Morgan Stanley: Computex Notes Bullish on Cloud Chips & Legacy Memory; PC Semi Outlook Beats Expectations

Overview

This report updates views on the Greater China technology semiconductor sector based on research from the 2026 Computex exhibition. The core conclusion is that data points to a stronger industry outlook, particularly in cloud semiconductors and legacy memory. The report suggests recent share price volatility provides a good entry point for investors. While Agentic PCs have not yet seen mass adoption, PC semiconductor orders and gross margins in 2H26 are expected to remain stable, with an outlook better than market fears. The firm explicitly recommends Overweight positions in cloud semiconductor and legacy memory names, while holding structural views on PC semiconductors.

Core views

In cloud semiconductors, demand is strong but supply is constrained. GPU servers continue to ramp up with minor spec adjustments; CPU server demand is primarily driven by AI and Agent applications. Notably, lower DRAM content in ARM servers is more of a supply-side issue rather than weakening demand; meanwhile, CXL adoption in x86 CPUs is increasing RDIMM usage. Pull-in timing for CPO-related components may advance from Q1 2027 to Q4 2026. In PC semiconductors, the Agentic PC thesis holds but pricing is high. Checks indicate AI PCs with N1X chips are priced at $2,899, and N1 models at $1,799; desktop configurations could reach 128GB DRAM and 4TB SSDs with starting prices of $10,000. Despite cost pressures from further foundry price hikes in Q3, downstream customers appear willing to accept memory price increases, implying PC semi gross margins may stabilize in 2H26, resulting in an overall outlook better than pessimistic market expectations. The Legacy Memory sector faces greater shortages. Legacy NAND (SLC) has a stronger growth outlook over the next two years, mainly driven by its application in high-performance data center eSSDs and enhancing HDD performance. Additionally, channel prices for high-density NOR flash have risen from $1 to $8, primarily driven by doubled content in VR devices. Kioxia noted that data center NAND bit demand will grow at a 34% CAGR from 2025-2031, with the AI inference segment reaching 56%.

Analysis framework

The report combines 'top-down' industry cycle analysis with 'bottom-up' company/technology validation. First, through primary channel checks at Computex, it gathers micro-data on CPU/GPU specs, AI PC pricing, and memory configurations to assess demand strength and cost pass-through capabilities across the value chain. Second, incorporating keynotes from industry leaders (e.g., NVIDIA CEO, Kioxia), it distills tech trends like 'Agentic AI' and 'GPU Direct Storage' to analyze their pull-through effect on specific niche products (e.g., SLC NAND, BMC chips, memory interface chips). Finally, based on supply-demand gaps (e.g., ARM server DRAM supply constraints, tight SLC capacity) and pricing trends (NOR price hikes, customer acceptance of memory inflation), it derives earnings prospects for sub-sectors and maps them to individual stock ratings.

Methodology notes

• Industry Analysis FrameworkSupply-demand framework

  Supply-Demand Mismatch Analysis

  The report analyzes mismatches between supply bottlenecks and demand growth in specific segments (e.g., ARM server DRAM, SLC NAND) to judge price trends and industry sentiment. For instance, it identifies low ARM server DRAM content as a supply issue rather than weak demand, correcting market misperceptions.

• Industry Analysis FrameworkValue Chain Transmission

  Cost Pass-Through and Bargaining Power

  Analyzes whether mid-stream semiconductor design firms can pass upstream foundry price hikes to downstream terminal brands. The report finds PC brands are willing to accept memory price increases, inferring stable gross margins for PC semi makers—a typical value chain profit distribution analysis.

• Industry Analysis Framework

  Incremental Demand Driven by Tech Architecture Shifts

  The report focuses on how new tech architectures (e.g., GPU Direct Storage, Agentic AI) create new hardware demands (e.g., SLC NAND as a caching layer, high bandwidth needs for Vera CPU). This methodology emphasizes identifying non-consensus incremental markets via technological evolution paths rather than relying solely on traditional cyclical demand.

Asset mapping & comparison

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

• Aspeed

  Beneficiary: Vera CPU server ramp increases BMC chip demand

  Strengths

  Higher BMC content value; direct beneficiary of millions of Vera CPU deployments

  Comparison

  Better positioned than Montage (neutral on Vera CPU)

  Risks

  Server shipments below expectations

• Montage

  Neutral/Beneficiary: Neutral on Vera CPU for memory interface chips; PCIe interconnect chips benefit from Agentic AI

  Strengths

  Global leader in memory interface chips; China leader in PCIe interconnect chips

  Weaknesses

  Vera CPU uses SOCAMM instead of RDIMM, reducing demand for some legacy products

  Comparison

  Core cloud semiconductor name

  Risks

  Technology route change risk

• Winbond / Macronix / Gigadevice

  Beneficiary: Legacy NAND (SLC) shortage and GPU Direct Storage demand

  Strengths

  Unique position of SLC NAND in data center eSSDs and GPU direct storage

  Comparison

  Collectively bullish on Legacy Memory sector

  Risks

  Capacity expansion leading to price declines

• Parade / Elan

  Beneficiary: PC semi companies with new growth drivers

  Strengths

  Possess new growth catalysts; better positioned to handle rising costs

  Comparison

  Preferred over Realtek, Novatek, Asmedia

  Risks

  Shrinking PC market demand

• Realtek / Novatek / Asmedia

  Negative/Cautious: Lack new growth drivers; facing cost pressures

  Weaknesses

  Lack new growth points; constrained by compressed demand and high costs

  Comparison

  Rated lower than Parade and Elan

  Risks

  Gross margin pressure

Key data

• Data Center NAND Bit Demand CAGR (2025-2031)34%AI inference segment CAGR is 56%
• High-Density NOR Flash Channel Price Change$1 -> $8Significant increase, primarily driven by doubled content in VR devices
• AI PC (N1X) Estimated Price$2,899Entry-level pricing indicated by checks
• AI PC (N1) Estimated Price$1,799Entry-level pricing indicated by checks
• High-End AI Desktop Configuration128GB DRAM + 4TB SSDStarting price approx. $10,000
• Vera CPU Core Bandwidth14 GB/s per core3x higher than DDR5 x86 CPUs, 40% lower latency

Impact & implications

For cloud semiconductors, large-scale deployment of NVIDIA Vera CPUs (estimated 2.5m-4m units) directly benefits BMC chip supplier Aspeed, while impact on Montage (using SOCAMM vs. RDIMM) is neutral, though Montage's leadership in PCIe interconnect chips remains supported by distributed computing needs for Agentic AI. For legacy memory makers (Winbond, Macronix, Gigadevice), widespread adoption of GPU Direct Storage technology will make their SLC NAND products a critical link in data center storage hierarchies, bringing long-term structural growth opportunities. For PC semiconductors, near-term caution is needed regarding potential volume suppression from high pricing, but medium-term, companies benefiting from new growth drivers (e.g., Parade, Elan) should outperform the industry average, while those lacking new narratives (e.g., Realtek, Novatek) may face valuation pressure.

Risks

• High AI PC pricing suppressing market demand
• Further foundry cost increases eroding margins
• Slowdown in cloud server capex
• Rapid capacity expansion in legacy memory leading to price drops

What to watch

• Actual 2H26 PC semi orders and gross margin performance
• Actual Vera CPU shipment volumes and server configuration mix
• Penetration progress of SLC NAND in data center eSSDs
• Confirmation of CPO component pull-in to Q4 2026