Optical interconnect is expected to become the next major trend in AI infrastructure
AI summary card
Optical interconnect is expected to become the next major trend in AI infrastructure
Goldman believes that upgrading AI servers from GB300 to Vera Rubin / Rubin Ultra will significantly raise the connection-dollar value of scale out and scale up, driving optical modules, CPO, copper cables, PCB midplanes, and light-source supply chains into a multiyear growth cycle.
- Goldman expects scale out and scale up connection value per compute-dollar to rise by 16x and 45x, respectively, from GB300 NVL72 to Rubin Ultra NVL576.
- Total scale up and scale out connection-value TAM is expected to rise from US$15bn at GB300 NVL72 to US$154bn at Rubin Ultra NVL576, an increase of about 9x.
- Within US$154bn TAM, scale up contributes about US$106bn, or 69%; CPO contributes about US$91bn, or 59%, assuming a 29% scale out CPO penetration rate.
- Goldman remains positive on its optical communication and PCB coverage names, including Innolight, Eoptolink, TFC Optical, Landmark, VPEC, Sumitomo, Mitsubishi, Furukawa, Victory Giant, WUS, EMC, and Shengyi.
Report interpretation
Overview
This report focuses on optical interconnect and network connectivity upgrades in AI infrastructure. Goldman believes that as AI server rack counts rise, per-rack compute increases, and network bandwidth demand expands, network connectivity will move from pure scale out to more complex scale up and cross-rack connectivity, increasing value in optical modules, CPO, copper cables, PCB midplanes, CW lasers, EML, and silicon photonics.
Core views
The core view is that network connectivity is the key to unlocking AI chip compute power, and in the future different connectivity architectures are not in a simple substitution relationship but instead grow together. New platforms such as Rubin Ultra will drive a significant increase in connection-dollar content; optical solutions are shifting from scale out toward scale up, CPO has power and TCO advantages in high-bandwidth and short-distance scenarios, and pluggable optical modules are expected to coexist with NPO / CPO while benefiting from speed upgrades and increasing silicon photonics penetration.
Analysis framework
The report is based on Nvidia roadmaps, GTC 2026 announcements, supply-chain checks, and Goldman calculations, decomposing connection architectures, per-rack dollar content, server rack shipment assumptions, TAM opportunities, and potential EPS impacts for key suppliers across GB300, Vera Rubin, and Rubin Ultra rack configurations.
Methodology notes
Estimate TAM by multiplying per-rack connection value by server rack shipments.
The report uses connection-content values for different GPU platforms and specifications, combined with GS rack-shipment assumptions, to derive value TAM for scale up, scale out, CPO, and optical modules.
Estimate supplier EPS or net income contribution using TAM, market-share, and margin assumptions.
In Exhibit 13, the report combines value TAM, supplier market-share assumptions, and margin assumptions to assess potential earnings contribution from key suppliers under a single server configuration.
Compare scale up and scale out connection architectures and component choices across different GPU platforms.
The report breaks down configurations such as Vera Rubin NVL72, Rubin Ultra NVL144, and Rubin Ultra NVL576, distinguishing applicable scenarios for copper cables, PCB midplanes, pluggable optical modules, and CPO under different distance, bandwidth, and cost constraints.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- Innolight, Eoptolink, TFC OpticalOptical modules and optical engines beneficiaries
- Strengths
- Benefit from 800G, 1.6T, and 3.2T speed upgrades, silicon photonics expansion, and incremental opportunities in optical engines in NPO / CPO.
- Weaknesses
- Pluggable optical module prices are expected to fall as scale rises, and the adoption pace for CPO / NPO is uncertain.
- Comparison
- Compared with traditional low-speed optical modules, high-speed silicon-photonic and CPO-related products have higher value content.
- Risks
- Customer adoption may be slower than expected, reliability requirements may delay technology shifts, and light-source supply bottlenecks may emerge.
- Landmark, VPEC, CW lasers and EML suppliersLight-source supply-chain beneficiaries
- Strengths
- AI server growth, speed upgrades, and expanding optical connectivity are increasing demand; InP substrate and capacity constraints support continued tightness in supply-demand.
- Weaknesses
- Capacity expansion takes time, and geopolitical factors and export controls may affect supply-chain stability.
- Comparison
- CW lasers have incremental opportunities in silicon photonics and CPO use cases, while EML will continue to coexist in long-distance transmission and mature reliability scenarios.
- Risks
- Supply-demand may balance after 2H28, or AI server specification upgrades may slow.
