A 3-Layer Portfolio Framework for the 9 AI Infrastructure Stocks — 40/40/20 Weighting

Knowing the 9 AI choke-point stocks doesn't mean you should buy all 9 equally weighted.

The way I look at this list is as a 3-layer portfolio — Foundation 40%, Silicon 40%, Materials 20%. Each layer responds to different market shocks on different timescales, so equal weighting doesn't make sense.

This piece walks through how to weight, sequence, and trigger your entries and exits across these 9 names — from one individual investor's perspective. It's a starting framework, not a prescription.

1. Foundation Layer — 40% (Vistra · Eaton · Vertiv)

One-line: What AI needs first to even function — electricity and cooling.

These three are lower volatility, compound steadily, and don't depend on any one chip company winning. That's why they get the largest allocation in my framework.

Vistra (3,800MW of contracted AI nuclear). Nuclear plants in Texas and the Midwest contracted for AI data centers — double the nearest competitor. Meta locked in 2GW for 20 years, AWS 1.2GW for 20 years. Fixed cost structure means rising natural gas prices expand its relative margin.

Eaton (11-year transformer backlog). Transformer lead times are 128–144 weeks. Half of all planned 2026 US data center builds are delayed or canceled because of this. Eaton owns the entire data center electrical chain — transformers, switchgear, distribution, busway. As long as data center capex stays alive, revenue keeps coming in.

Vertiv (effective monopoly on liquid cooling). New Nvidia AI racks pull 132kW. Ten times a traditional server rack. Air cooling can't handle it. Liquid cooling is the only answer, and Vertiv is the company Nvidia named as their official AI factory infrastructure partner globally. $15B backlog — more than a full year of revenue already locked in.

Why 40%: These three earn revenue regardless of which LLM wins or which chip becomes the standard. Lowest volatility, lowest cycle risk in the basket. Heaviest weight to steady compounding makes sense.

2. Silicon Layer — 40% (Micron · Amkor · Broadcom · Marvell)

One-line: Largest upside in the basket — also largest volatility.

I covered these four in detail in my AI silicon stack 4-bottleneck piece. Just the highlights here.

Micron — 21% of HBM supply, asymmetric Hormuz beneficiary. Revenue +56%, EPS +175%, PEG 0.25. Fastest growth, but also most cyclical.

Amkor — #1 overflow source for TSMC CoWoS. Advanced packaging revenue tripling this year. Small base means highest volatility.

Broadcom — $73B AI backlog. 60–70% of custom AI chips, 80% of Ethernet switching. Pre-locked 18–24 months of revenue is the safety margin.

Marvell — #1 in optical transceiver chips. 63M transceiver shipping base. 800G → 1.6T → 3.2T transitions don't reverse once they happen.

How I'd split inside that 40% — my framework:

Stock	Share within 40%	Reason
Broadcom	35%	Backlog safety, largest AI revenue mix
Marvell	30%	Optical cycle just beginning
Micron	20%	Highest beta, also highest cyclicality
Amkor	15%	Highest volatility, smallest weight

Why 40%: This is where I see the largest upside over the next five years. But cyclicality is too high to match the Foundation weight. Same 40% allocation, but tiered by safety inside.

3. Materials Layer — 20% (Southern Copper · Corning)

One-line: Longest time horizon, most value-investing flavor.

These two have weak seasonal momentum and aren't sensitive to a single quarter of data center capex shifts. Instead, they ride 5–10 year supply deficit trends that compound steadily.

Southern Copper — 51.1M ton reserves, $0.42/lb production cost. S&P Global projects a 330k ton copper deficit in 2026, widening to 10M tons by 2040. Each 100MW data center uses 27–33 tons of copper. Copper is the substrate every other 8 names sit on.

Corning — $6B Meta fiber deal. AI data centers use 36x more fiber than traditional racks. Ribbon fiber lead times exceed 60 weeks. New fiber preform capacity takes 18–24 months to come online.

Why 20%: Cycles are long, so multi-baggers in the next 1–2 years are unlikely. But the supply deficit is genuinely long — it doesn't resolve in 2027 either. Small long-term core position makes sense.

4. You Don't Need to Buy All 9 — Phased Entry Sequence

If buying all 9 at once feels like too much, here's the entry sequence I'd recommend.

Phase 1 (if you must start with two): Vistra + Broadcom. Highest-safety foundation plus deepest backlog silicon. Just these two captures 70% of AI infrastructure exposure.

Phase 2 (add when you're showing returns): Eaton + Vertiv + Marvell. Transformers, cooling, optics — fill out foundation and bridge into the optical cycle.

Phase 3 (only after the portfolio is stable): Micron + Amkor + Southern Copper + Corning. Higher-beta names and long-duration value plays go in last.

The sequence runs stability → growth → asymmetry. Betting on asymmetry from day one means you don't have the constitution to ride out the volatility.

5. Triggers — When to Add, When to Trim

The 5 triggers I watch most closely:

1. Qatar helium export volume changes → Trigger to ADD Micron + Vistra weight 2. TSMC announcing CoWoS capacity expansion → Trigger to TRIM Amkor (overflow demand declines) 3. Adoption rate of new optical transceiver standards (1.6T) → Trigger to ADD Marvell 4. US data center grid-connection wait times shortening → Trigger to TRIM Eaton + Vistra 5. Copper ETF inflows accelerating → Trigger to rebalance Southern Copper weight

Monitor these five quarterly and you can dynamically adjust relative weights across the 9 names.

Closing — A Framework, Not a Prescription

40 / 40 / 20 is what I see as a reasonable starting point — not the right answer. Your risk tolerance, time horizon, and existing portfolio positions all change the right weights.

Whatever weights you land on, one thing is clear — AI infrastructure exposure isn't tied to one Nvidia ticker. Keep even 3–4 of these 9 in your head and you'll move a beat ahead of the market the next time Hormuz hits the news.

This is for educational purposes, not investment advice.