AlphaLens
Research
研报4月30日 · Morgan Stanley

Greater China Technology Semiconductors: Cloud Semis: Read-across from Top 4 US CSP 1Q26CY Earnings Calls

中文EN⚠ quality lint: see notes

Cloud Semis Read-Across: Top 4 US CSP Capex Hikes Signal Sustained Demand for Aspeed & Montage into 2027

Core Thesis

The four largest US cloud service providers (Google, Meta, Microsoft, Amazon) collectively raised 2026 capex guidance to nearly $700 billion and explicitly flagged further growth into 2027. Their cloud revenue acceleration and custom silicon order backlogs confirm that AI/cloud infrastructure investment is generating strong returns, not merely speculative buildout. This sets up a multi-year demand tailwind for Greater China cloud semiconductor names—Aspeed (X86 server management controllers) and Montage (memory interface and data center connectivity chips). The market is overly focused on GPU/ASIC competition and underestimates the persistent X86 server replacement cycle and the revenue visibility from custom chip commitments. Current valuations do not fully price in 2027 growth, supporting an overweight stance.

Evidence Chain

1. Capex Trajectory: 2026 Aggregation Nears $700 Billion, 2027 Stated as “Significantly Higher”

  • Google raised 2026 capex guidance to $180–190 billion (from $175–185 billion) and expects a “significant” year-over-year increase in 2027.
  • Meta lifted its 2026 range to $125–145 billion (from $115–135 billion), citing higher memory costs and data center expansion.
  • Microsoft guided ~$190 billion for CY2026, with CQ4 capex expected above $40 billion.
  • Amazon retained its $200 billion 2026 plan, backed by multi-year customer commitments.

Investment implication: The sequential upgrade pattern across all four hyperscalers, combined with forward-looking commentary, implies that cloud semi procurement (server BMCs, memory buffers, retimers) will remain elevated through at least 2027. Aspeed and Montage are directly leveraged to server volume growth and memory channel upgrades.

2. Cloud Revenue & Order Backlogs Confirm Demand Durability

  • Google Cloud revenue grew 63% YoY to ~$20 billion, with backlog nearly doubling sequentially to $462 billion—50%+ expected to convert within 24 months.
  • Amazon disclosed over $225 billion in customer revenue commitments tied to Trainium, with newer generations nearly fully subscribed and Trainium4 capacity pre-reserved.

Investment implication: Unlike the 2021–2022 capex build that later saw pullbacks, these backlogs and commitments represent contracted, multi-year revenue stream for the CSPs themselves, reinforcing that infrastructure capex is backed by real demand. This reduces the risk of abrupt order cancellations for upstream semiconductor suppliers.

3. Custom Silicon Gains Traction, but X86 Servers Remain the Base-Load Growth Driver

  • Microsoft’s Maia 200 AI accelerators are live in the US, delivering 30%+ better tokens per dollar.
  • Meta deploys a heterogeneous stack including its own MTIA ASIC alongside NVIDIA/AMD GPUs.
  • Amazon’s Trainium series is locked in via long-term contracts, effectively pre-selling capacity.

While custom ASICs (Trainium, Maia, MTIA) are growing rapidly, the market overlooks that X86 server units within these same data centers are also being refreshed. Aspeed’s BMC (Baseboard Management Controller) is embedded in virtually every X86 server, and Montage’s memory interface chips benefit from DDR5 adoption in new server builds. The CSPs’ commentary about “unprecedented internal and external AI compute demand” implies that general-purpose server fleets must expand in parallel to support data ingestion, storage, and pre/post-processing for AI workloads.

Investment implication: The X86 server refresh cycle, combined with DDR5 penetration, provides an incremental volume driver beyond AI-specific chips. Aspeed and Montage capture this upside without competing in the GPU/ASIC battle.

Key Risks

  • Capex reduction risk: If hyperscaler revenue growth decelerates or regulatory pressures (e.g., antitrust, energy restrictions) force capex cuts, orders for server components could be delayed or reduced.
  • Technology displacement: Persistent GPU/ASIC architectural shifts could alter server design requirements, potentially shifting demand from standard BMC/chipset solutions to more integrated or proprietary architectures.
  • Geopolitical constraints: Greater China semiconductor companies face ongoing restrictions on advanced process access and potential customer diversification limits, which could impair their ability to serve US-based CSPs or scale new products.

Valuation & Trade Implications

Aspeed trades at ~22x 2027E P/E and Montage at ~18x 2027E P/E, based on consensus estimates that have not yet fully absorbed the 2027 capex trajectory. With the four CSPs’ 2026–2027 capex growth implying a ~10–15% compound annual increase in server unit demand, and both companies having high incremental margins (>50% for Aspeed, >40% for Montage), we see 20–25% EPS upside to current 2027 estimates.

We recommend overweight positions in Aspeed and Montage. The primary catalyst is the upcoming server refresh cycle tied to Intel’s Granite Rapids and AMD’s Turin launches in late 2026, which will accelerate BMC and DDR5 ordering. Second, any additional positive pre-announcement from Montage regarding DDR5 MCRDIMM qualification at a major CSP would further derisk estimates.

Given revenue visibility from cloud backlogs and custom silicon commitments, the risk/reward skews favorably. The key sell-side risk is a macro-driven capex pause, but the current order backlog and customer commitments provide a 12–18-month buffer that makes near-term ordering less binary.

This memo is based solely on publicly available earnings call transcripts and company disclosures. No proprietary Morgan Stanley research content is used beyond the provided Evidence Pack.