- Victory Giant, WUS, EMC, ShengyiPCB / CCL supply-chain beneficiaries
- Strengths
- Rubin Ultra scale up may lift demand for PCB midplanes and higher-layer M9 CCL materials.
- Weaknesses
- Initial costs are higher and requirements for manufacturing complexity and yield are tighter.
- Comparison
- PCB midplanes offer a better cost-performance profile than traditional wiring approaches for high-bandwidth, shorter distances and assembly efficiency.
- Risks
- Architectural roadmap changes, slower-than-expected customer adoption, and weaker-than-expected cost declines.
- Copper cablesShort-distance scale up connection option
- Strengths
- In rack or short-distance links, they retain low-cost, low-power advantages, with DAC / ACC / AEC evolving toward longer distances.
- Weaknesses
- Signal quality degradation becomes significant as distance increases or rates rise.
- Comparison
- For short-distance links, copper has a cost advantage over fiber, but optical solutions are superior in high-bandwidth, longer-distance scenarios.
- Risks
- Gradual substitution by optical connectivity in higher-bandwidth and cross-rack scenarios.
- CPOCore technology for high-bandwidth short-distance optical connectivity
- Strengths
- Shortens electrical transmission paths, lowers power and latency, saves DSP and retimer costs, and improves integration and form-factor efficiency.
- Weaknesses
- Packaging is complex, maintenance costs are higher, and supply chains require broad architecture migration rather than component-by-component upgrades.
- Comparison
- Compared with pluggable optical modules, CPO is more attractive on TCO in high-bandwidth scenarios such as 6.4T and 12.8T, but carries larger reliability and maintainability challenges.
- Risks
- Technology maturity, yield, maintenance cost, customer adoption pace, and ecosystem compatibility.
Key data
- Total connection-value TAMUS$154bnTotal value TAM for scale up plus scale out under the Rubin Ultra NVL576 cycle, per Goldman; about 9x higher than US$15bn for GB300 NVL72.
- Scale up TAM share69% / US$106bnWithin total TAM of US$154bn, scale up contributes about US$106bn.
- CPO TAM contributionUS$91bn / 59%Assuming scale out CPO penetration of 29%, CPO contributes about 59% of total TAM.
- Per-compute-unit dollar content increasescale out 16x / scale up 45xGoldman estimate from GB300 NVL72 to Rubin Ultra NVL576.
- Pluggable optical module count216 units to 2.5k unitsThe number of 1.6T-equivalent pluggable optical modules per compute unit rises from 216 in GB300 NVL72 to about 2.5k in Rubin Ultra NVL576.
- Silicon photonics penetration6% in 1Q24 to 45% in 4Q28The report expects silicon photonics adoption in optical transceiver modules to continue rising.
- Optical module supplier gross margins48%-55%The report expects product mix improvement to be supported by higher-speed products and silicon photonics migration.
- Light source supply-demandtight through 2027; more balanced in 2H28EML and CW laser supply is expected to remain tight through 2027 due to AI server growth, InP substrate constraints, and ramping capacity.
Impact & implications
The investment implication is that value-chain upgrading in AI infrastructure is extending from GPU logic itself to the network connectivity system. Optical modules, CPO optical engines, CW lasers, EML, PCB midplanes, CCL, and copper cable suppliers may obtain significant incremental revenue and EPS leverage. Leading global CSPs are expected to drive the migration first, with China CSPs following, which may support strong network supply-chain growth for the next five years.
Risks
- The adoption pace for CPO, NPO, and silicon photonics may be slower than Goldman expects.
- If AI server shipments or specification upgrades fall below expectations, connection TAM and supplier EPS leverage could weaken.
- Light source, InP substrate, and key manufacturing capacity bottlenecks may constrain short-term deliveries.
- CPO maintenance costs, reliability, and supply-chain migration difficulty may affect customer adoption.
- Geopolitics, export controls, and China-related government restrictions may disrupt light-source and semiconductor material supply.
- Technology-路线 competition among pluggable optical modules, CPO, copper cables, and PCB could alter value allocation.
What to watch
- Actual launch timing, rack shipments, and customer adoption pace for Vera Rubin and Rubin Ultra platforms.
- Whether actual CPO penetration in scale out and scale up remains close to the assumed 25%-29%.
- Progress of migration from 800G to 1.6T, 3.2T, and above.
- Whether silicon photonics penetration in optical transceiver modules rises from 6% in 1Q24 to 45% in 4Q28.
- The expansion pace of EML, CW lasers, InP substrates, and MOCVD capacity.
- Procurement and deployment progress by global CSPs and China CSPs for high-bandwidth network architectures.