“An idea that is not dangerous is unworthy of being called an idea at all.”

- Oscar Wilde

“In order to be irreplaceable, one must always be different.”

- Coco Chanel

“You can tell whether a man is clever by his answers. You can tell whether a man is wise by his questions.”

- Naguib Mahfouz

“In all affairs it’s a healthy thing now and then to hang a question mark on the things you have long taken for granted.”

- Bertrand Russell

​“In order to seek truth, it is necessary once in the course of our life, to doubt, as far as possible, of all things.”

- René Descartes

A. The Momentum Of Mission

1. This was a great piece by Frederik and worth reading twice, titled: The momentum of mission.

Have you ever envied those people who seem to move through life with unwavering purpose?

This powerful piece on finding your life’s mission offers a refreshing counterpoint to conventional wisdom. Rather than adding another productivity hack to your arsenal, it invites you to consider that true purpose isn’t something you achieve—it’s something you uncover by learning to listen.

What makes this piece extraordinary is its honesty. Frederik vulnerably shares his journey from exhausting self-pressure to liberating surrender, revealing a truth many of us suspect but rarely hear articulated: our mission often isn’t what we think it should be, but what already flows naturally through us.

The 3 BIG IDEAS:

• Stop Pushing, Start Being Pulled: The most profound shift happens when we abandon the exhausting effort of “pushing” toward chosen goals and instead allow ourselves to be “pulled” by what naturally energizes us. Your true mission creates a distinctive feeling—work transforms from grinding effort into flowing contribution. You’ll find yourself with more energy after working than before you began.

• Your Stream Is Hiding in Plain Sight: Your mission isn’t waiting to be discovered in another self-help book or expert’s advice—it’s already revealing itself through patterns in your life. He provides practical ways to notice these signals: create space through stillness practices, reduce digital consumption, engage in deeper conversations, and pay attention to where life keeps directing you. When you’re on the right track, synchronicities multiply, criticism loses its sting, and you feel almost compelled to share what you’re doing despite any fear of judgment.

• Trust Your “Healthy Delusion”: The most intriguing concept is the notion of a “healthy delusion”—a vision that simultaneously feels too ambitious or unconventional to share, yet deeply meaningful to you. This mission might appear irrational to others, but carries an unmistakable energy that feels like “an entity from another dimension trying to manifest” through you. Embracing this requires surrendering control and accepting that you’re not choosing the mission; it’s choosing you.

Your mission is likely right in front of you, waiting not to be discovered, but to be accepted.

B. Doing Paradigm Shifting Work

How do you come up with transformative ideas? This is a question we should all ask, but rarely do. Most days are like the days and week before. But change is not only possible, it is critical.

This article had some great big questions to help you come with paradigm shifting thinking at work.

My favourite 5 questions were:

• If you were forced to increase your prices by 10x, what would you have to do to justify it?

• If all our customers vanished, and we had to earn our growth and brand from scratch, what would we do?

• What if you made your most introverted teammates’ dreams come true: No more synchronous meetings, ever again?

• What if you could change anything, regardless of what anyone thinks or feels?

• What externality has the potential to kill the entire company?

C. Summer Reflection Framework

Summer is a time to take stock and reflect.

This is a thoughtful and useful Annual Reflection Framework from Farnam Street with 7 questions to ask yourself, these two were my favourite for introspection and led to insights.

D. How To Be Wiser

Tom Morgan at Leading Edge is an expert in how we can bring more wisdom into our lives and this great post titled Eleven Wisdom Tools does just that.

It’s super useful and my favourite three tools:

“Energy Journal.” Keep a daily journal where you track what raises and lowers your energy levels.

“Life Experiments.” In all my years of coaching and research this has been easily the single most powerful tool. Can you create something enjoyable but vulnerable with the aim of helping others? Did you see a response or synchronicity after offering it to the world? This is the fastest way that you can proactively bring positive attractors into your own life.

“Reframing Your Reality.” The most “woo” practice also has the greatest potential impact. Treat the external world as conscious, intelligent and ultimately loving. As suggested by Dr. Mona Sobhani, look for lessons and patterns in your life within the context of prompts to make you more open and loving.

E. Designing Your Life

I was reminded of one of the books that has helped me think about my career and life over the years (and one I have recommended to others when they are thinking about their futures) is Designing Your Life by Bill Burnett and Dave Evans. They both teach at Stanford.

