early engagers with COUPE development
Production timeline remains uncertain; meaningful volume unlikely before late 2027

Nuanced Impact

COUPE may not "collapse" the discrete supply chain so much as shift bargaining power toward the foundry. Key considerations:

Foundry integration can expand the pie while still requiring external lasers, fiber harnessing, test, burn-in, and system-level qualification
Hyperscalers often prefer multi-sourcing; COUPE may standardize interfaces rather than fully internalize assembly
Most exposed are pure assembly/value-add steps that don't control IP (e.g., Fabrinet)
Least exposed are companies with proprietary laser technology or vertically integrated platforms

Investment Implication COUPE pressures margins and shifts bargaining power. It's a headwind for contract assemblers and a tailwind for foundries. The "existential" framing is too binary—but the directional pressure is real.

Valuation Framework

As of February 2, 2026

To avoid mixing noisy GAAP P/E with forward estimates, we use a consistent framework: EV/LTM Sales and Gross Margin (relevant for optical component cycles), plus EV/EBITDA where meaningful.

Ticker	Price	EV	LTM Rev	EV/Sales	Gross Margin	EV/EBITDA	Rating
MRVL	$79	$65B	$5.5B	11.8x	46%	27x	Equal Weight
FN	$500	$17B	$4.0B	4.3x	12%	22x	Underweight
GLW	$110	$100B	$14B	7.1x	34%	18x	Equal Weight
COHR	$222	$38B	$5.8B	6.6x	37%	32x	Equal Weight
LITE	$423	$29B	$1.8B	16.1x	35%	65x	Avoid
AAOI	$44	$2.8B	$0.5B	5.6x	22%	NM	Speculative
POET	$5.87	$0.9B	~$0	NM	NM	NM	Avoid

Note: EV and revenue figures rounded; gross margins reflect recent quarters. AAOI EBITDA negative.

Valuation Context The optical component sector historically trades at 2-4x EV/Sales through-cycle. Current multiples of 5-16x embed aggressive growth assumptions. LITE at 16x sales and 65x EBITDA is the most extreme; MRVL at 12x sales reflects broader AI semiconductor premium.

Company Analysis

Marvell Technology (MRVL)EQUAL WEIGHT

The Bull Case: Most direct CPO integration story through custom XPU architecture and electro-optics. AWS partnership for Trainium accelerators. Nova 2 800G PAM4 DSPs power most CPO optical engines. Data center segment growing 40%+ annually.

The Bear Case: Custom silicon wins concentrated (AWS ~40% of data center revenue). Celestial AI acquisition ($3.25B, up to $5.5B) is pre-revenue photonics—meaningful revenue not expected until H2 FY28. If AWS shifts to Alchip or internal silicon, Marvell faces stranded asset risk.

Valuation: At 12x EV/Sales and 27x EV/EBITDA, MRVL trades at premium to optical peers but in-line with AI semiconductor comps. Valuation assumes continued custom silicon wins and successful Celestial integration.

Position Sizing: Core position (equal-weight). Best risk-managed CPO exposure given diversification, but customer concentration warrants monitoring.

Fabrinet (FN)UNDERWEIGHT

The Bull Case: Precision optical manufacturing moat; Thailand footprint provides tariff insulation. Customer roster includes Lumentum, Coherent, NVIDIA. Consistent 20%+ operating margins.

The Bear Case: Contract manufacturer economics cap upside—Fabrinet doesn't own IP. At 4.3x EV/Sales and 22x EV/EBITDA, valuation is elevated vs. historical 2-3x sales range. TSMC COUPE shifts bargaining power toward foundry; pure assembly steps that don't control IP face margin pressure.

Valuation: Multiple expansion reflects CPO assembly optimism. If COUPE succeeds in standardizing interfaces, Fabrinet's value-add compresses even if volumes grow.

Position Sizing: Underweight. Structural headwinds at elevated multiple.

Corning (GLW)EQUAL WEIGHT

The Bull Case: Every CPO switch connects via fiber. Meta deal ($6B through 2030) provides revenue anchor. Corning supplies fiber harnesses, connectors, and fiber array units (FAUs) into Broadcom's Bailly CPO ecosystem. Balance sheet fortress.

The Bear Case: At 7x EV/Sales, valuation is elevated for a diversified materials company. Display segment (~25% revenue) faces structural headwinds. Commodity fiber pricing limits margin expansion.

CPO Exposure Clarification: Corning's CPO participation is primarily the fiber/connectivity layer (harnesses, connectors, FAUs), not lasers/PIC/EIC. Real but limited exposure—Corning supplies infrastructure, not optical engines.

Position Sizing: Core position (equal-weight) as portfolio ballast. Preferred hold for diversified AI infrastructure exposure, not CPO pure-play.

Coherent Corp (COHR)EQUAL WEIGHT

The Bull Case: Vertical integration from compound semiconductor wafers through finished transceivers. 800G/1.6T transceiver leadership with hyperscaler design wins. Silicon carbide and InP laser platforms provide moats.

