AI infrastructure bottlenecks across semiconductors and technology Report Interpretation

Overview

Morgan Stanley assesses where AI infrastructure bottlenecks may create the most attractive semiconductor opportunities. Its central preference is for AI compute and related constrained components, followed by networking; it remains constructive but selective on memory and favors analog as a cyclical diversification exposure.

Core views

Morgan Stanley argues that Astra GPT-6 strengthens the AI bottleneck narrative because improvements in reasoning, engineering, computer use and physical-world capabilities could make more inference workloads commercially viable. Rather than focusing solely on how much infrastructure is needed for existing demand, the report shifts attention to how improved model usefulness per dollar can increase the number, duration and complexity of workloads. In its view, this makes physical constraints—limited capacity, rising content intensity and manufacturing complexity—the central investment question. It favors businesses exposed to those constraints rather than a broad, indiscriminate AI trade. AI compute is the report's highest-conviction bucket because it combines strong unit growth with continued earnings upside across GPUs, CPUs, ASICs, leading-edge foundry capacity and supporting components. TSMC is cited as evidence: it raised 2026 revenue-growth guidance to more than 40% year on year, Morgan Stanley considers a 70-80% CAGR for its AI semiconductor business reasonable, and AI semiconductors could account for more than 30% of its 2026 revenue. TSMC's US$60-64 billion 2026 capex plan is viewed as another signal of stronger-than-expected AI demand. The opportunity extends into ABF substrates, where server, GPU/ASIC and networking demand has displaced PC demand and Morgan Stanley expects an undersupply from 2027 onward, and MLCCs, where higher AI-server power density raises content value. Taiwan monthly MLCC sales were up 36% year on year in June, while Samsung Electro-Mechanics is expected to deliver a 29% earnings CAGR in 2026-28e. The report is constructive on networking as the second-best AI infrastructure exposure, but challenges the assumption of an immediate copper-to-optics shift. It estimates the scale-up networking opportunity at more than US$70 billion by 2030, over four times its estimate a year earlier, as accelerator domains expand; NVIDIA Rubin and Rubin Ultra are cited as increasing scale-up domains to 144 and 576 GPUs, respectively. Copper should remain viable for short intra-rack links because of latency, power and cost advantages, supported by improvements in SerDes, retimers, connectors and active copper. Morgan Stanley expects a progression from copper to NPO or hybrid architectures and eventually CPO, with limited scale-up CPO adoption in 2028 and stronger adoption from 2029 onward. It forecasts CPO switch shipments rising from 23,000 units in 2026 to 59,000 in 2027 and about 200,000 by 2030; scale-out CPO switches are projected to rise from 5,000 units in 2025 to 200,000 in 2030, implying a 144% CAGR over 2023-30. Memory remains structurally supported by AI demand, but Morgan Stanley judges its relative risk/reward less attractive as the commodity cycle matures. It expects a transition to late cycle by 4Q26, with year-on-year monthly DRAM pricing peaking, sequential price increases moderating and earnings-revision momentum slowing. Third-quarter 2026 DRAM contract pricing was tracking at about 20% quarter on quarter, above the prior 15-20% expectation for both DRAM and NAND, while fourth-quarter pricing was tracking at roughly 5-10% quarter on quarter with limited visibility beyond. DRAM bit growth reached 36% year to date based on June SIA data and was tracking 31% for 2026 versus a prior 25% estimate; 2026 DRAM WFE estimates have been raised 47% over nine months. Morgan Stanley models 32-38% bit-supply growth in 2026-27e and HBM bit-supply growth above 50% in 2027-28e. The report also identifies reasons memory tightness could last longer. HBM4E's back-end-of-line complexity could delay DRAM supply growth, divert DRAM capex toward BEOL expansion in 2026-27e and delay material front-end migration and DRAM GB-shipment acceleration until 2H27. Priority equipment slots for DRAM makers could also defer NAND capacity expansion. Morgan Stanley therefore favors structural share gain, localization and downstream AI exposure over generic commodity-memory exposure, highlighting CXMT's China AI and server localization opportunity and its forecast 46% bit-shipment CAGR for 2025-28e. It prefers packaging and test beneficiaries over incremental generic front-end memory exposure, while noting that new investment coming online from late 2027 could ultimately pressure the pricing cycle. Outside AI, Morgan Stanley favors global analog as an early-cycle hedge. After more than three years near an L-shaped bottom, it sees lean channel inventory, stabilizing prices, selective tightness in mature-node and power products, improving industrial bookings and recovering utilization as signs of a demand-led upcycle rather than merely distributor restocking. It expects the recovery to continue through at least the latter half of 2026, supported by price improvement, inventory replenishment and healthier book-to-bill trends. By contrast, it remains cautious on traditional consumer end markets: global smartphone shipments are expected to decline about 13% year on year in 2026, including about 15% for Android, as higher memory costs increase OEM prices. PCs are described as stagnant to soft amid affordability pressure and heavy pre-building, while conventional enterprise-server replacement remains weaker than hyperscaler and AI-cloud infrastructure demand. Morgan Stanley emphasizes that earnings delivery, rather than valuation cheapness, should drive technology performance. It says investors now expect near-perfect execution from AI names, and that relative earnings-revision momentum must improve before technology can outperform. Consensus forecasts 56% earnings growth for AI beneficiaries in 2027 and assumes margin expansion despite decelerating 2026 revenue growth; Morgan Stanley considers 53% more likely. Its revision-breadth analysis indicates the strongest forward earnings momentum in AI compute, networking and non-AI areas, and the weakest in memory. Key headwinds include insufficient upside surprises relative to elevated expectations, a likely deceleration in AI capex growth from 2027, and macro pressures from oil, inflation, Federal Reserve rates and US elections.

