Optical networking will become the next major trend in AI infrastructure

Overview

This report focuses on the optical networking upgrade within AI infrastructure. Goldman Sachs believes that as GPUs, server racks, and network architectures evolve from GB300 to Vera Rubin and Rubin Ultra, networking is no longer just a support function for compute power, but a key bottleneck for unlocking multi-chip and multi-rack collaborative computing. The report examines scale up, scale out, CPO, pluggable optical modules, silicon photonics, light sources, PCB midplanes, and copper cables, and assesses the connection architecture, dollar content, TAM opportunity, and EPS impact on the supply chain across future generations of AI server platforms.

Core views

The core view is: first, the expanding scale of AI clusters and higher bandwidth requirements will drive growth across all major network configurations rather than simple substitution; second, the Rubin Ultra platform will significantly increase connection value per rack and per compute unit, creating multiple opportunities for optical modules, CPO/NPO, PCB midplanes, and copper cables; third, CPO has strong TCO appeal in high-bandwidth, short-distance, and power-constrained scenarios, but due to maintenance, packaging, and supply-chain migration complexity, it will coexist with pluggable optical modules for the long term; fourth, rising silicon photonics penetration, tight light-source supply, and the migration from 800G/1.6T to 3.2T will extend the industry upcycle for the optical communications supply chain.

Analysis framework

Based on Nvidia GTC 2026 post-event technical roadmaps, company announcements, supply-chain research, and Goldman Sachs estimates, the report breaks down the server connectivity architecture of four Vera Rubin and Rubin Ultra configurations, and converts assumptions on ASPs and unit usage for connection components into per-rack dollar content, full-lifecycle value TAM, and EPS impact for key covered companies. The analysis also distinguishes between scale up and scale out, and builds scenario assumptions for a 25%-29% CPO penetration rate in scale out.

Methodology notes

• market_sizingTAM estimation

  Estimate market size using connection value content and server rack shipments

  The report multiplies the dollar content of connection components per rack or per compute unit by Goldman Sachs' shipment assumptions for Vera Rubin and Rubin Ultra racks to derive annual and full-lifecycle value TAM.

• scenario_analysisSpec A / Spec B scenarios

  Construct high and low scenarios using different platform configurations and CPO penetration rates

  The low-end scenario is mainly based on Spec A, while the high-end scenario is mainly based on Spec B; the Vera Rubin NVL72 CPO scale out assumption uses a 25% CPO penetration rate, and Rubin Ultra-related configurations assume a 29% CPO penetration rate.

• supply_chain_analysisSupply-chain research and EPS mapping

  Map component upgrades to supplier revenue and profit leverage

  The report maps demand growth for optical modules, CPO optical engines, CW lasers, EML, PCB/CCL, and copper cables to potential EPS contributions for covered companies, and emphasizes that the EPS contribution from a single configuration may exceed a company's full-year 2025 EPS.

Asset mapping & comparison

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

• Innolight / Eoptolink / TFC Optical

  Beneficiaries in the optical module and optical engine supply chain

  Strengths

  Benefit from 800G, 1.6T, and 3.2T speed migration, expanding demand for pluggable optical modules, and new opportunities in optical engines, FAU, and ELS created by NPO/CPO.

  Weaknesses

  As CPO replaces part of the traditional pluggable module market, product mix and customer qualification timing may create uncertainty.

  Comparison

  Compared with suppliers that rely only on traditional optical modules, manufacturers with silicon photonics and high-end optical engine capabilities are better positioned to participate in the next-generation platform upgrade.

  Risks

  CPO adoption falling short of expectations, price declines outpacing cost declines, delayed customer validation, or optical source supply constraints.

• Landmark / VPEC / Sumitomo / Mitsubishi / Furukawa

  Beneficiaries in light sources, materials, and upstream optical communications

  Strengths

  Benefit from CW lasers, EML, InP substrates, and tight optical source supply; the report expects light-source supply to remain tight through 2027.

  Weaknesses

  Capacity ramp-up takes time, and geopolitics and export controls may affect supply-demand balance and regional allocation.

  Comparison

  In a period of rapid AI server optical interconnect demand growth, upstream scarce segments may enjoy stronger pricing support than downstream assembly segments.

  Risks

  Supply moving toward balance in the second half of 2028, slower AI server specification upgrades, or weaker demand intensity due to a shift from training to inference.

• Victory Giant / WUs / EMC / Shengyi

  Beneficiaries related to PCB, CCL, and PCB midplanes

  Strengths

  In configurations such as Rubin Ultra NVL144, PCB midplanes extend from intra-server connections to rack-level connections, potentially using M9 CCL materials and high-end boards with 78 layers or more.

  Weaknesses

  Initial costs are high, and manufacturing complexity and yield requirements increase.

  Comparison

  Compared with traditional PCB applications, AI server PCB midplanes rely more on high-layer-count, high-speed materials, and system-level assembly capabilities.

  Risks

  Customers ultimately choosing copper cable or optical interconnect solutions, PCB midplane penetration coming in below expectations, or intensifying cost competition.

