Morgan Stanley: Strong AI Interconnect Demand; Latest Updates on ALAB/MRVL/INTC
AI summary card
Morgan Stanley: Strong AI Interconnect Demand; Latest Updates on ALAB/MRVL/INTC
Morgan Stanley meeting minutes indicate that AI infrastructure is accelerating toward interconnect-intensive architectures; Astera Labs has established advantages in PCIe and optical interconnect, Marvell is gaining customer favor through its full-stack layout, and Intel's foundry and CPU execution remain key watchpoints.
- AI compute demand drives continuous growth in interconnect content, with single-XPU interconnect value at approximately $1,000 and still rising
- PCIe protocol retains growth resilience over the coming years, especially in card-based Chinese AI markets
- Marvell is viewed as an important alternative to NVIDIA, leveraging full-stack capabilities from switch chips to optical modules
- Intel's new CEO emphasizes foundry execution, with 14A process yield improvements and risk production targeted for 2028
- Institutions are optimistic on ALAB's long-term prospects but express concern over MRVL's current valuation at 2.5x NVDA
Report interpretation
Overview
This report summarizes key discussion points from Morgan Stanley's recent meetings with management at Astera Labs (ALAB), Marvell (MRVL), and Intel (INTC). The core view is that the development focus of AI infrastructure is shifting from pure compute units toward highly complex interconnect systems, creating significant incremental opportunities for vendors offering full-stack interconnect solutions. ALAB has a clear and competitive positioning in PCIe and optical interconnect; MRVL benefits from supply chain diversification trends, with its full-stack product portfolio emerging as a new competitive advantage; Intel is at a critical juncture in its foundry strategy transformation, with management demonstrating greater confidence in the 14A process and CPU market share recovery.
Core views
Interconnect technology becomes the core driver of the AI era: Management teams generally agree that AI infrastructure is becoming increasingly interconnect-intensive. Currently, interconnect content per XPU (accelerator processor) is approximately $1,000, with room for growth as system complexity increases. While some workloads may migrate to other protocols, PCIe will maintain growth momentum over the next several years. Particularly in China, where a large number of AI systems adopt card-based architectures, PCIe holds a native advantage; although the scale is smaller than the US, it remains an important growth point. Astera Labs' strategic focus and competitive moat: ALAB positions itself as the 'NVSwitch' in an open ecosystem, with comprehensive coverage across PCIe, UALink, and optical interconnect. Management notes that the switch chip market is not winner-take-all, with initial deployment expected next year and larger-scale volume in 2028. In optical interconnect, NPO (passive optical) may see earlier deployment in 2027, and if successful could even delay CPO adoption. Despite market enthusiasm for optical interconnect, intra-rack copper connections will remain important due to cost and maturity advantages. ALAB's acquisition of aiXscale has provided the capability to produce complete optical engines, building a defensive moat. Marvell's full-stack advantage and valuation divergence: Marvell positions itself as one of the few suppliers capable of providing full-stack services from DCI, Die-to-Die IP, switch chips, optical interconnect to custom silicon. In a supply-constrained environment, this integration capability has become its core advantage against competitors (such as Broadcom). In the Scale-up (intra/inter-die interconnect) space, Marvell is expected to lead the commercial optical interconnect market and benefit from customer trends seeking non-NVIDIA alternatives. However, Morgan Stanley notes that despite Marvell's exciting business prospects, its current valuation reaches 2.5x that of NVIDIA, which raises institutional concerns for a company whose long-term growth has not exceeded NVIDIA's. Intel's foundry transformation and product roadmap: New CEO Lip-Bu Tan has shifted focus toward business growth and product roadmap execution. On the CPU side, the reintroduction of SMT and P-core/E-core differentiation; management believes that if executed smoothly, market share losses could be halted in approximately 9 months. On the foundry side, the 14A process is a critical milestone, with defect density and yield metrics currently improving; the target is to reach lower defect density by Q1 next year. Risk production is scheduled for 2028, with mass production in 2029. Additionally, EMIB advanced packaging technology is viewed as a 'secret weapon,' but also faces complexity and reliability challenges.
Analysis framework
Morgan Stanley's analytical logic follows the main thread of 'technology evolution - commercial deployment - financial mapping.' First, by dissecting technical bottlenecks in AI infrastructure (such as the interconnect wall after compute saturation), it identifies interconnect bandwidth and power consumption as key variables constraining system performance. Second, it tracks the commercialization timelines and customer validation progress of various sub-technologies (such as PCIe, CPO, NPO, UALink) to judge market share shifts among different vendors. Finally, it evaluates companies' full-stack capabilities (whether they possess end-to-end delivery capability from electrical to optical) to assess their risk resilience and bargaining power in supply chain restructuring, and on this basis examines whether current valuation multiples have already priced in future growth expectations.
