The world is built out of a few narrow places.

I did not start with the idea of building a website. I started, like a lot of people, by trying to understand what was actually happening with AI. Not the demos. Not the launch threads. Not the model leaderboards that get reshuffled every six weeks. The thing underneath all of that. The physical thing. The seven hundred billion dollars a year of hyperscaler capex that has to convert, somehow, into wafers and fiber and concrete and electricity in the real world.

The further I pulled on that thread, the stranger the picture became. Almost everyone is talking about the same five companies. Almost no one is talking about the layers below them. And the layers below them are where the actual constraints live. Not in the clouds, not in the models, not in the keynote slides. In a few buildings, in a few towns, run by a few thousand engineers most of the world has never heard of.

The night I fell into the EUV wormhole

I can tell you exactly when this stopped being a casual interest and became something closer to an obsession. It was late, I was reading about why China cannot make leading-edge chips even after spending hundreds of billions of dollars trying, and the answer kept coming back to one Dutch company in a town called Veldhoven that almost nobody outside the industry has heard of. ASML. I assumed it was the usual story of a company with a clever patent or a good lobbyist. It was not. It was something much weirder.

What ASML actually sells, when you strip away the acronyms, is a machine the size of a city bus that costs around four hundred million dollars and is, by a fair distance, the most complicated object humans have ever built and shipped commercially. To etch a modern chip, the machine fires a tiny droplet of molten tin into a vacuum chamber fifty thousand times a second. A laser hits each droplet twice. The first pulse flattens it into a pancake. The second pulse vaporizes the pancake into a plasma at hundreds of thousands of degrees, hotter than the surface of the sun. That plasma emits a flash of extreme ultraviolet light at exactly thirteen point five nanometers, a wavelength so short that it is absorbed by air, by glass, by almost anything you might naively try to make a lens out of.

So they cannot use lenses. They use mirrors. Mirrors so smooth that if one of them were scaled up to the size of Germany, the largest bump on its surface would be less than a millimeter tall. Those mirrors are made by Zeiss in southern Germany, in a clean room that took a decade to build and that the company guards like a piece of national infrastructure, which it now effectively is. The light bounces off a stack of these mirrors and lands on a silicon wafer, where it carves features smaller than a strand of DNA, while the wafer itself is moving at meters per second and being held in alignment to within a few atoms by a stage levitated on magnets.

Every step of that sentence sounds made up. None of it is. There are roughly five thousand suppliers in the chain that produces a single EUV machine. The light source comes from a Californian company that ASML had to buy because nobody else could get it to work. The vibration isolation comes from another firm in upstate New York. The optics come from Germany. The wafers and the photoresists come from Japan. The final assembly happens in the Netherlands. The customers, almost all of them, end up in Taiwan and South Korea. Not a single country on Earth could build one of these machines on its own. Not even close. It is, in the most literal sense, a planetary artifact. A thing that the species, collectively and almost by accident, learned to make.

I sat with that for a long time. There is something about the existence of EUV that I find genuinely moving. A machine that fires lasers at tin droplets to vaporize them into artificial sunlight that bounces off mirrors smoother than anything else humans have ever polished, in order to draw circuits finer than viruses, so that someone in a different country can write an email more quickly. Most of the people who use the end product will never know it exists. Most of the people who built the parts will never see the final machine. And yet the chain holds. Year after year. Wafer after wafer.

Once you have seen one of these chains end to end, you cannot unsee them. You start to look at every object in the room and wonder what its EUV is. What is the one strange step, in the one strange place, that almost shouldn't exist, without which the whole rest of the thing falls over. The phone in your pocket has at least a dozen of those steps. The car outside has hundreds. The data center training the model that wrote the autocomplete on your inbox has thousands. Civilization, viewed from this angle, looks less like a smooth surface and more like a cathedral held up by a few load-bearing miracles that almost nobody is paying attention to.

A small number of narrow places

If you sit with the AI supply chain long enough, you start to notice a pattern. Each loud, expensive thing at the top of the stack rests on something quieter and much smaller underneath it. A frontier model rests on a GPU. A GPU rests on a packaged die with stacks of high bandwidth memory glued onto an interposer. The interposer rests on a slot of CoWoS capacity in Taiwan. The high bandwidth memory rests on a hybrid bonder built by one or two companies in Europe. The transceivers that move data between racks rest on indium phosphide wafers grown in furnaces that take two weeks to cycle. The whole thing rests on transformers and gas turbines and a power grid that was not built for any of this.

At each of those layers, the number of credible suppliers shrinks. By the time you reach the bottom you are often staring at two or three companies in the world. Sometimes one. Sometimes a single fab in a single country, run by a few thousand people, whose output decides whether a trillion dollars of capex turns into anything real.

