Scale-up networking opportunity has been raised significantly; copper interconnect lasts longer while optical interconnect arrives later but with greater certainty
AI summary card
Scale-up networking opportunity has been raised significantly; copper interconnect lasts longer while optical interconnect arrives later but with greater certainty
Morgan Stanley now estimates the scale-up networking opportunity at around $70bn+ by 2030, more than quadruple versus last year’s estimate, with copper solutions still dominant in the short term and optical interconnect penetration from CPO/NPO expected to accelerate progressively after 2028.
- The scale-up market opportunity estimate for 2029 was about $17bn last year, now revised to about $73bn for 2030.
- Copper interconnect is expected to remain longer in in-rack scale-up connectivity due to lower latency, lower power, lower cost, and ecosystem openness.
- Key catalysts for optical interconnect include multi-rack scale-up architectures, rising electrical I/O power consumption, higher bandwidth density requirements, and growing AI workload communication demand.
- The report views meaningful adoption of CPO in scale-up as more likely around 2029, with NVIDIA’s Feynman generation potentially an important catalyst.
- KEYS.N was upgraded to Overweight because coexistence of multiple architectures should increase testing and measurement demand.
Report interpretation
Overview
This report updates Morgan Stanley’s framework for AI scale-up networking. As AI model complexity, accelerator cluster size, and capital spending continue to rise, scale-up networking extends from server-internal connections to rack-level and multi-rack-level links, allowing multiple GPUs/XPUs to work as one system. The report estimates scale-up opportunity at about $70bn+ in 2030 and analyzes fabric protocols, copper versus optical transmission paths, the adoption timeline for CPO/NPO, and stock-level implications.
Core views
Core views are: first, scale-up networking TAM has been significantly revised up due to multiple XPU lines including NVIDIA, AMD, AWS Trainium, Google TPU, and custom ASICs; second, architectures are unlikely to unify quickly, with NVLink, NVLink Fusion, Ethernet/SUE, UALink and PCIe likely coexisting; third, copper interconnect will not exit soon, and SerDes, PAM, DSP, retimers and active copper cables are expected to continue extending copper’s life; fourth, as multi-rack scale-up, I/O power, and bandwidth density pressures rise, optical solutions such as CPO/NPO will eventually enter the scale-up domain; fifth, from an investment perspective, NVIDIA, Broadcom, Astera, Marvell, Semtech, Lumentum, Coherent, Corning, and Keysight are exposed across different chain points, with current rating divergence driven mainly by valuation, execution, and adoption timing.
Analysis framework
The report uses a combined top-down approach of TAM revision, XPU roadmap, interconnect protocol share, transmission technology evolution, and company EPS sensitivity. On the technical side, it compares NVLink, UALink, Ethernet/SUE, PCIe, CPO, NPO, AEC and ACC. On the investment side, it maps architecture adoption timing to earnings and revenue sensitivity for semiconductor, switching, optical device, laser, glass, and test-and-measurement companies.
Methodology notes
Revisions scale-up networking opportunity based on AI capex, cluster size, and XPU unit counts.
The report revises the opportunity estimate from about $17bn for 2029 last year to about $73bn in 2030, mainly driven by larger clusters, higher interconnect complexity, and more XPU platforms.
Copper interconnect remains preferred over short distances, while optical interconnect becomes progressively necessary at multi-rack and high bandwidth density levels.
Copper still has clear advantages in latency, power, and cost, but higher SerDes rates shorten electrical signal reach and increase signal conditioning complexity, ultimately pushing adoption of NPO/CPO.
Assessing fabric ecosystems by tracing XPU roadmaps for NVIDIA, AMD, AWS, Google, and custom ASIC lines.
NVLink remains dominant in the NVIDIA ecosystem, while non-NVIDIA ecosystems may compete among PCIe, Ethernet, UALink, and NVLink Fusion, with 2027–2028 commercial silicon and customer choices as key variables.
Estimate FY28 EPS sensitivity for LITE, COHR, and GLW via a matrix of scale-out and scale-up CPO adoption rates.
