The AI revolution is constrained not by software but by physical infrastructure. Hyperscalers have committed $650B+ in 2026 capex for AI data centers, yet nearly half of planned U.S. projects are delayed or canceled due to supply chain bottlenecks. The shortages are structural, not cyclical—CoWoS packaging, HBM memory, power transformers, and critical minerals face multi-year constraints. This creates a seller's market where incumbent suppliers have unprecedented pricing power.
The investment framework: capital flows from abundant top-layer services down to scarce bottom-layer resources. The highest-conviction opportunities sit at the foundation layers where supply constraints cascade throughout the entire chain.
| Layer | Category | Priority | Key Tickers |
|---|---|---|---|
| LAYER 5 | AI Applications & Services | LOWEST | MSFT, AMZN, GOOG, PLTR |
| LAYER 4 | Networking & Interconnects | MEDIUM | COHR, AVGO, LITE, GLW |
| LAYER 3 | Power & Cooling Infrastructure | HIGH | CEG, VRT, POWL, SMR |
| LAYER 2 | Semiconductors & Memory | HIGHEST | 000660.KS, MU, ASX, TSMC |
| LAYER 1 | Critical Minerals & Materials | HIGHEST | FCX, MP, RIO, BHP |
Investment edge: own what AI giants cannot operate without.
| Supplier | Market Share | HBM4 Status | Pricing Power |
|---|---|---|---|
| SK Hynix | 62% | Volume production early 2026; exceeds JEDEC spec by 25% | Very High |
| Micron | 21% | Samples at 11 Gb/s (highest performance); mass production 2026 | Very High |
| Samsung | 17% | Mass production H1 2026; AMD MI450 supplier | High |
- HBM market: $38B (2025) → $58B (2026) → ~$95B (2028)
- SK Hynix: "Already sold out our entire 2026 HBM supply"
- Micron: "HBM capacity for 2025 and 2026 is fully booked"
- Samsung: Raising HBM prices by high-teens to low-twenties % for 2026 contracts
Investment Angle: SK Hynix offers the purest HBM play. Micron offers the most dramatic earnings leverage. Samsung offers a contrarian recovery.
Without CoWoS packaging, even perfectly fabricated 3nm wafers cannot become functional AI chips.
| Company | Role | Investment Merit |
|---|---|---|
| ASE Group (ASX) | CoWoS outsourcing partner | Advanced packaging sales doubling in 2026 |
| Amkor (AMKR) | Alternative packaging provider | Building new capacity aggressively |
| Intel (INTC) | EMIB/Foveros alternative | Domestic U.S. packaging advantage |
- TSMC: "CoWoS capacity is very tight and remains sold out through 2025 and into 2026"
- NVIDIA: "CoWoS assembly capacity is oversubscribed through at least mid-2026"
- TSMC scaling CoWoS from ~35k wafers/month to 130k wafers/month by end of 2026—nearly 4x
U.S. data center electricity: 176 TWh (2023) → 325-580 TWh by 2028. AI will drive a 165% increase in data center power demand by 2030 (Goldman Sachs). Transformer delivery now takes up to 5 years (was 24-30 months pre-2020).
The Nuclear Pivot
Big tech signed contracts for 10+ GW of new nuclear capacity in the past year:
| Company | Deal | Capacity | Timeline |
|---|---|---|---|
| Microsoft | Three Mile Island restart | 835MW | 2027 |
| Kairos Power SMRs | 500MW | 2030 | |
| Amazon | Susquehanna campus + X-energy | 5GW SMR projects | 2028-2030s |
| Meta | RFP for nuclear generation | 1-4GW | TBD |
Nuclear Investment Targets
| Company | Ticker | Key Positioning |
|---|---|---|
| Constellation Energy | CEG | Largest U.S. nuclear fleet; $1B DOE loan |
| Vistra Energy | VST | Comanche Peak PPA; 2,400MW through 2053 |
| Talen Energy | TLN | Direct AWS partnership; 1,920MW contracted |
| NuScale Power | SMR | First U.S. SMR NRC approval (May 2025) |
| Metric | Value |
|---|---|
| 2025 Market Size | $4.58B |
| 2035 Market Size | $44.4B |
| CAGR | 25.5% |
| Two-phase immersion share (2026) | 66% |
| Liquid cooling PUE | 1.05-1.15 vs. 1.4-1.8 for air |
Key players: Vertiv (VRT), Schneider Electric (SU), CoolIT Systems, LiquidStack (TT), Daikin (6367.T)
Copper: The Workhorse Metal
- A single hyperscale facility uses ~2,200 tonnes of copper (~27 tonnes per MW)
- Data center copper demand: 375,000 tonnes/year by 2030
- New mines take 17 years from discovery to production
- BHP: Copper demand will rise 72% by 2050
China Control Risk
| Mineral | China Control | AI Use |
|---|---|---|
| Gallium | 98% of primary production | GaAs/GaN semiconductors |
| Germanium | 60% of refining | Fiber optics, high-performance chips |
| Rare Earths | 70% production, 90% processing | Magnets, motors |
| Copper | ~15% of mining | Cabling, power, cooling |
| Bottleneck | Severity | Duration | Investment Play |
|---|---|---|---|
| CoWoS Packaging | 9.5/10 | 36 months | ASE Group (ASX), Amkor (AMKR) |
| HBM Memory | 9.0/10 | 24 months | SK Hynix, Micron (MU) |
| 3nm/2nm Wafers | 8.5/10 | 24 months | TSMC (TSM) |
| Power Transformers | 8.0/10 | 18 months | Powell Industries (POWL) |
| EUV Lithography | 8.0/10 | 24 months | ASML (ASML) |
| Liquid Cooling | 6.5/10 | 12 months | Vertiv (VRT) |
| Optical Transceivers | 6.0/10 | 12 months | Coherent (COHR), Lumentum (LITE) |
Tier 1 — Critical Bottlenecks (Highest Conviction)
| Company | Ticker | Thesis | Key Catalyst |
|---|---|---|---|
| SK Hynix | 000660.KS | 62% HBM share; sold out through 2026 | HBM4 ramp, price hikes |
| ASE Group | ASX | CoWoS outsourcing; packaging sales doubling | TSMC capacity constraints |
| Vertiv | VRT | Liquid cooling leader; 25%+ CAGR market | Mandatory liquid cooling transition |
Tier 2 — Infrastructure Enablers (Medium-High Conviction)
| Company | Ticker | Thesis |
|---|---|---|
| Constellation Energy | CEG | Largest U.S. nuclear fleet; multiple hyperscaler PPAs |
| Coherent Corp. | COHR | 300M transceivers; DCI segment growing 34% CAGR |
| Freeport-McMoRan | FCX | Largest U.S. copper miner; structural demand surge |
Tier 3 — Emerging Opportunities (Higher Risk/Reward)
| Company | Ticker | Thesis | Risk Level |
|---|---|---|---|
| NuScale Power | SMR | First U.S. NRC-approved SMR | High — pre-revenue |
| MP Materials | MP | Only U.S. rare earth processing | High — China competition |
| Powell Industries | POWL | Small-cap switchgear; 30% supply gap | Medium — execution |
The defining characteristic of the AI supply chain is that shortages are architectural, not cyclical. Past semiconductor cycles resolved in 12-18 months. The AI cycle is different because demand grows exponentially while supply is constrained by physical limits.
The playbook: Own the bottlenecks. The companies controlling scarce resources in CoWoS, HBM, nuclear power, liquid cooling, and critical minerals will determine AI's pace, shape its economics, and capture disproportionate value.
Standard Kepler Research | standardkepler.com