AI scale-up network TAM raised to approximately $70bn+, with copper interconnects remaining more durable but optical interconnects eventually arriving
AI summary card
AI scale-up network TAM raised to approximately $70bn+, with copper interconnects remaining more durable but optical interconnects eventually arriving
Morgan Stanley believes that expanding AI accelerator clusters and diversifying XPU architectures will significantly enlarge the scale-up networking opportunity. Copper still has advantages in short-reach scenarios in cost, power consumption, and latency, but optical interconnects such as CPO/NPO will gradually become critical directions after 2028.
- The scale-up market opportunity is expected to reach approximately $70bn+ by 2030, about 4x larger than the prior year's estimate of $17bn for 2029.
- Copper interconnects may still maintain a multi-generation advantage in in-rack scale-up connections due to low latency, low power consumption, low cost, and ongoing life extension from technologies such as SerDes, retimers, and PAM.
- Key catalysts for optical interconnects include multi-rack scale-up architectures, rising electrical I/O power consumption, higher bandwidth density requirements, and growing communication demand from larger AI workloads.
- NVIDIA is still expected to dominate the scale-up ecosystem. Rubin Ultra is more likely to adopt a hybrid copper-plus-optics approach, while Feynman and the period after 2029 are more likely to be the points when CPO sees meaningful volume adoption.
- The non-NVIDIA ecosystem remains unsettled, with PCIe, Ethernet, UALink, and NVLink Fusion competing in parallel. Vendors such as Broadcom, Marvell, and Astera Labs have greenfield opportunities.
- The report upgrades KEYS.N to Overweight on the view that architectural diversification will increase test and measurement demand; GLW, LITE, and COHR are viewed as clear beneficiaries once optical interconnects enter scale-up.
Report interpretation
Overview
This report is Morgan Stanley’s updated deep-dive research on scale-up networking in the North American technology and communications networking equipment sector. The core view is that continued increases in AI model complexity and cluster scale are expanding high-speed interconnects between accelerators from within-rack to cross-rack, rapidly enlarging the scale-up networking opportunity. The report estimates the 2030 market opportunity at approximately $70bn+, a substantial increase from the prior year's estimate of $17bn for 2029.
Core views
The report argues that the industry principle is “use copper when you can, use optics when you must.” Copper interconnects still have advantages in short-reach scale-up in terms of low latency, low power consumption, and low cost, while technologies such as SerDes, DSP, retimers, PAM4/PAM6, and active copper cables continue extending copper’s lifecycle. But as I/O speeds move from 100G to 200G and 400G, and AI clusters expand from 8 GPUs and 72 GPUs to 144 GPUs, 576 GPUs, and eventually more than 1,000 accelerators, electrical signal transmission distances shorten while communication distances lengthen, ultimately driving optical interconnect solutions such as NPO and CPO into scale-up. The report expects small-scale optical interconnect adoption in 2028 and sees CPO becoming more materially meaningful in scale-up after 2029.
Analysis framework
The report analyzes scale-up networking across multiple dimensions including the TAM revision upward, AI cluster size, interconnect protocols, transmission media, XPU vendor roadmaps, optical interconnect form factors, and single-stock EPS sensitivity. It compares fabric paths such as NVLink, NVLink Fusion, SUE/ESUN, PCIe, and UALink, as well as transmission technologies such as DAC/AEC/ACC, NPO, and CPO, and maps these technology paths to assets including NVIDIA, Broadcom, Marvell, Astera Labs, Lumentum, Coherent, Corning, and Keysight.
Methodology notes
Copper barrier and migration to optical interconnects
Copper is preferred for short-reach connections to reduce latency, power consumption, and cost; when multi-rack distance, bandwidth density, and electrical I/O power exceed copper’s economic and physical limits, optical interconnect solutions such as NPO and CPO are gradually adopted.
Interconnect fabric share and ecosystem competition
The report breaks scale-up fabric into NVLink, NVLink Fusion, Ethernet/SUE/ESUN, PCIe, and UALink, and evaluates how NVIDIA’s closed ecosystem, open standards, hyperscaler choices, and commercial chip availability affect vendor opportunities.