The Bear Case: At 6.6x EV/Sales and 32x EV/EBITDA, valuation assumes continued share gains. $5B debt from II-VI merger limits flexibility. Chinese competition (Innolight, Eoptolink) creating structural pricing pressure—hyperscalers actively qualifying Chinese suppliers.

EML Competitive Dynamics: Coherent has announced and demonstrated 200G-class D-EML capability and is moving toward broader production availability. Additional InP/EML capacity is being developed by multiple suppliers, including in Asia, which will pressure pricing as supply normalizes.

Position Sizing: Core position (equal-weight) given vertical integration moat. Monitor debt covenants and datacom pricing.

Lumentum (LITE)AVOID

The Bull Case: EML supply leadership for 800G/1.6T transceivers. NVIDIA Silicon Photonics Ecosystem Partner.

The Bear Case: At 16x EV/Sales and 65x EV/EBITDA, valuation is extreme. EML leadership reflects temporary supply constraints, not structural moat. As additional EML capacity comes online (Coherent, Asian suppliers), pricing power erodes. Stock up ~400% in past year.

Historical Pattern: Component shortages attract capacity investment → supply normalizes → pricing power evaporates. This cycle typically runs 18-24 months. NVIDIA's EML procurement for 2026 switches is largely complete; 2027-2028 volumes will face expanded supply.

Position Sizing: Avoid. Risk/reward deeply asymmetric. This is a short candidate at current levels.

Applied Optoelectronics (AAOI)SPECULATIVE

The Bull Case: 800G transceiver ramp with first volume order from major hyperscaler (December 2025). Amazon relationship structured as a transaction/warrant agreement with expected purchases up to a cap over 10 years. New 210,000 sq. ft. Texas facility under construction.

The Bear Case: Company burned cash for six consecutive years before recent stabilization. Customer concentration extreme. China manufacturing creates tariff/geopolitical risk. AAOI has introduced a 400mW laser marketed for silicon photonics and CPO—early-stage positioning, not proven CPO revenue.

Position Sizing: Speculative allocation only (<1%). Turnaround credible but execution risk high.

POET Technologies (POET)AVOID

The Bull Case: Optical Interposer platform promises integrated photonics without complex hybrid bonding.

The Bear Case: Effectively pre-revenue (~$300K TTM). Technology 2-3 years behind leaders in integration density. Management has missed multiple commercialization timelines. No hyperscaler design wins.

Position Sizing: Avoid. Pre-revenue speculation with execution history problems.

Scenario Framework

Bull Case (20% probability)

CPO adoption accelerates. NVIDIA pulls forward Feynman. LPO proves inadequate at 1.6T+. TSMC COUPE encounters yield issues. Hyperscaler capex elevated through 2028.

Implications: Current valuations justified; upside in diversified names.

Base Case (50% probability)

CPO deploys per current timelines—scale-out 2025-2026, scale-up 2028+. LPO captures meaningful short-reach share. COUPE ramps late 2027, pressuring discrete assemblers. Chinese suppliers gain commoditized transceiver share.

Implications: Current valuations already price in base case. Limited upside, moderate downside.

Bear Case (30% probability)

NVIDIA copper persistence extends through Feynman (2029+). LPO adequate for 1.6T. COUPE succeeds, shifting value to foundry. Hyperscaler capex cycle turns.

Implications: Sector derating of 30-50%. Most exposed: LITE, FN.

Position Summary

Ticker	Rating	Position	Rationale
MRVL	Equal Weight	Core	Best risk-managed exposure; customer concentration
GLW	Equal Weight	Core	Preferred hold for diversified AI infra
COHR	Equal Weight	Core	Vertical integration; debt/competition headwinds
FN	Underweight	None	Structural COUPE pressure at elevated multiple
LITE	Avoid	None/Short	Extreme valuation on temporary dynamics
AAOI	Speculative	<1%	Turnaround credible; execution risk
POET	Avoid	None	Pre-revenue; execution history

What We're Watching

NVIDIA GTC 2026 (March): Feynman details, NVLink CPO timeline
TSMC Quarterly Calls: COUPE yield/timeline updates
LPO Adoption Data: LightCounting/Dell'Oro quarterly reports
EML Capacity Announcements: Supply normalization signals
Hyperscaler In-House Efforts: Meta/Google internal silicon photonics

Conclusion

The CPO thesis is valid, but most hype valuations are not.

At 5-16x EV/Sales across most names, the sector already embeds aggressive adoption curves. The best risk-managed exposures are diversified plays (Marvell, Corning) where CPO is upside optionality rather than the entire thesis. The most dangerous positions are single-catalyst stories (Lumentum EML, Fabrinet assembly) at extreme valuations.

The base case isn't bullish enough to justify current prices. Size accordingly.

Disclosures: Positions may exist in mentioned securities. This analysis reflects publicly available information as of February 2, 2026, and should not be construed as investment advice.