The report rebuilds models for LITE, COHR, and GLW and incorporates CPO adoption rates, showing profit leverage for optical component and materials suppliers versus changes in adoption rate.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- NVIDIA Corp. (NVDA.O)Dominant driver in scale-up fabric and optical interconnect roadmap
- Strengths
- The NVLink ecosystem is closed but high-performing; its GPU share and roadmap support continued leadership of scale-up spend.
- Weaknesses
- A closed ecosystem may push hyperscalers to seek more open alternatives.
- Comparison
- Compared with non-NVIDIA ecosystems, NVLink has stronger maturity and deployment depth.
- Risks
- CPO adoption pace, custom ASIC performance of customers, and speed of open-fabric alternatives.
- Broadcom Inc. (AVGO.O)Beneficiary in Ethernet/SUE, Tomahawk switching chips, and custom ASIC networking demand
- Strengths
- Well positioned for Ethernet scale-up and ASIC customer networking pull-through, and argues that SerDes will continue to extend copper lifespan.
- Weaknesses
- More conservative on CPO; rapid customer shifts to other protocols could affect share.
- Comparison
- Compared with UALink pathways, Broadcom is more aligned with an open Ethernet ecosystem.
- Risks
- UALink commercialization success, expansion of NVLink Fusion, and shifts in customer protocol selection.
- Astera Labs Inc (ALAB.O)Beneficiary in PCIe Scorpio, UALink, and NVLink Fusion ecosystems
- Strengths
- As a higher-purity beneficiary of scale-up connectivity, Scorpio-X serves AWS Trainium 3 and has multi-customer traction.
- Weaknesses
- Valuation is high, and PCIe may be only a transitional fabric.
- Comparison
- Compared with Broadcom and Marvell, Astera is more concentrated in connectivity chips and emerging fabric.
- Risks
- UALink latency, lower-than-expected customer adoption, and valuation compression.
- Marvell Technology Group Ltd (MRVL.O)Beneficiary in Ethernet, UALink, NVLink Fusion, and AI connectivity collaborations
- Strengths
- Expected to gain share in scale-up and strengthened AI connectivity positioning through collaboration with the NVIDIA ecosystem.
- Weaknesses
- The report remains cautious due to valuation and limited XPU visibility.
- Comparison
- Compared with Astera, Marvell has a more diversified business; compared with Broadcom, XPU/ASIC visibility is seen as relatively more limited.
- Risks
- Unfavorable protocol choice, customer ramp pace, and valuation.
- SemtechBeneficiary exposure in copper interconnect and discrete optics
- Strengths
- Benefits from prolonged copper lifespan and dissimilar/discrete optical trends.
- Weaknesses
- Still needs to prove execution capability.
- Comparison
- Compared with LITE/COHR/GLW, Semtech is more exposed to copper and dissimilar optical chains.
- Risks
- Execution shortfalls, delayed optical adoption, or increased competition.
- Lumentum Holdings Inc (LITE.O)Beneficiary in laser and optical component supply chain
- Strengths
- CPO/NPO adoption would raise demand for lasers, external light sources, and optical components.
- Weaknesses
- Near-term concerns include scale of CPO adoption in 2028 and supply constraints.
- Comparison
- Together with COHR, LITE is an optical-component beneficiary, with operating leverage dependent on customer adoption and supply capabilities.
- Risks
- Later-than-expected CPO adoption, supply limitations, customer concentration, and valuation.
- Coherent Corp (COHR.N)Beneficiary in optical components, lasers, and materials supply chain
- Strengths
- NPO/CPO could drive demand for high-power CW lasers, InP content, and optical engine modules.
- Weaknesses
- The price target has limited upside versus then-current price, with ratings constrained by valuation and adoption pace.
- Comparison
- Like LITE, COHR is an optical-component beneficiary; versus GLW it is more focused on lasers and optical module components.
- Risks
- Delayed CPO/NPO penetration, insufficient margin expansion, and cyclical CAPEX pacing.
- Corning Inc (GLW.N)Beneficiary in optical interconnect materials and glass-related segments
- Strengths
- AI data-center bandwidth density and growth in optical connectivity can support long-term demand.
- Weaknesses
- In this report, the target is below the then-current price, so near-term risk/reward is not compelling.