EPS sensitivity to CPO penetration
The report builds CPO scale-out and scale-up adoption scenarios for companies such as LITE, COHR, and GLW, observing changes in FY28 EPS or core EPS under different penetration-rate combinations.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- NVIDIA Corp. (NVDA.O)scale-up fabric and accelerator ecosystem leader
- Strengths
- Dominant NVLink ecosystem, leading GPU share, and NVLink Fusion expands third-party CPU, XPU, and ASIC connectivity.
- Weaknesses
- The ecosystem is relatively closed, and non-NVIDIA XPUs and open standards may divert demand in some hyperscaler scenarios.
- Comparison
- Compared with open Ethernet or UALink, NVLink has clear bandwidth and latency advantages but is less open.
- Risks
- Faster-than-expected progress in ASICs or open fabric, changes in the pace of CPO adoption, and uncertainty in hyperscaler architecture choices.
- Broadcom Inc. (AVGO.O)Ethernet/SUE and custom ASIC networking beneficiary
- Strengths
- Strong positioning in Tomahawk switch chips and Ethernet scale-up solutions, benefiting from AMD MI400/Helios and ASIC customer ramps.
- Weaknesses
- More conservative on CPO and may rely on SerDes to extend the lifecycle of copper interconnects.
- Comparison
- Compared with UALink, the Ethernet ecosystem is more mature; compared with NVLink, it is more open but still must compete on performance and software ecosystem.
- Risks
- Uncertain customer protocol choices, acceleration of UALink or NVLink Fusion, and ASIC ramp timing below expectations.
- Astera Labs Inc (ALAB.O)PCIe scale-up, Scorpio, UALink and NVLink Fusion beneficiary
- Strengths
- Scorpio-X has already shipped for AWS Trainium 3 and is expected to accelerate in 2H26, with engagement from more than ten Scorpio customers, providing exposure to open-fabric growth.
- Weaknesses
- Valuation is high, and PCIe may be a transitional fabric.
- Comparison
- Compared with Broadcom/Marvell, Astera is closer to a pure-play scale-up connectivity beneficiary.
- Risks
- Delays in commercial UALink chips, customer shifts toward Ethernet or NVLink Fusion, and valuation compression.
- Marvell Technology Group Ltd (MRVL.O)Ethernet, UALink and NVLink Fusion connectivity participant
- Strengths
- Its strategic partnership with NVIDIA strengthens its AI connectivity positioning, and it may also benefit from the Ethernet and UALink ecosystems.
- Weaknesses
- The report says its XPU visibility is limited and valuation makes the stance more cautious.
- Comparison
- Like Broadcom, it benefits from the Ethernet path; like Astera, it benefits from UALink and NVLink Fusion opportunities.
- Risks
- Share gains below expectations, customer architectures failing to converge, and valuation and execution uncertainty.
- Keysight Technologies Inc (KEYS.N)test and measurement beneficiary from architecture diversity
- Strengths
- Multiple fabrics, transmission media, and AI/semiconductor investment cycles increase testing demand, and the report upgrades it to Overweight.
- Weaknesses
- The opportunity comes more from indirect testing demand rather than direct scale-up interconnect content.
- Comparison
- Compared with optical or semiconductor suppliers, KEYS benefits from architectural divergence itself, with a more diversified path.
- Risks
- AI, semiconductors, A&D, and edge AI/6G investment falling short of expectations, leading to insufficient estimate revisions.
- Lumentum Holdings Inc (LITE.O)laser and optical component beneficiary of CPO/NPO adoption
- Strengths
- High-power CW lasers, external laser source modules, and InP laser content can benefit from NPO/CPO.
- Weaknesses
- It has recently been affected by concerns that 2028 CPO adoption may be delayed or small in scale, and near-term earnings or supply constraints may limit catalysts.
- Comparison
- Like COHR and GLW, it is a clear beneficiary of optical interconnects entering scale-up.
- Risks
- Later-than-expected CPO adoption, changes in customer design choices, and supply-chain maturity and cost pressure.
- Coherent Corp (COHR.N)optical components and lasers beneficiary
- Strengths
- Benefits from demand for optical components such as lasers, VCSELs, and InP in NPO/CPO.
- Weaknesses
- Near-term CPO adoption timing and supply constraints may affect earnings realization.
- Comparison
- Like LITE, it has direct exposure to optical components; compared with GLW, it is more focused on optical module and laser content.
