AI’s Hidden Leverage Points
--
The scarce layers beneath the AI trade.
Everyone is watching the model layer.
But the AI trade is not really about intelligence. It is about constraint.
Every time one layer gets solved, the bottleneck moves somewhere else. Chips became packaging. Packaging became memory. Memory became power. Power becomes optical wiring, cooling, permits, and regulation.
The opportunity is not where AI is visible. It is where AI becomes physically hard to scale.
Hundreds of billions of dollars are going into AI infrastructure. Revenue is still a fraction of that. The gap between spending and return is not new. It is the shape of every infrastructure bubble. What is new is how fast the constraint rotates.
Where pricing power migrates as each bottleneck gets solved.
The market misprices AI because it misreads expertise
AI is a cross-field bridge. It connects knowledge no single human career could span. But it is built on an architecture that rewards sounding right over being right.
Most people treat fluency as intelligence. The market has no way to price that difference.
A weak user asks AI to replace judgment. A strong user uses it to widen the search space, cross-check assumptions, and connect fields faster than a single career could. The output may look similar on the surface. The economic value is not.
This matters because it shapes where capital flows. Money moves toward what looks like AI :- chatbots, model companies, application layers and away from what makes AI physically possible. The model is visible. The constraint is not.
The bottleneck keeps moving down the stack
The bottleneck does not disappear. It migrates.
In 2023, the constraint was chips. Nvidia’s GPUs were allocation-only. In 2024, it was advanced packaging. TSMC couldn’t assemble chips fast enough. In 2025, it moved to high-bandwidth memory. Now it is power. Next it will be optical interconnects and cooling.
Each shift reshapes which industries have pricing power. The companies that matter in one phase are not always the ones that matter in the next.
This rotation is why the AI trade is not one trade. It is a sequence of trades, each defined by whatever is hardest to scale at that moment.
Power
AI demand does not stop at GPUs. It pushes into turbines, substations, transformers, grid equipment, and interconnection capacity. The grid was not built for this.Companies exposed to this layer include GE Vernova, Quanta Services, Constellation Energy, and Eaton.Power is slower to scale than software. Permitting alone takes years. That gives the layer pricing power even after the market starts looking beyond chips.
Optical & Cooling
As clusters scale past current limits, the constraint moves to moving data between chips and removing heat from density. AI facilities require far more fiber than traditional data centers, while liquid cooling demand is rising quickly as density increases.Companies exposed to this layer include Coherent, Lumentum, Corning, Vertiv, Modine, and Schneider Electric.These layers are capacity-constrained with few qualified suppliers. That is the definition of pricing power.
Memory
High-bandwidth memory is the current bottleneck between compute and performance. Three companies make it. SK Hynix has held a leading share and qualified early with Nvidia, which in semiconductors can mean designed-in as default for that generation.Companies exposed to this layer: SK Hynix, Micron, Samsung.
Packaging
TSMC manufactures chips and does advanced packaging in-house. Very few companies can combine leading-edge manufacturing and advanced packaging at this density. ASE handles overflow when TSMC is capacity-constrained, which current demand has made increasingly common.Companies exposed to this layer: TSMC, ASE Technology, Amkor.
Compute
Nvidia designs GPUs and CUDA, the programming layer most AI labs have spent a decade optimizing against. Switching chips means rewriting years of custom code for worse performance, so almost nobody does.The hardware matters less than the ecosystem lock-in.
That is why a correction in AI does not mean the AI buildout stops. It means the market starts repricing which layers were essential and which were only riding the narrative.
The correction will not hit every layer equally
A 50 to 70 percent drawdown in exposed AI equities would not be abnormal if this follows prior infrastructure bubbles. Telecom in 2000. Railways in 1845. Canals before that. The pattern is consistent: massive overinvestment, followed by a collapse that separates the structurally necessary from the narratively inflated.
The correction does not destroy the whole stack. It separates scarcity from narrative.
What survives sits on one of three things: physical scarcity (you cannot download a turbine or permit a nuclear plant in under a decade), switching costs (when your customers have spent years building against your platform, leaving costs more than staying), or regulatory walls (permits, grid interconnection queues, nuclear licenses, time advantages that compound).
Everything else is margin waiting to be absorbed.
What survives
The winners are not always the companies making AI look intelligent. They are the companies making AI possible.
The spread between structurally scarce and narratively inflated is where the opportunity sits. Not in picking which chatbot wins. Not in guessing which model company raises the next round. It sits in mapping which physical layers the entire stack depends on, and which of those layers cannot scale fast enough to meet demand.
The bottleneck keeps moving. The leverage stays with whoever sits on the constraint.
Download the full memo with source notes, company map, and detailed metrics
Download Ai’s Hidden Leverage Points.pdf
The public thesis is the map. The PDF is the evidence layer.
Source note: Goldman Sachs estimates more than $500B of AI infrastructure investment in 2026; other estimates put total AI capex closer to $600–700B. The full source list is in the PDF.