Methodology notes
Upstream-Midstream-Downstream Transmission
The report analyzes the value transmission process from the underlying physical layer (optical/copper interconnect) to the protocol layer (PCIe/CXL), and then to the system layer (switch/server), helping to understand why the interconnect segment has captured an increasingly large share of value in AI development.
Value Chain Analysis
By comparing the coverage scope of each company across design, manufacturing, and packaging (such as Marvell's full-stack layout vs ALAB's interconnect focus), it reveals how 'vertical integration' translates into core competitiveness for enterprises amid supply chain fragmentation.
Inflection Point Analysis
The report closely tracks the yield ramp milestones for Intel's 14A process (2027-2028) and the deployment timeline for NPO technology (2027); these specific time nodes constitute potential inflection points for related companies' earnings explosions.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- Astera Labs (ALAB.O)Core AI interconnect (PCIe, optical interconnect, switch chips) beneficiary, positioned to capture exponential growth in AI system interconnect bandwidth and power consumption requirements.
- Strengths
- Holds a position analogous to 'NVSwitch' in the PCIe ecosystem; possesses full-stack production capabilities spanning electrical IC, PIC, and optical engines; clear strategic focus.
- Weaknesses
- Stock has risen substantially, with extremely high market expectations; iteration path for optical interconnect technologies (NPO/CPO) carries uncertainty.
- Comparison
- Compared to giants like Broadcom, ALAB is more focused on the interconnect segment, but faces broader competition in areas such as Ethernet switching.
- Risks
- AI capital expenditure slowdown; CXL or 1.6T port speed delays hindering new product rollout; intensifying competition leading to market share loss.
- Marvell Technology (MRVL.O)Full-stack interconnect and custom chip supplier, benefiting from AI customization demand and supply chain de-concentration trends.
- Strengths
- Extremely broad product portfolio (spanning optical, electrical, switching, and custom silicon), with strong integration and customer stickiness advantages amid supply constraints.
- Weaknesses
- Current valuation (2.5x NVIDIA) is considered insufficiently supported by earnings growth momentum; traditional storage and networking businesses may drag on overall performance.
- Comparison
- As the most important interconnect alternative provider outside of NVIDIA, its integration of software, silicon, and optics is the key differentiator from pure hardware vendors.
- Risks
- AI custom project scale falling short of expectations; storage and networking market demand decline; ongoing pressure on enterprise data center spending.
- Intel Corporation (INTC.O)Typical representative of IDM-to-foundry transformation, facing dual challenges of CPU market competition and advanced process breakthroughs.
- Strengths
- Possesses massive x86 ecosystem foundation; new management has strengthened execution on foundry and CPU roadmaps; advanced packaging EMIB holds significant potential; near-term CPU shortages provide positive catalyst.
- Weaknesses
- Foundry business still in investment phase with heavy cost structure; advanced process (14A) yield and customer acquisition still need validation; memory supply subject to external constraints.
- Comparison
- Catching up with TSMC in the foundry space is extremely difficult, but Intel retains unique x86 licensing barriers in the CPU domain.
- Risks
- Foundry business failure leading to cost失控; AMD and other competitors capturing more server and desktop market share; advanced process node delays.
Key data
- Single-XPU Interconnect Content~$1,000Current level, with continued growth as AI infrastructure becomes more interconnect-intensive
- Marvell Custom Business Revenue Expectation~$4bn to ~$10bnReflects explosive growth in AI custom chip demand
- Intel 14A Process Target Yield/Defect Density~40% yield / 0.5 defect densityEarly metrics; target is to reduce defect density to 0.1-0.2 by Q1 next year
- Marvell Valuation Multiple2.5x NVDAMorgan Stanley considers this valuation elevated relative to its historical growth rate
Impact & implications
For investors, the focus of AI investment is shifting from single GPU procurement toward building complete interconnect ecosystems. Astera Labs and Marvell, as core beneficiaries of interconnect infrastructure, have relatively high long-term growth certainty, but vigilance is warranted against Marvell's current high valuation correction risk. Whether Intel's foundry story can be delivered will directly determine the magnitude of its valuation re-rating in coming years; in the near term, the earnings recovery from CPU shortages is worth watching.
Risks
- AI capital expenditure pause or significant reduction, leading to stalled data center investment
- New technologies (such as CXL, NPO) experiencing significant delays in commercialization
- Intensifying industry competition eroding interconnect chip vendors' market share
- Intel foundry yield falling short of expectations, preventing mass production
- AMD and other competitors adopting aggressive strategies in the CPU market, squeezing Intel's share
What to watch
- Astera Labs Amazon Trainium 3 Scale-up ramp progress and subsequent UALink/NVLink Fusion orders
- Marvell's actual market share in the Scale-up optical interconnect market and adoption of non-AVGO alternative solutions
- Intel 14A process internal testing in 2027 and yield data for 2028 risk production
- Global hyperscalers' ultimate technology route selection in AI interconnect architecture (e.g., preference for CPO vs NPO)