C
ChaoBro

日本語版: Morgan Stanley Again Raises AI Capex Forecast: $805B in 2026, $1.1T in 2027

日本語版: Morgan Stanley Again Raises AI Capex Forecast: $805B in 2026, $1.1T in 2027

この記事は日本語版です。言語ルートを完全にするため、本文は既存の標準原稿をベースにしています。


Intelligence Brief

Morgan Stanley has once again raised its forecast for global AI infrastructure capital expenditure. The five hyperscale tech giants — Amazon, Google (Alphabet), Meta, Microsoft, and Oracle — are expected to reach a combined AI Capex of $805 billion in 2026, with a new 2027 forecast soaring to $1.1 trillion.

Data Breakdown

2026 Capex Budgets by Company

Company 2025 Actual 2026 Budget YoY Increase Primary Allocation
Amazon $83B ~$200B +141% AWS AI infrastructure, Anthropic investment
Microsoft $80B ~$190B +138% Azure AI, OpenAI investment, data centers
Alphabet $75B $180-190B +140-153% Google Cloud AI, TPU, DeepMind
Meta ~$65B ~$135B +108% Llama training, AI advertising, AR/VR
Oracle ~$20B ~$80B +300% OCI GPU clusters, AI databases
Total ~$323B ~$805B +149%

Key Trends

1. Everyone Doubling Down Every giant has doubled its Capex. This is no longer a single company's strategic bet — it's a consensus action across the entire industry.

2. Oracle's Surge Oracle's increase is the most dramatic (+300%), leaping from roughly $2B to $8B. Its OCI (Oracle Cloud Infrastructure) is aggressively building large-scale GPU clusters, aiming to carve out a share of the AI cloud services market.

3. What Does the $1.1T 2027 Forecast Mean? If the 2027 projection holds true, the five giants' combined AI Capex will exceed the global semiconductor industry's total revenue in 2025 (approximately $600B). This means AI infrastructure investment has surpassed the entire chip manufacturing industry's output scale.

Where Is the Money Going?

GPU/Accelerator Chips (~40%)

  • NVIDIA H100/H200/B100/B200 series
  • AMD MI300/MI350 series
  • In-house chips: Google TPU v5/v6, AWS Trainium/Inferentia, Meta MTIA

Data Centers & Power (~30%)

  • New AI-specific data centers
  • Power infrastructure (transformers, transmission lines)
  • Cooling systems (liquid cooling becoming mainstream)

HBM Memory (~15%)

  • SK Hynix, Samsung, Micron's HBM3E/HBM4
  • Supply shortages driving continuous price increases

Networking & Interconnect (~10%)

  • InfiniBand/RoCE networks
  • Optical modules (800G → 1.6T upgrade cycle)

Other (~5%)

  • Software and toolchains
  • Talent and R&D

Impact on the Supply Chain

Direct Beneficiaries:

  • NVIDIA: Most certain GPU demand; B-series chips in short supply
  • SK Hynix: Leading HBM market share, strong pricing power
  • Arista Networks: Core supplier for AI data center networking
  • Vertiv: Leader in data center power and cooling equipment

Potential Beneficiaries:

  • AMD: MI300 series starting to win large-scale orders, but share still far below NVIDIA
  • Broadcom: Growing demand for custom AI chips (ASICs)
  • TSMC: Advanced process capacity remains tight

Risk Signals:

  • HBM supply bottlenecks may constrain GPU shipments
  • Power infrastructure (especially the US power grid) is becoming a constraint
  • Long data center construction cycles may reduce capital expenditure efficiency

Landscape Assessment

Short Term (6-12 months): AI infrastructure investment won't slow down. Capex guidance in the giants' Q2/Q3 earnings will only get more aggressive. NVIDIA's FY2026 revenue is highly likely to break through $200 billion.

Medium Term (1-2 years): Whether the $1.1T 2027 forecast materializes depends on two key variables:

  1. Whether the commercialization returns of AI applications can keep pace with infrastructure investment
  2. Whether US AI regulatory policies will impose limits on data center construction

Long-term Risks: If AI application revenue growth cannot match Capex growth, 2028-2029 could face the risk of "AI infrastructure oversupply." This bears similarities to the fiber optic infrastructure oversupply during the 2000 internet bubble.

Investment Implications

Most certain direction: GPUs, HBM, advanced packaging — these are the "toll roads" of AI Capex. Regardless of which giant's AI strategy succeeds, they all need to buy this hardware.

Direction requiring caution: Pure data center REITs — if Capex growth slows or AI application returns fall short of expectations, data centers could face oversupply risks.

Signals to watch: Closely monitor Capex guidance changes in each giant's quarterly earnings. If any single company starts scaling back AI investment, it could signal an industry inflection point.