That is the part I find difficult to look away from. Most of human civilization, at any given moment, is balanced on a surprisingly small number of narrow places. You only really see them when something breaks. A volcano in Iceland. A ship in a canal. An export control notice on a Tuesday afternoon. The rest of the time the system feels smooth, and the narrow places stay invisible, and the people who own them get to keep being mispriced.

The four chokes

If I had to compress the thesis into something you could explain on a napkin, it is this. The 2026 to 2030 AI buildout is gated by four physical constraints, and almost nothing else.

Indium phosphide wafers. Every coherent optical transceiver above 800 gigabit per second runs on InP. The world has roughly two credible non-Chinese suppliers of polished four inch and six inch InP substrates. One of them sits in Sumitomo's chemicals division. The other is a small American company most investors have never opened a 10-K for. Their backlog is at record highs. Their capacity is doubling. The market still values them like a sleepy materials business with cyclical exposure to telecom, because that is what they were until eighteen months ago.

Advanced packaging. The world makes a lot of leading-edge silicon. The world makes very little leading-edge packaging. CoWoS, ABF substrates, hybrid bonders, the tools that align two dies to within a few nanometers and fuse them into a single piece of working logic, this is where the actual ceiling on Blackwell, Rubin and everything that comes after gets set. There are perhaps four companies on Earth that matter at this layer, and one of them trades on the pink sheets in the United States because European listings still confuse American brokerages. That mispricing is not subtle. It is a structural inefficiency hiding in plain sight.

Power. A gigawatt of new data center load is not a software problem. It is a turbine problem, a transformer problem, a transmission problem, and a permitting problem. The order book for industrial gas turbines is sold out into 2030. There are three companies in the world that build them at scale. Backup generation, switchgear, and medium voltage transformers are in the same shape. None of this is fixable in a quarter. None of it is fixable in a year. The lead times are physical.

Critical minerals and the rules around them. In late 2025, China temporarily suspended its export controls on gallium, germanium and antimony. The suspension expires on November 27, 2026. That single date is the largest known catalyst on the calendar between now and the end of the decade, and almost no one in mainstream financial media is treating it that way. If those controls re-impose, the only ex-China suppliers of the materials that go into III-V semiconductors and night vision optics and a long list of defense components become the entire global market overnight. If they do not re-impose, the optionality stays priced into a handful of small caps and you wait. Either way, you want to know exactly who sits on the right side of that line.

Why the stock market

I came to all of this through markets, which is an embarrassing thing to admit and also the truth. I have always been the kind of person who picks up a subject and follows it until it stops being interesting, which usually takes a long time. Biology, history, languages, design, philosophy, urbanism. The stock market turned out to be the rare topic that never finished. Every company is a small window into an industry. Every industry is a small window into a country. Every country is a small window into the species.

You can tell yourself you are studying earnings, but you are really studying the world. What people want, what they are scared of, what they are willing to pay for, what they cannot let go of. The price is just the place where all of those things meet on a given day. A 10-K is partly an accounting document and mostly a history book about how a particular group of humans decided to bend a piece of physics or chemistry or distribution into something the rest of us are willing to pay for.

What surprised me about the chokepoint thesis is how cleanly it lined up with everything else I cared about. Geopolitics. Resource scarcity. The slow century-long migration of supply chains. Cold war logic dressed in new clothes. The strange way a single material, mined in one province, can quietly underwrite the entire frontier of a technology that everyone insists is about software. The way a backup diesel generator made in Wisconsin can suddenly become the gating factor on whether a Texas data center comes online in time to train next year's model.

What no one is really talking about

The discourse around AI has settled into two camps. One side talks about model capabilities and benchmarks, as if intelligence were the only constraint. The other side talks about safety and alignment, as if intent were the only constraint. Both groups, in their own way, are arguing about software.

The thing they leave out is the body. The wafer that has to be grown in a furnace for two weeks. The substrate that has to be co-fired in a kiln in central Japan. The bonder that has to align two dies to within nanometers, built by a company most people have never heard of in the south of the Netherlands. The gas turbine that has to be ordered four years in advance from one of three vendors on Earth, because a hyperscaler decided in 2024 that it needed another gigawatt by 2027. The retimer chip that lets a PCIe Gen 6 link run at full rate without burning the bus. The single laser cavity that has to lase cleanly at 1.6 terabits per second or the entire optical link falls back to last generation speeds.

None of this is hidden. It is all in filings, transcripts, trade press, export-control notices, fab roadmaps, hyperscaler RFP leaks. It is just spread across hundreds of small places that almost nobody is reading at the same time. The interesting work is not finding new information. It is connecting the information that already exists.

The arithmetic that nobody runs out loud

Here is the kind of math that quietly drives the thesis, and that almost nobody bothers to do in public.

