All essays

Essay - 2026-05-14

How to play the AI bottleneck: a field guide to the physical stack behind the model race

Stop trying to pick the winning model and start mapping the physical floor under all of them. The picks-and-shovels framework, the four chokepoints, and the dates on the calendar.

The question gets asked the same way every week. Someone reads about Nvidia, looks at the price chart, decides they are too late, and then asks me, in some form: how do you actually play the AI bottleneck if you do not want to own the index?

The short answer is that you stop trying to pick the winning model and start mapping the physical floor under all of them. The longer answer is what this essay is for.

I run inference at scale for a living. The numbers below come from earnings reports, supply chain disclosures, and what I see inside the bill. None of it is investment advice. All of it is the framework I would have wanted in front of me three years ago.

What is the AI bottleneck, exactly?

The AI bottleneck is the layer of physical and logistical inputs that the entire AI buildout depends on, but which cannot be scaled by writing a check faster than it can be scaled by building a factory.

Compute is not the bottleneck. Compute is the output. The bottleneck is the set of inputs that turn into compute: indium phosphide wafers for the lasers that link the racks, hybrid bonders that stack memory onto logic, gas turbines that feed the grid, transformers that step the power down, switchgear that protects the racks, cooling that pulls 120 kilowatts of heat off a single chassis, and the skilled trades who install all of it.

Every one of those layers has a supplier base that fits on a single sheet of paper. Some of them fit on a Post-it note.

That is the whole game. You are not trying to forecast whether OpenAI or Anthropic wins. You are trying to identify the layers of the stack where the supplier base is two companies wide, the order book is full, and the customer is price-insensitive because the customer's competitor will buy the capacity if they do not.

The framework: bottoms-up, not top-down

Most AI investing in 2024 and 2025 was top-down. Pick the model lab. Pick the hyperscaler. Pick the GPU vendor. That trade worked because the entire stack was levered to the same exponential curve.

The trade that works now is different, because the curve has bent. Compute capex is going up. Token prices have dropped roughly 100x in a year. Margins are thinner. Hyperscalers are starting to discriminate inside their capex budgets.

In that environment, the picks-and-shovels math beats the model-lab math, for a simple reason. A picks-and-shovels supplier gets paid the same dollar amount whether its customer's gross margin is 50% or 20%. A model lab's revenue is exposed to every quarter of price compression in the layer above it.

So the question to ask in front of every name is: what would have to be true for this company to lose pricing power?

For a fab tool vendor like ASML, the answer is "EUV is no longer required." For BESI, "hybrid bonding is replaced by something cheaper." For AXTI, "China removes its export controls permanently and Sumitomo doubles capacity in eighteen months." For GE Vernova, "hyperscalers stop building gigawatt campuses."

None of those answers come true on a one-year horizon. Some of them never come true. That is what a chokepoint looks like.

The four physical constraints that gate the 2026 to 2030 buildout

After two years of pulling on this thread, I keep coming back to the same four layers. They are not the only chokepoints in the stack, but they are the ones with the highest convexity, the cleanest supplier base, and the most legible catalysts.

1. Compound semiconductor substrates

Every laser inside a modern AI rack starts on a polished wafer of indium phosphide. Without the substrate, the laser cannot lase. Without the laser, the transceiver cannot drive the link. Without the link, the rack cannot scale beyond a few hundred GPUs.

The ex-China supplier base is two names: $AXTI in Fremont, California, and Sumitomo Electric in Japan. AXTI's record InP backlog crossed $100 million in early 2026 and the company is planning another capacity doubling in 2027. The China policy catalyst is on the calendar: the suspension of gallium, germanium, and antimony export controls expires November 27, 2026.

The silicon-on-insulator side of the same chokepoint is $SOI (Soitec), which has near-monopoly market share on photonics-SOI wafers for silicon photonics. $IQE on the merchant InP epitaxy side. The substrate layer captures co-packaged optics adoption without depending on which transceiver vendor wins.

2. Advanced packaging

The memory wall is the new compute wall. Every AI accelerator is memory-bound, not FLOP-bound. The bottleneck is not how fast the GPU can multiply, it is how fast HBM can feed it. And HBM is gated by advanced packaging: TSV etch, hybrid bonding, sub-10-micron bump inspection.

There are roughly four companies on Earth that matter in this layer. $TSM owns CoWoS, sold out through 2026. $BESI is the cleanest hybrid-bonding convexity, with hybrid-bonder unit orders more than doubling sequentially in Q1 2026. $ONTO and $CAMT split 3D metrology between them. $LRCX owns the etch step.

The convexity here is that hybrid bonding moves from pilots to volume in 2026 and 2027. BESI is the public name with the steepest revenue ramp into that transition. The bear case (TCB stays competitive longer than expected) is real but already in the multiple.

3. Power and grid equipment

The chokepoint is not electricity. The chokepoint is the equipment that delivers electricity at industrial scale: gas turbines, transformers, switchgear, busways, and the trades labor to install all of it.

$GEV is the most legible name in the basket. The Q1 2026 backlog crossed 100 GW. Pricing is up 10-20% sequentially. $2.4B in data-center electrification orders was booked in a single quarter, more than the full prior year. Gas turbines from the three majors (GE Vernova, Mitsubishi Power, Siemens Energy) are sold out into 2030.

