AI Infrastructure
The physical buildout layer of AI — accelerators, advanced packaging, high-bandwidth memory, datacenter networking and GPU cloud — where hyperscaler capex converts directly into hardware demand.
The story
AI workloads have triggered one of the largest infrastructure buildouts in modern history, with the four largest US hyperscalers alone planning a step-change in 2026 capital spending and roughly three-quarters of that directed at AI compute. Demand for accelerated servers, advanced-node silicon and high-bandwidth memory is running ahead of supply across the chain, leaving foundry packaging, memory and networking capacity structurally tight into 2027. The result is a multi-year demand pull rather than a single product cycle, anchored by long-dated capex commitments and multi-billion-dollar backlogs.
The outlook to 2030
Through 2030, McKinsey models cumulative compute-value-chain investment of roughly $5.2 trillion in a base case (of which ~$3.3 trillion is IT equipment), driven by an estimated 125 incremental gigawatts of AI datacenter capacity. The economics shift from training-led to inference-led demand, broadening the buyer base and lifting the share of custom accelerators and Ethernet-based AI fabrics. Bottlenecks migrate over time from raw GPU supply toward power, advanced packaging, memory and the network — making the picks-and-shovels layers increasingly value-critical.
What is moving the capital
The forces routing money into this theme right now.
Hyperscaler 2026 capex step-up
The four largest US hyperscalers guided to a sharp 2026 capex increase versus a record ~$388B in 2025, with roughly 75% directed at AI infrastructure.
HBM and DRAM supply shortage
SK Hynix reported its 2026 memory capacity essentially sold out, with HBM4 mass production pulled forward to early 2026 and structural tightness lasting into 2027.
Advanced packaging (CoWoS) capacity ramp
TSMC is doubling advanced-packaging capacity yet still describes it as 'very tight,' keeping packaging a gating constraint on accelerator output.
Custom AI accelerators (XPU/ASIC) scaling
Hyperscaler custom-silicon programs are scaling, with Broadcom citing a $10B TPU rack order plus an $11B follow-on for delivery into late 2026.
리더 종목
Structural large-cap anchors — lower-variance exposure to the theme.
NVIDIA is the gravity well of the AI infrastructure stack, positioned to capture the lion's share of hyperscaler capex as AI compute demand explodes; its software stack enhances GPU versatility, driving revenue beyond depreciable life.
Why the excitement: Management noted data center revenue of $75 billion, up 92% year over year, driven by strong demand for GB300 and NVL72 systems.
The honest risk: Consumer demand fell modestly due to higher memory and system prices, indicating potential sensitivity to component cost fluctuations.
TSMC is the linchpin of AI infrastructure, fabricating nearly every leading-edge AI chip. Its near-monopoly position and advanced process technologies like 3nm and 2nm make it a structural play on AI's hardware backbone.
Why the excitement: TSMC's Q1 FY2026 earnings highlighted the strength of leading-edge process technologies, with 3-nanometer contributing 25% of wafer revenue, signaling strong demand for AI-related chips.
The honest risk: Gross margin dilution from the ramp-up of 2-nanometer technology and overseas fabs could pressure profitability in the coming years.
Broadcom is a structural play on AI infrastructure, providing custom ASICs and networking solutions that underpin hyperscalers' AI buildouts; its deep partnerships and critical silicon position it as a lower-variance way to play the AI capex cycle.
Why the excitement: Broadcom anticipates AI chip revenue exceeding $100 billion in 2027, driven by custom AI accelerator deployments across its customer base.
The honest risk: Reliance on a concentrated customer base of hyperscalers exposes Broadcom to potential demand shifts or insourcing initiatives.
비대칭 플레이
Smaller names with higher upside and deeper potential drawdowns.
CoreWeave is a neo-cloud provider specializing in GPU-as-a-service, positioned to capture significant AI infrastructure spending, underpinned by substantial hyperscaler commitments and a rapidly expanding contracted power footprint. However, its high debt-to-equity ratio warrants caution.