A modern AI training cluster needs, very roughly, one optical transceiver for every GPU, often more if you are running a fat-tree topology. A million-GPU cluster therefore needs on the order of several million transceivers. At 1.6 terabits per second, each of those transceivers needs a small piece of indium phosphide. The total annual world capacity for non-Chinese InP substrate, even after the doubling currently underway, is small enough that you can write it on a single line of a spreadsheet. The arithmetic does not balance. Either prices move, or volumes get rationed, or both. Markets that do not balance on physical inputs eventually balance on price.

Same exercise on power. A gigawatt-scale data center campus might need eight to twelve large industrial gas turbines, plus equivalent backup. Three companies in the world build them. Their factories are sold out. Their lead times are measured in years. Hyperscaler 2026 capex guidance now starts with a seven hundred billion dollar handle. That money cannot turn into watts faster than the turbines can be cast and shipped. The bottleneck is not capital. The bottleneck is the people who make the part.

You can run this exercise on every layer of the stack and you keep arriving at the same answer. The constraint is never at the loud end. The constraint is always two or three steps upstream, in a company whose investor day has fewer attendees than a mid-tier B2B SaaS earnings call.

Geopolitics is just supply chain with a flag on it

One of the things that becomes obvious once you spend long enough at the bottom of the stack is that geopolitics and supply chain are not really two separate subjects. They are the same subject, looked at from different desks. Export controls, tariffs, sanctions, friend-shoring, reshoring, all of it is governments noticing what specialists noticed years earlier, that a few narrow places control more than they should.

China understood this about rare earths in the early 2000s and built a dominant position over the next decade. The United States understood it about EUV lithography and used it to lock China out of the leading edge. The European Union is currently trying to understand it about photonics and quantum, with mixed results. Japan understood it about specialty chemicals decades ago and has quietly enjoyed the rents ever since. Whichever country owns the narrow place gets a non-trivial say in the shape of the next economy.

What that means for an investor, or honestly for any curious person, is that the geopolitical news cycle is mostly a lagging indicator of supply chain reality. By the time a chokepoint is on the front page, the move is largely over. The prize goes to whoever was patient enough to map the chain when it was boring.

A prism, not a portfolio

AI Bottlenecks is the result of doing this kind of work for myself for a couple of years and eventually deciding that other people might want to look through the same prism. It is not a portfolio. It is not a recommendation. It is not a newsletter trying to flatter anyone into a subscription. It is a public surface where the chokepoints, the names that own them, the live prices, the conviction tiers, the catalysts and the written thesis sit next to each other in one place, so anyone curious can walk the chain themselves and disagree with me in detail.

I wanted it to be free, with no login, because the people I would have wanted to read this when I was younger could not have afforded a Bloomberg terminal and would have been actively excluded by a paywall. Students. Engineers. Journalists who are trying to understand what they are writing about. Retail investors who are tired of being sold someone else's conviction. Policy people who need a map of who actually controls what. Curious teenagers in countries where the local financial press is twenty years behind the actual frontier.

The bet is simple. If you give people a clean view of the physical layer of the AI buildout, with the names and the numbers and the reasoning all visible at once, a meaningful number of them will start asking better questions than the ones being asked on cable television. That alone is worth building.

What I think this exercise is really about

I think a lot about the fact that almost every important shift in the last three centuries can be told as a story about a chokepoint nobody was paying attention to until suddenly everyone was. Saltpeter and gunpowder. Coal and steam. Suez and oil. Silicon and memory. Lithium and cobalt. Each time, a small group of people noticed the constraint early, built around it, and ended up with a disproportionate share of the next era.

AI is going to be the same story. The question is whether you want to read it after the fact in a business school case study, or while it is still being written and the names are still small enough to matter. I would rather read it now, with the physics in front of me, even if I get parts of it wrong. Being wrong about a real thing is more interesting than being right about a fashionable one.

If you take nothing else away from any of this, take this. The story of the next decade is not really about models. It is about whether the physical world can keep up with what we are asking of it. The answer to that question is being decided, right now, by a few thousand engineers in a few dozen factories, most of which you have never heard of. They deserve at least to be visible. The chain deserves to be walked.

A note on curiosity

Somewhere along the way I stopped trying to choose between things. Markets, geopolitics, materials science, history, design, philosophy. They are all the same subject seen from different angles. What humans build, why they build it, what it costs, who gets to decide, what gets quietly sacrificed at the edges so that the loud thing in the middle can exist.

This project is one expression of that. There will be others. If any of it makes the world a slightly less confusing place for one curious person, that is more than enough reason for it to exist.

Open the dashboard, walk the supply chain, argue with the thesis. That is the whole invitation.

Educational, not investment advice.