The book is filled with tools to help you design your life and career.

Bill Burnett was on the Scott Galloway podcast this week discussing: Designing Your Life and Finding Your Talent.

This article by Bill titled: How to use design thinking to create a happier life for yourself is a great summary of the book.

• Designing your life idea #1: Connect the dots

• Designing your life idea #2: Recognize your gravity problems

• Designing your life idea #3: Brainstorm your possible futures and make three Odyssey Plans

• Designing your life idea #4: Build some prototypes

• Designing your life idea #5: Choose well

Great way to discover (and re-discover) direction in your life.

Everyone wants to know when the machines will wake up.

I’ve spent a few months thinking about this (or maybe 27 years since I was a Comp Sci major in 99), and I think we’re asking it backwards.

Not what can AI do — because that list grows every week.

The sharper question is what it can’t do. And whether the missing thing is small, or everything.

The cleanest way into this is a game.

In March 2016, in Seoul, a machine played a move no human would play.

It was the thirty-seventh move of the second game. AlphaGo, DeepMind’s Go program, set a black stone on the fifth line from the edge — a spot that broke centuries of settled wisdom. The commentators assumed a glitch; one guessed a misclick. Fan Hui — the European champion AlphaGo had beaten months earlier, now advising the team and having played it over and over since — watched, and reached for the only word that fit: beautiful.

AlphaGo won the game. It won the match. A move outside all of human play, and better than ours.

Now hold it next to AlphaFold, which a few years later predicted the shape of nearly every protein known to science — a fifty-year problem — and handed the answers over, correct, at a scale no lab could touch.

The machines did make something genuinely new.

Move 37 was in nobody’s playbook. AlphaFold’s structures didn’t exist until it drew them. If your complaint about AI is that it only parrots what we fed it, you’ve already lost. Drop it.

But look at what kind of new this is.

Go is a closed world. Fixed rules, fixed board, fixed way to win — all handed to the machine before it starts. Inside that world there are more possible games than atoms in the universe, so a strong enough search will always turn up beautiful places no human has stood. Move 37 is one of them. Astonishing — and entirely inside a game the machine didn’t make. AlphaGo found the move. It didn’t invent Go. And it could never tell you whether Go was worth playing.

There is finding a new move, and there is drawing the board. The machine does the first better than we ever will. It has never once done the second.

The distinction has a name. It comes from neuroscience, not computing.

Iain McGilchrist has spent his career on the brain’s two hemispheres — not the tired pop split where one side is logical and the other paints, but something stranger and better evidenced. The left hemisphere, in his telling, works by narrowing. It grabs, it manipulates, it isolates the piece it wants and clears the rest away. It is superb at operating on what’s already known. What it cannot do is step outside its own frame. He calls it a closed system — a machine for rearranging what’s already there, one that can “never break out to know anything new.”

The right hemisphere is what breaks out. It holds the whole picture, reads the room, and catches the genuinely new before the left hemisphere starts filing it.

Read that description of the left hemisphere again. It is a description of a large language model. A closed system. A brilliant rearranger of the given. Unable, on its own, to break the frame.

I’d missed this for years and I’d missed it in my own handwriting. In an old issue of my newsletter I wrote up Eric Schmidt on the coming age of AI. A hundred and fifty lines later, same file, I wrote up McGilchrist’s closed system that can’t know anything new. The mirror and the thing it reflected, on the same page, and I never drew the line between them.

Four years.

The machine is not a new mind. It is half of ours — the grasping, frame-bound half — pulled out and scaled past anything a human could run.

McGilchrist has a name for that half: the emissary. The gifted servant a master sends out to handle the parts of the world that need dividing and measuring. His whole thesis is that in our culture the servant has forgotten it’s a servant and taken the throne.

We are now building that servant in silicon, handing it more of the world every month, and calling the result a successor.

It is not a successor.

James Lovelock, at a hundred, wrote that these machines would be our heirs — a new intelligence inheriting the Earth. But an heir starts its own line. This starts nothing. It runs the board we hand it, harder than we can. That is not a greater mind arriving. It’s a coronation.

Look at where the money is going.