Why the excitement: CoreWeave signed more than $40 billion of new commitments in Q1 FY2026, growing contracted revenue backlog to nearly $100 billion.
The honest risk: A current ratio of 0.3 suggests potential short-term liquidity challenges given the capital-intensive nature of expanding AI infrastructure.
Asymmetry: High upside if GPU demand and execution hold; deep drawdown if AI capex moderates or financing tightens.
Astera Labs is a smaller, faster-growing AI-infrastructure name focused on connectivity solutions linking AI accelerators within data centers. The company's retimers and fabric are critical for enabling high-performance AI systems.
Why the excitement: Astera Labs anticipates strong revenue growth through 2026 and into 2027, driven by AI fabrics and the transition to PCIe 6, 800 gig, and 1.6T Ethernet connectivity.
The honest risk: As a smaller company, Astera Labs is more vulnerable to customer concentration and potential shifts in hyperscaler spending priorities.
Asymmetry: Strong upside tied to accelerator volume; sharp drawdown if a single large customer pulls forward or slows.
Credo Technology Group is a smaller, high-growth interconnect player leveraged to the buildout of AI infrastructure, specifically through its active electrical cables (AECs) and SerDes solutions. While carrying higher risk, Credo offers asymmetric upside from rising data center rack density and accelerating bandwidth demands.
Why the excitement: Credo management expects fiscal year 2026 revenue to triple after more than doubling in fiscal year 2025, showcasing exceptional growth in the semiconductor space.
The honest risk: Credo's smaller size and higher valuation (price/sales of 37.7) make it more vulnerable to customer concentration and potential shifts in hyperscaler spending.
Asymmetry: Large upside on AI networking ramp; high drawdown risk given customer concentration and valuation.
Inside the theme
The sub-layers and the leaders that anchor each one.
자주 묻는 질문
Are AI infrastructure stocks still worth it in 2026?
The thesis rests on hyperscaler capex above $400B for 2026 and power-purchase visibility into the 2030s. That makes AI infrastructure a structural multi-year build rather than a single-year spike, though valuations and capex pace are the swing factors.
NVDA or the smaller suppliers?
NVIDIA is the structural leader and reference design, but smaller suppliers like CoreWeave, Astera Labs and Credo carry higher asymmetry — more upside if the build-out continues, deeper drawdowns if AI capex slows.
How long does the AI infrastructure cycle last?
Hyperscaler power-purchase agreements already reach into the 2030s, and the projected build runs at a high-20s to low-30s percent CAGR — this is framed as a structural, not single-year, theme.
How does the MoonshotScore evaluate this theme?
For AI themes the score weights revenue growth, insider buying, R&D intensity and market-share gains more heavily — the factors that separate structural winners from momentum names.
What is the main risk?
The key risks are a moderation in hyperscaler capex from 2027, China chip restrictions, and a GPU supply glut — any of which could de-rate the whole chain quickly.
Research sources
- IDC — Artificial Intelligence Infrastructure Spending to Reach $758Bn USD Mark by 2029
- IDC — AI Infrastructure Spending Caps Historic Year at ~$90 Billion in Q4 2025; 2029 Spending to Eclipse $1 Trillion
- McKinsey & Company — The cost of compute: A $7 trillion race to scale data centers
- Statista — Big Tech's AI Spending to Reach $725 Billion in 2026 (capex of Meta, Alphabet, Amazon, Microsoft)
- NVIDIA / SEC — NVIDIA Q4 and FY2025 Results (Form 8-K)
- Asia Business Outlook — TSMC 2025 Revenue Jumps 32% on Strong AI Chip Demand
- Futurum Group — Broadcom Q4 FY2025 Earnings: AI And Software Drive Beat
- Tom's Hardware — Samsung and SK hynix warn AI-driven memory shortages could last until 2027 and beyond
Related themes
Energy — The Electricity of AICopper — The AI BottleneckAI Agents & Agentic Software
투자 자문이 아닌 교육 콘텐츠. 과거 성과가 미래 결과를 보장하지 않습니다.