The Layers of the AI Trade
How AI infrastructure layers turn chip demand into second-order opportunities across power, cooling, networks and capital for global investors.
Home / Map
The Layers of the AI Trade
AI deployment is not one trade. Chips trigger spending across power, cooling, networks, data centres and capital, so the investable opportunity sits with the suppliers that remove bottlenecks around accelerators. This map explains those layers, the companies exposed to each, and how international investors can access them across global markets.
Key takeaway: Everyone is watching Nvidia. This letter watches what Nvidia's spending does to everyone else.
The first-order AI trade is priced. You know the names, you know the multiples, and so does everyone else. That is not a criticism of the names. It is a description of where the easy money went.
Second Order exists for what happens next. The 14 largest publicly owned data-centre operators are on track to spend close to $750 billion of capital this year, up from a little under $450 billion in 2025, according to BloombergNEF, with more than 23 gigawatts of capacity under construction at the last count. That money does not stop at the chip. It flows outward through a stack of industries, and at each layer it lands on companies that most investors have never filed under "AI".
This is the map of that stack. Nine layers, one sentence each on why it matters, and the second-order question that decides who profits.
- 1. Semiconductors: design, foundry, equipment, memory
The chip layer is where the trade starts, and it is more than one trade. Design (Nvidia, AMD, Broadcom), foundry (TSMC), equipment (ASML, Applied Materials, Lam, KLA, Advantest, Disco) and memory (SK Hynix, Samsung, Micron) each have their own cycle. The second-order question: which sub-layer is capacity-constrained rather than demand-constrained? Right now it is high-bandwidth memory, which is why the memory makers are behaving like AI stocks.
- 2. Power and energy
Electricity, not silicon, is now the binding constraint on the buildout. Grid interconnection queues in major US markets run three to seven years, and in some markets seven to ten. Texas paused new data-centre development in August pending an audit of an interconnection queue holding roughly 474 gigawatts of requests. The second-order question: who can deliver power in 2027, when the grid says 2029? Independent power producers with contracted nuclear (Vistra, Constellation), the grid-equipment makers (GE Vernova, Eaton) and the engineering firms that build the sites (Quanta) are the answer the market is slowly arriving at.
- 3. Cooling and thermal
AI racks draw 100 to 300 kilowatts against 30 to 40 for a traditional rack, and air cannot carry that heat away. Fewer than one in ten data centres is liquid-cooled today, and every new GPU generation pushes that share up. The second-order question: when does liquid cooling go from optional to mandatory? That is the inflection for Vertiv, nVent, Modine, Schneider Electric and the Taiwanese cold-plate supply chain.
- 4. Data-centre operators and REITs
The Asian data-centre REITs and the broader regional buildout are the part of this layer international investors can track across markets.
- 5. Neoclouds
The specialist GPU-rental operators, CoreWeave, Nebius, Applied Digital, and the crypto miners pivoting to high-performance compute. The most speculative and the most capital-hungry layer, with a queue of names lining up to list. The second-order question: is the GPU rental price falling faster than the cost of capital? Whoever answers that correctly will make or lose the most money in the entire stack.
- 6. Networking and optical
Bigger clusters mean the data has to move between chips, and that is its own industry: Arista for switching, Coherent and Lumentum for transceivers, Corning for fibre, Amphenol and TE Connectivity for the connectors. The second-order question: does bandwidth scale with compute, or faster than it? So far, faster.
- 7. Hyperscalers and cloud platforms
Microsoft, Amazon, Alphabet, Meta and Oracle are the ones writing the cheques, and they are the only public-market route into the private model labs. The second-order question is the one the whole trade turns on: does the capex earn a return, and when does the market start asking?
- 8. Servers and hardware assembly
Super Micro, Dell and HPE assemble the boxes. Thin margins, huge volumes, and a layer that lives or dies on the cadence of the chip layer above it.
- 9. The application and software layer
Palantir, ServiceNow, Salesforce and the enterprise-software field bolting AI onto products. This is where the argument about whether AI actually makes money gets settled, and it is the layer that will tell us, eventually, whether layers one to eight were a bubble.
Which link binds first
The method behind this letter is simple. A supply chain runs at the speed of its tightest link, and the tightest link moves. In 2024 it was accelerators. In 2025 it was high-bandwidth memory. Through 2026 it has been power. Every week I will publish a tracker that rates each of the nine layers Loosening, Stable or Tightening, and flags the one whose status changed most. The tightening layer is where the pricing power is, and pricing power is what a second-order investor is looking for.
For international investors
Most of the writing on this trade is American, for Americans. Second Order is written for international readers, which means two things. First, every piece names how you actually get exposure: the US listing, the ADR, the Hong Kong or Taiwan line, the ETF. Second, the Asian half of the supply chain, the foundries, the memory makers, the equipment and cold-plate suppliers, the regional data-centre REITs, gets the attention it deserves rather than a footnote.
What to expect
Three posts a day on the site, across three sections. The Wire: what moved in the stack today, and the second-order read. The Map: one layer, one company or one term explained, building this reference library out one piece at a time. Research: single-name and single-layer analysis, with the argument, the exposed tickers, how an international investor gets access, and an honest note on what would change my mind. Research is the paid tier and it is free during the launch season.
One email a week, on Thursdays, carrying the flagship Deep Dive, the bottleneck tracker, and the three pieces worth your time. Nothing else lands in your inbox.
Second Order is made by TKN, the team behind The Kopi Notes.
Not investment advice. Do your own research.
Written by Second Order by TKN for international investors. Educational research only — not investment advice. Do your own research.
Home / Map / The Layers of the AI Trade
Read the AI Stock Map: The Stack · Related layer: Power & energy
FAQ
What are the layers of the AI trade? Chips, power, cooling, networks and capital each create second-order beneficiaries.
Why focus on second-order effects? The best opportunities often sit with suppliers and infrastructure that make AI deployment possible.