Analysis framework

The report sequences the semiconductor cycle across AI compute, memory, networking, analog and consumer end markets. It combines capex and demand indicators, supply-capacity constraints, earnings-revision breadth, pricing trends, component-content intensity, technology road maps and company-specific exposure to identify where AI spending can translate into sustained earnings upside.

Methodology notes

• Industry AnalysisSupply-demand framework

  Supply-demand analysis of AI infrastructure components

  Morgan Stanley links AI workload and capex demand with limited ABF, MLCC, memory and advanced-packaging capacity to assess pricing power and earnings potential.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  AI infrastructure supply-chain transmission

  The report traces demand from accelerators and inference workloads through foundries, substrates, capacitors, memory, packaging, testing, power and networking components.

• Cycle and Business ConditionsBusiness-Cycle Inflection Analysis

  Cycle-stage and earnings-momentum assessment

  Morgan Stanley distinguishes compute growth, late-cycle memory and an early-cycle analog recovery using pricing, inventory, utilization, bookings and estimate revisions.

• Event-Driven and Behavioral FinanceExpectation Gap and Expectation Management

  Earnings deliverability versus elevated market expectations

  The report argues that AI stocks require upside surprises relative to already high expectations, so relative earnings-revision momentum matters more than a valuation-only argument.

Asset mapping & comparison

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

• NVIDIA (NVDA)

  Preferred AI-compute GPU exposure

  Strengths

  Direct exposure to AI-compute demand and expanding accelerator domains

  Weaknesses

  High earnings expectations require continued delivery

  Comparison

  Included in the report's most-preferred AI-compute bucket

  Risks

  AI-capex growth deceleration and insufficient upside surprise

• Keysight Technologies (KEYS)

  Architecture-agnostic networking and test-complexity beneficiary

  Strengths

  Exposure across optical, electrical, scale-up and scale-out networks

  Comparison

  Morgan Stanley upgraded KEYS to Overweight within its networking beneficiaries

  Risks

  Slower-than-expected network architecture deployment

• Corning (GLW)

  Networking beneficiary

  Strengths

  Relatively architecture-agnostic optical exposure

  Comparison

  Preferred alongside LITE and Coherent in scale-up networking

  Risks

  Optical deployment may transition more gradually than expected

• Coherent (COHR)

