Optical Content Growth Is Certain; Architecture Debate Is Just Noise

Overview

This research report focuses on the core controversy in the AI data center optical interconnect field—the route competition between Co-Packaged Optics (CPO) and Near-Packaged Optics (NPO). Morgan Stanley points out that the market is overly fixated on technical routes and adoption timing, while ignoring the more fundamental trend: regardless of the final architecture form, the surge in bandwidth demand driven by AI computing power expansion will deterministically drive optical content growth. Based on this judgment, the firm holds a more positive attitude towards Lumentum (LITE), Coherent (COHR), and Corning (GLW), which have recently pulled back.

Core views

The core argument of the report is that 'optical content growth is independent of architecture.' Currently, each GPU is configured with only 2-4 pluggable optical modules (for scale-out networks), whereas the NVL576 architecture introduces CPO, requiring about 17 optical engines just for scale-up interconnects (Lumentum estimates ~18), with a fully optical configuration reaching 35. If SerDes rates remain at 200G, the number of optical engines in the next-generation Feynman architecture could double to 70 per GPU. Regarding technical routes, the report details four major challenges facing CPO: higher failure rates of optical components compared to electronic ones, manufacturing costs 8-10 times higher than traditional pluggable solutions, an immature ecosystem (highly concentrated around Nvidia and Broadcom), and difficulty in hot-swapping for maintenance. As a transitional solution, NPO separates optical packaging from the chip, offering advantages in production yield, assembly difficulty, and supply chain openness, and is seen as an intermediate path to CPO. Regarding adoption pace, Morgan Stanley maintains a relatively conservative judgment: optical engine demand in 2027 will be approximately 6-7 million units (corresponding to about 59,000 switches), significantly lower than some investors' optimistic expectations of 20-30 million; large-scale adoption of CPO/NPO is expected to appear in 2028-2029. The report believes that this conservative expectation limits stock performance in the short term, but it is essentially a timing issue rather than a directional one.

Analysis framework

The firm's analysis follows the main line of 'bandwidth demand → optical content → technical route → impact on individual stocks.' First, it establishes a core cognitive framework: the fundamental driver of optical content growth is bandwidth demand, not specific packaging forms. On this basis, by comparing the changes in the number of optical engines under different architectures (current pluggable → NVL576 hybrid → NVL576 full CPO → Feynman), it quantifies the certainty of 'content growth.' For the analysis of technical routes, the report adopts a 'challenge-path' framework: first identifying the bottlenecks of CPO in yield, cost, ecosystem, and maintenance, then arguing the rationality of NPO as a transitional solution, finally pointing to the conclusion that 'regardless of the path, optical content is growing.' At the individual stock level, through Nvidia's equity investment commitments to LITE/COHR/GLW, it demonstrates the capacity binding relationship of these three suppliers, explaining that demand will not be simply substituted.

Methodology notes

• Industry/Industrial Analysis FrameworkUpstream-Midstream-Downstream Transmission

  Optical Content Quantification Method: Establishing a quantitative model linking architecture evolution to optical device usage through the mapping relationship of GPU → Switch → Optical Engine

  The report breaks down AI data center architecture into three layers: GPU, network equipment, and optical devices. By using parameters such as port rates, SerDes speeds, and oversubscription ratios under different architecture schemes, it calculates the number of optical engines required per GPU, thereby transforming technical route debates into quantifiable demand forecasts

• Industry/Industrial Analysis FrameworkSupply and Demand Framework

  Demand Elasticity Analysis: The elasticity coefficient of optical content to bandwidth demand is greater than 1, meaning bandwidth growth leads to disproportionate growth in the number of optical devices

  When scale-up shifts from copper cables to optical interconnects, not only does the transmission medium change, but more importantly, the 'attach rate' of optical engines increases significantly—from 2 per GPU to 17-35. This non-linear growth is the core logic for assessing the growth space of optical suppliers

• Valuation MethodPE/PEG valuation

  Growth Premium Valuation: Assigning valuation multiples higher than historical ranges to AI-related optical suppliers to reflect their growth attributes

  The report applies a 28-30x FY28e PE to COHR, a 35x FY28 PE to GLW, and a 30x CY28 PE to LITE, explicitly stating that these are premiums over historical trading ranges due to AI growth prospects. This 'theme premium' valuation method is common practice for current tech growth stocks

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Lumentum Holdings Inc (LITE)

  Core supplier of AI optical engines, benefiting from optical content growth

  Strengths

  Leading capacity in the optical engine field, equity cooperation binding with Nvidia

  Risks

  EML pricing decline due to slowing AI demand; CW (continuous wave laser) replacing EML; Cloud Light acquisition dragging down gross margin; industrial demand affected by macro factors

• Coherent Corp (COHR)

  Core supplier of AI optical engines, benefiting from optical content growth

  Strengths

  High demand visibility, accumulating leverage for gross margin improvement

  Risks

  Difficulty in raising prices due to macro headwinds or tariffs; fierce competition in SiC market; AI construction pauses due to factors like DeepSeek; competitive pressure in data center communications

• Corning Inc (GLW)

  Core supplier of AI optical engines, benefiting from optical content growth

  Strengths

  Gross margin improvement brought by fiber optic business entering data center opportunities

  Risks

  Delays in fiber deployment/AI plans; emergence of CPO alternative solutions; macro disturbances affecting pricing and cash flow; loosening of pricing discipline in display business

Key data

• Optical Engines per GPU (Current)2-4 unitsTraditional pluggable optical module configuration
• Optical Engines per GPU (NVL576 Hybrid Architecture)~17 unitsIntroducing CPO for scale-up interconnects
• Optical Engines per GPU (Full Optical Configuration)~35 unitsFully optical interconnects between GPU and switch
• Lumentum Value per Optical Engine~$150-200Data disclosed by LITE itself
• 2027 Optical Engine Demand Forecast6-7 million unitsMorgan Stanley expectation, lower than some investors' optimistic expectation of 20-30 million
• CPO Switch 2024-2028 CAGR144%Based on Yole and Morgan Stanley forecasts
• CPO ASP Premium vs. Traditional Solutions8-10xSignificantly higher manufacturing costs
• CPO SolC Yield50-60%One of the key bottlenecks
• CPO Downstream Assembly Yield20-50%Another key bottleneck

Impact & implications

The report believes that recent market concerns about CPO adoption timing have led to sharp volatility in LITE, COHR, and GLW stock prices, but this concern ignores the structural trend of optical content growth. For these three suppliers, Nvidia's equity investment commitments have formed capacity binding, meaning that even if technical routes adjust, demand will not be simply substituted. From a longer-term perspective, as AI clusters evolve from scale-out to scale-up, optical content will enter an early volume ramp phase, and the current pullback provides a better entry opportunity.

Risks

• CPO adoption slower than expected, prolonging the resolution time for technical bottlenecks (yield, cost, ecosystem)
• AI capital expenditure slows down due to macroeconomic factors or technological breakthroughs (e.g., DeepSeek)
• NPO as a transitional solution extends, delaying CPO demand release
• Intensified competition among optical device suppliers, leading to increased pricing pressure
• Geopolitical/tariff policies affecting supply chain and cost structure

What to watch

• Updates to architecture roadmaps from AI chip manufacturers like Nvidia
• Progress in mass production yield and cost reduction for CPO/NPO
• Deviation of actual optical engine shipments in 2027-2028 from institutional forecasts
• Capacity expansion and customer certification progress for Lumentum, Coherent, and Corning
• Sustainability of capital expenditure for AI data center construction