- Risks
- Optical interconnect adoption below expectations, cost and manufacturing complexity, and customer concentration.
- Corning Inc (GLW.N)optical connectivity and glass/fiber beneficiary
- Strengths
- Expansion of optical connections in scale-up and scale-out creates opportunities in optical materials and connectivity.
- Weaknesses
- Value capture in CPO/NPO may be less direct than for laser or optical engine suppliers.
- Comparison
- Like LITE and COHR, it benefits from the migration to optical interconnects, but its exposure is more concentrated in connectivity and materials.
- Risks
- Delayed CPO adoption, volatility in AI data center capex, and competitive and pricing pressure.
Key data
- Scale-up market opportunityapproximately $70bn+ by 2030The report says this opportunity is about 4x larger than the prior year's estimate; the body also mentions an increase from last year’s $17bn by 2029 to $73bn by 2030.
- KEYS.N rating changeEqual-weight -> OverweightThe price target was raised from $350.00 to $400.00; the current price is $322.05 as of 2026-07-10.
- Assessment of CPO adoption timingsmall-scale in 2028, more meaningful after 2029The report believes Rubin Ultra may be more of a copper-plus-optics hybrid, while the Feynman generation is more likely to catalyze meaningful CPO adoption.
- CPO energy-saving potentialapproximately 50%-80% optical interconnect power reductionThe report says CPO reduces interconnect power consumption by moving the optical engine close to the switch ASIC in the same package, shortening the electrical path.
- Active copper cable power advantageAEC power consumption is approximately 50% lower than pluggable optical modulesTypical AEC reach is about 7-9 meters, while ACC is more oriented toward short-reach ~3 meter, low-power, and low-latency scenarios.
- AI cluster scale roadmap72 GPU Blackwell, 144 GPU Vera Rubin, 576 GPU Rubin Ultra, potential 1,152+ with FeynmanExpanding cluster scale is a key driver of the rising number and complexity of scale-up connections.
Impact & implications
From an investment perspective, the report remains constructive on scale-up networking semis: NVIDIA remains the top semiconductor pick; Broadcom benefits from the 2027-2028 ASIC ramp and networking tailwinds; Astera Labs is a pure-play beneficiary of PCIe/UALink/Scorpio but valuation is high; Marvell may gain share but valuation and XPU visibility temper the stance; Semtech has exposure to copper and linear optics but needs to prove execution. In the optical chain, GLW, LITE, and COHR are viewed as clear beneficiaries as scale-up shifts toward optical interconnects; KEYS.N is upgraded due to greater testing demand created by architectural divergence.
Risks
- CPO/NPO adoption in scale-up may arrive later than the report expects, especially if copper interconnects continue extending their life through 448G SerDes, PAM6, 200G retimers, and similar technologies.
- Hyperscalers and ASIC customers have not yet converged on Ethernet, UALink, NVLink Fusion, or PCIe, leaving significant uncertainty in vendor share.
- If AI capex, XPU shipments, and cluster-scale expansion fall short of expectations, the revised TAM of approximately $70bn+ may not materialize.
- The optical supply chain is still maturing. CPO involves complex integration of optical engines, silicon photonics, external lasers, advanced packaging, and thermal management, with elevated cost and yield risks.
- Some beneficiary names carry high valuations or still need to prove execution, which may cause a disconnect between fundamental improvement and share-price performance.
- Near-term earnings may not be a strong catalyst, as supply constraints and market concerns about the scale of 2028 CPO adoption still remain.
What to watch
- Whether OCP October brings signals on CPO/NPO roadmaps and customer adoption.
- The scale of NVIDIA Rubin Ultra, Feynman, and subsequent NVLink domains, as well as the pace of CPO deployment.
- Fabric choices of AMD MI400/Helios, AWS Trainium 3/4, Microsoft Maia, and Broadcom ASIC customers.
- Availability and performance of commercial UALink chips and switches, and hyperscaler adoption progress.
- Whether Broadcom’s next-generation SerDes and copper interconnect reach continue extending copper’s lifecycle.
- Order flow, design wins, and EPS sensitivity realization for LITE, COHR, and GLW in CPO/NPO.
- Whether KEYS.N orders and estimate revisions reflect testing demand from AI, semiconductors, A&D, and edge AI/6G.