The under-discussed layer is medium-voltage switchgear and busway, where $ETN, $SU.PA (Schneider), and $POWL have structural pricing power that nobody outside of utility procurement teams understands. Storage adjacency goes through $FLNC.

The MEP construction layer captures the same demand from the labor side. $PWR has a $48.5B record backlog. $EME has $15.6B in record RPO. $FIX has seen its backlog double to $12.45B. These are demand-uncapped, supply-capped businesses.

4. Critical minerals and rare earths

The April 2025 China export controls on heavy rare earths turned the magnet supply chain into a strategic asset. Every AI motor, every actuator, every defense application that uses neodymium-iron-boron magnets depends on a supplier base that is roughly 90% Chinese.

$MP Materials is the only US rare-earth producer at scale. The market still treats it as a mining cyclical. The thesis treats it as a strategic asset whose end markets (AI plus defense) are price-insensitive at the margin.

The November 27, 2026 calendar date matters here too, because the gallium, germanium, and antimony controls cover the same policy surface. If the controls are re-imposed at expiration, the substrate names and the rare-earth names re-rate at the same time.

How to size the layers

The temptation, when someone draws a picture like this, is to put 10% of a book into every layer. That is the wrong sizing.

The right sizing comes from asking three questions about each layer:

First, how thin is the supplier base? Two names is a chokepoint. Five names is a competitive market. The book should weight toward the two-name layers.

Second, what is the catalyst path? A chokepoint with no calendar catalyst pays you optionality over years. A chokepoint with a date on the wall pays you a re-rating event. The book should overweight the dated chokepoints when the date is inside twelve months.

Third, what is the substitution risk? Silicon photonics could displace discrete EMLs, compressing the InP-vertical moat. TCB could remain competitive against hybrid bonding for longer than the bull case assumes. The book should fund substitution risk with optionality on the substituting technology, not by walking away from the chokepoint.

The version of this framework I run myself looks like roughly 12% positions in the highest-conviction chokepoints (substrate duopoly, advanced packaging convexity, gas turbine cartel, retimer monopoly), 6% positions in the second tier (hybrid bonding metrology, liquid cooling pure plays, custom silicon duopoly), and 3% to 4% positions in the asymmetric tail names (single-catalyst stories like $AAOI, $AEHR, $KLIC).

That is not a recommendation. It is a shape. The shape says: concentrate where the supplier base is thinnest, fund the rest from the layers where the supplier base is competitive.

The catalysts to mark on the calendar

The patience that a chokepoint thesis requires is not infinite patience. It is patience that ends on specific dates.

The dates I have on the calendar for the next eighteen months:

A chokepoint thesis is mostly the work of doing nothing in the gaps between those dates. The hard part is not finding the names. The hard part is holding them through the quarters where the chart goes sideways.

What can go wrong

The skeptical version of the chokepoint thesis is worth taking seriously, because some of it will be right.

The first risk is that hyperscaler capex normalizes. If the 2027 capex guides come in flat instead of up, the entire stack compresses at once. The defense is to overweight the layers where the customer base is broader than the four hyperscalers (industrial power, advanced packaging that serves auto and consumer too, substrate layers that feed telecom as well as AI).

The second risk is that substitution arrives faster than the bull case assumes. Silicon photonics could ramp ahead of EMLs. TCB could compete with hybrid bonding for a generation longer. The defense is to hold the substituting technology as optionality inside the same theme.

The third risk is that China relaxes export controls permanently. That would compress the InP and rare-earth premium meaningfully. The defense is sizing: do not let any single policy outcome become more than 15 to 20% of the book.

The fourth risk is the one most people underweight: the AI financing market itself. Neoclouds, leveraged GPU buyers, and the SPV structures that fund hyperscale debt have not been tested through a real drawdown. A financing wobble would hit the demand side of the chokepoint stack before it hit the supply side.

The posture, not the trade

The hardest thing to teach in this kind of investing is the posture. The chokepoint thesis is not a trade. It is a multi-year posture you hold against a market that wants to narrate AI as a model race because the model race is what makes good copy.

The narrative wants to be about Sam Altman and Dario Amodei. The mechanics are about indium phosphide and hybrid bonders. The narrative is sold every day. The mechanics are sold a few weeks per year, when the supplier prints earnings and the chokepoint becomes visible for a single news cycle.

Almost everyone can name five hyperscalers. Almost nobody can name five substrate suppliers. That asymmetry of attention is the entire opportunity. The market is now learning, maybe too fast, maybe not fast enough, that the floor underneath the model race is thinner than it looks.

How do you play the AI bottleneck? You stop watching the model race and you start counting the suppliers. You hold two thoughts at once: the bottleneck can be real, and the stock can still become crowded. The substrate can matter, and valuation can still matter. The supply chain can be structurally tight, and the trade can still reverse violently.

You walk the chain.

The free public dashboard at [aibottlenecks.app](https://aibottlenecks.app) maps every layer with live prices, the calendar catalysts, and the written thesis behind each name. No login. Not a portfolio. Not a recommendation. A prism.

Educational, not investment advice.