- Comparison
- Compared with LITE/COHR, GLW is more weighted toward materials and systemic optical connectivity infrastructure.
- Risks
- Adoption pace, valuation, and non-AI business cycles.
- Keysight Technologies Inc (KEYS.N)Beneficiary in test and measurement
- Strengths
- Coexistence of multiple scale-up architectures, protocols, and transmission routes increases validation needs, supporting the upgrade.
- Weaknesses
- Subject to broader capex cycles and timing of upward revision realization.
- Comparison
- Unlike optical-component companies, KEYS benefits from architecture diversity itself rather than a single winning path.
- Risks
- AI, semiconductor, A&D, and edge AI/6G spending may be below expectations.
Key data
- 2030 scale-up market opportunityaround $70bn+; in-report text references around $73bnRevised to more than 4x versus last year’s estimate of around $17bn for 2029.
- NVIDIA cluster scale evolutionBlackwell 72 GPUs; Vera Rubin 144 GPUs; Rubin Ultra 576 GPUs; Feynman potentially over 1,000 acceleratorsLarger scale-up domains increase interconnect count, distance, and complexity.
- Optical interconnect adoption timingPossible small-scale introduction in 2028, more meaningful adoption likely after 2029The report does not see 2028 adoption as delayed, but expects more substantive CPO adoption to be more likely in the Feynman generation.
- Potential CPO power improvementOptical interconnect power could improve by about 50%-80%By moving or integrating optical engines closer to, or into, the switch ASIC package, electrical path length is shortened.
- AEC typical reacharound 7-9 metersAEC uses retimers, equalization, and clock data recovery, suitable for longer copper links but with higher power and latency.
- ACC typical reacharound 3 metersACC uses lower-power redrivers and is more oriented to short in-rack connections.
- KEYS.N rating changeEqual-weight to Overweight; target $350.00 to $400.00Coexistence of multiple architectures increases testing and measurement demand.
- COHR.N target price and priceTarget $330.00; 2026-07-10 price $324.50CPO/NPO adoption and optical device uptake are the main related drivers.
- GLW.N target price and priceTarget $180.00; 2026-07-10 price $190.89Glass and optical interconnect materials benefit, but the target is not above the then-current price.
Impact & implications
The investment implication is that the scale-up optical interconnect theme should not be interpreted as an immediate replacement of copper. The copper chain may retain a longer runway in in-rack and some between-rack scenarios; larger-scale optical opportunities may materialize as multi-rack scale-up and CPO/NPO commercialization unfolds after 2028-2029. Semiconductor and networking chip companies benefit from XPU and fabric diversification, optical component companies benefit from the eventual shift to optics, and test-and-measurement firms benefit from the validation complexity created by architecture dispersion.
Risks
- CPO/NPO adoption in scale-up could be later than expected, delaying optical-component EPS leverage.
- Continued copper innovation may further extend life, suppressing meaningful optical interconnect penetration in the near to medium term.
- Protocol choice among UALink, Ethernet, NVLink Fusion, and PCIe is uncertain and could alter supplier shares.
- Hyperscaler concentration is high; single-customer roadmap shifts can materially affect supplier opportunities.
- Supply constraints, manufacturing complexity, external laser and advanced packaging capacity may limit CPO volume ramp.
- Valuations for the relevant names are relatively elevated; if upside revisions or technology adoption do not materialize, drawdown risk exists.
What to watch
- Whether October OCP brings CPO/NPO architecture, customer, and supply-chain catalysts.
- Actual deployment scope of optical interconnect and CPO in NVIDIA Rubin Ultra and Feynman generations.
- Fabric choices by AMD MI400/Helios, AWS Trainium 3/4, Microsoft Maia, and Broadcom custom ASIC customers.
- Progress on UALink 2.0 commercial silicon, switching-chip availability, and performance validation.
- Whether copper technologies like 448G SerDes, PAM6, and 200G retimers continue to extend copper’s usable distance.
- Updates to CPO adoption assumptions, FY28 EPS sensitivity, and supply capability for LITE, COHR, and GLW.
- Whether KEYS.N follow-on orders and revenues reflect higher multi-architecture testing demand.