  Networking beneficiary

  Strengths

  Broad exposure across the CPO/NPO component stack

  Comparison

  Preferred scale-up networking exposure alongside GLW and LITE

  Risks

  CPO adoption is expected to remain limited in 2028

• CXMT

  Preferred China memory localization exposure

  Strengths

  China AI/server demand, HBM and DDR5 localization, and rapid capacity expansion

  Weaknesses

  Memory remains a late-cycle commodity market overall

  Comparison

  Preferred for structural share gain and localization over generic memory exposure

  Risks

  Future supply additions could pressure the memory pricing cycle

• Advanced Energy Industries (AEIS)

  Preferred US semiconductor-equipment exposure

  Strengths

  Morgan Stanley sees valuation reflecting a CY27 cycle peak; upside if WFE overshoots

  Weaknesses

  Dependent on capital-equipment cycle duration

  Comparison

  Preferred over Lam Research and KLA within US Overweights

  Risks

  WFE demand may not reach the US$230bn-plus CY27 upside scenario

• MKS Inc. (MKSI)

  Preferred US semiconductor-equipment exposure

  Strengths

  Morgan Stanley sees valuation reflecting a CY27 cycle peak; upside if WFE overshoots

  Weaknesses

  Dependent on capital-equipment cycle duration

  Comparison

  Preferred over Lam Research and KLA within US Overweights

  Risks

  WFE demand may not reach the US$230bn-plus CY27 upside scenario

• ASML

  Preferred European semiconductor-equipment exposure

  Strengths

  Dominant supplier of critical semiconductor manufacturing toolsets and beneficiary of DRAM, logic and EUV demand

  Comparison

  Morgan Stanley's Top Pick in European SPE

  Risks

  Dependence on improving 2027 DRAM and logic demand

• Texas Instruments (TXN)

  Analog-cycle indicator and exposure

  Strengths

  Provides the clearest read-through to broad-cycle analog demand

  Comparison

  Part of the report's preferred analog diversification group

  Risks

  Recovery could prove to be distributor restocking rather than sustainable sell-through

Key data

• TSMC 2026 revenue-growth guidance>40% Y/YRaised guidance, driven by stronger AI demand
• TSMC 2026 capexUS$60-64bnCited as evidence of AI-demand strength
• TSMC AI semiconductor revenue share>30% of 2026e revenueMorgan Stanley estimate
• Scale-up networking opportunity>US$70bn by 2030More than 4x the estimate a year earlier
• 3Q26 DRAM contract pricingApproximately +20% Q/QAbove the prior 15-20% expectation for DRAM and NAND
• 4Q26 DRAM pricingAbout +5-10% Q/QLimited visibility beyond the quarter
• AI-beneficiary 2027 consensus earnings growth56%Morgan Stanley considers 53% more likely
• Global smartphone shipmentsApproximately -13% Y/Y in 2026Android shipments expected to decline roughly 15%

Impact & implications

Morgan Stanley's preferred positioning is a barbell of constrained AI-compute and networking exposures plus early-cycle analog recovery. It favors capacity- and complexity-linked beneficiaries, is selective on memory despite strong fundamentals, and remains cautious on consumer hardware demand pressured by higher component costs.

Risks

• AI beneficiaries may fail to produce sufficient upside surprises relative to elevated market expectations.
• AI capex growth is likely to decelerate from 2027, even if it remains positive.
• Oil, inflation, Federal Reserve rates and US elections are identified as macro headwinds.
• Late-cycle memory supply growth and new investment from late 2027 could pressure pricing.
• Analog recovery could reflect short-lived distributor restocking rather than sustainable demand.

What to watch

• Relative earnings-revision momentum and companies' ability to meet elevated earnings expectations.
• TSMC AI demand, capex execution and demand across supporting ABF and MLCC components.
• The timing of DRAM pricing moderation, memory supply additions and HBM4E-related capacity constraints.
• Progression from copper to NPO and CPO architectures, including the pace of CPO adoption after 2028.
• Analog inventory replenishment, pricing, industrial bookings, utilization and book-to-bill trends.
• Consumer demand trends as higher memory costs affect smartphone, PC and electronics affordability.