• Prysmian

  Global cable manufacturer and related beneficiary of connectivity infrastructure

  Strengths

  The appendix identifies it as the world's largest cable manufacturer and includes it in the discussion of connectivity technologies and the optical networking supply chain.

  Weaknesses

  Compared with the core optical module and CPO value content in AI servers, direct leverage may depend on specific product exposure.

  Comparison

  Its opportunity is more concentrated in connectivity infrastructure and cable segments rather than the core optical engine or switch-chip packaging segments.

  Risks

  AI data center build-out pace, cable price cycles, and project delivery uncertainty.

Key data

• scale out / scale up dollar content increase16x / 45xFrom GB300 NVL72 to Rubin Ultra NVL576, the dollar content of scale out and scale up connections per compute unit increases, respectively.
• optical module / optical engine TAM expansion13xWhen the serviceable scope expands from scale out to scale up, the market size per compute unit expands.
• pluggable optical module value market expansion10xFrom GB300 NVL72 to Rubin Ultra NVL576, even assuming 29% CPO penetration, the value market for pluggable optical modules in scale out still expands significantly.
• number of pluggable optical modules216 units to 2.5k unitsOn a 1.6T equivalent basis, the number per compute unit rises from 216 in GB300 NVL72 to about 2.5k in Rubin Ultra NVL576.
• total connection dollar contentUS$315k to US$9.4bnThe combined dollar content of scale up and scale out connections per compute unit disclosed in the report rises by about 29x from GB300 NVL72 to Rubin Ultra NVL576.
• connection value TAMUS$15bn to US$154bnThe value TAM increases from about US$15bn in 2026, mainly corresponding to GB300 NVL72, to about US$154bn in 2028, mainly corresponding to Rubin Ultra NVL576.
• scale up contribution in Rubin Ultra NVL576US$106bn / 69%Within the US$154bn value TAM, scale up accounts for about 69%.
• CPO contribution in Rubin Ultra NVL576US$91bn / 59%Under the assumption of 29% CPO penetration in scale out, CPO contributes about US$91bn, or 59% of the US$154bn value TAM.
• high-end CPO TAM estimateUS$97bn in 2026-28EThe report's high-end estimate suggests total CPO TAM of about US$97bn in 2026-2028E.
• 2028E scale out CPO switch demand110kThe report's high-end estimate points to about 110k scale out CPO switches in 2028E.
• silicon photonics penetration6% in 1Q24 to 45% in 4Q28Goldman Sachs expects silicon photonics adoption in optical transceivers to keep rising.
• optical module supplier gross margin48%-55%A product mix shift toward higher speeds and more silicon photonics solutions is expected to lift gross margins.

Impact & implications

For investors, AI server network upgrades will extend the beneficiary chain from traditional pluggable optical modules to CPO/NPO optical engines, CW lasers, EML, PCB midplanes, CCL, copper cables, and related manufacturing segments. Leading global CSPs are expected to adopt first, followed by Chinese CSPs, thereby supporting continued growth in the network supply chain over the next five years. For suppliers, specification upgrades, volume expansion, and speed migration may together drive revenue and EPS leverage, but the final degree of benefit will depend on the pace of technology adoption, maintenance costs, supply constraints, and customer selection.

Risks

• CPO/NPO adoption progresses more slowly than Goldman Sachs assumes, causing optical engine and CPO switch TAM to fall short of expectations.
• CPO maintenance costs, yield issues, packaging complexity, or system reliability problems slow customer adoption.
• Technology choices among pluggable optical modules, NPO, and CPO change, affecting how benefits are allocated across different supply-chain segments.
• AI server shipments, GPU platform launch timing, or CSP capex come in below expectations.
• The duration or easing pace of bottlenecks in optical sources, InP substrates, EML, and CW lasers differs from expectations.
• Price declines outpace cost improvements, compressing gross margins for optical module and related component suppliers.
• Geopolitical factors, export controls, or regional supply-chain restrictions affect upstream materials and key component supply.
• As the market shifts from training to inference, the pace of AI server specification upgrades may slow.

What to watch

• The actual specifications, rack shipment pace, and customer deployment timeline for Nvidia's Vera Rubin and Rubin Ultra platforms.
• The true penetration rate of CPO in scale out and scale up, especially whether the 25%-29% assumption can be achieved.
• The migration progress of 800G, 1.6T, 3.2T, and faster optical modules.
• Whether silicon photonics penetration in optical transceivers rises as expected from 6% in 1Q24 to 45% in 4Q28.
• The ramp-up progress of CW lasers, EML, InP substrates, and MOCVD capacity expansion, and whether optical source supply moves toward balance in the second half of 2028.
• The AI network architecture choices and order pace of leading global CSPs and Chinese CSPs.
• Cost-performance comparisons among PCB midplanes, copper cables, AOCs, NPO, and CPO under different distance, bandwidth, and power consumption scenarios.
• The actual realization of high-end product mix, gross margin, and EPS contribution at the main beneficiaries.