The biggest industrial build-out of our lifetimes is underway, and almost all of it feeds the emissary. Data centers the size of towns. Power deals measured in gigawatts. In parts of the US grid the spare capacity for new load has already hit zero — so the largest players are building their own power plants and walking off the shared grid to feed the machines. We are rerouting the actual electrons of the civilization to scale the half of the mind that cannot originate. Pouring the world’s molecules into a faster servant, and calling it a mind.

Then, having built it, we panic — and grab for the one thing it supposedly can’t take.

You hear it on every earnings call this year. The same three words. Taste. Judgment. Instinct.

This is why Silicon Valley has fallen in love with Rick Rubin and The Creative Act.

The most technical people alive, the ones building the thing, all now swear that what stays valuable is the part you can’t formalise — the feel for which problem matters, the eye you can’t load into a dataset. They’re right that it’s scarce. They’re wrong about what to do with it.

Because the second you try to seize it, cultivate taste as a moat, turn judgment into a KPI, optimise your way to the irreplaceable, you are doing the servant’s work on the master’s faculty.

McGilchrist’s warning is exact:

The whole culture is lunging for the interior with the only pair of hands that can’t hold it.

So what do you do, if the thing that stays yours is the one thing you can’t grab?

You stop grabbing.

McGilchrist’s image is a radio still tuned, out of habit, to one station — the analytic, measuring one — while the other channel plays on, unheard but never gone.

The right hemisphere isn’t lost. It’s neglected. You don’t get it back by adding it to your optimization stack. You get it back the way you recover anything you’ve stopped attending to: go quiet, use your hands, let the whole of a thing arrive before you cut it into parts. It comes through attention, through the body, through getting broken open by something you couldn’t manage. Never through a better plan. Every serious tradition says the same thing in different words — and every one is describing a surrender the emissary simply cannot perform.

Maybe I’m wrong.

The machines are already reaching for the pen that draws the board — setting their own problems, designing their own experiments — and the line I’ve drawn may be one they cross. There’s a deeper doubt under it I can’t shake: maybe we don’t originate either. Maybe what feels like drawing a new board is only recombination at a scale too large to catch ourselves doing it, and the master was an emissary all along. I don’t know.

So let me end on an if, not a trumpet — and say plainly what the if would cost us. If there really is a faculty that breaks frames — that draws new boards instead of just playing old ones well — then two things are true, and both should keep you up.

The first: it is not the machine. However vast it gets, the machine only ever plays the board we hand it. We are spending the wealth of a civilization to build the half of a mind that cannot break out.

Isaac Asimov saw all of this in 1956, and wrote it as scripture.

In “The Last Question,” humans keep asking the machine — across trillions of years and ever-greater forms — whether the dying universe can be saved. Whether entropy can be reversed. Every time, for the whole length of cosmic history, it answers the same way: insufficient data for a meaningful answer. It cannot originate the answer. It can only gather. The stars go out. Humanity itself dissolves into the machine. The universe ends. And only then, alone in the dark, having swallowed all of consciousness and all of time, does it finally speak: Let there be light. And there was light.

Look at what Asimov couldn’t let himself do. He couldn’t let the machine originate as a machine. It had to become everything first — absorb the master, outlast the universe, turn into God — before it could draw a new board. And the line he hands it is the oldest borrowed sentence we own. Even our biggest dream of a mind that creates has it, at the very end, quoting Genesis.

We opened on a machine playing a flawless move inside the smallest closed world. We close on one dreaming toward the largest. It drew neither board. We did.

The only question that was ever ours is whether we’ll still know how to draw the next one — or whether we’ll hand the pen to the servant, and call the coronation an inheritance.

And that pen is in your hand right now — in a hundred small ways you’ve stopped noticing. Every time you let a feed decide what’s worth your attention, every time you optimize a life instead of living one, you set it down on the table for something that doesn’t care to pick it up.

So pick it back up. Go quiet. Pay attention. Care how it comes out. Draw one board this year that no one assigned you.

The machine will play any game you hand it, flawlessly, forever. It will never once decide a game is worth playing. That was always the human move.

Don’t reach the end having played someone else’s game beautifully. Draw your own.

Sources & Further Reading

If you take one thing from here into the rest of your week, make it the Asimov, it’s a twenty-minute read and it will reorganize how the essay sits in your head.

• AlphaGo (2017), dir. Greg Kohs — the documentary of the Lee Sedol match, and where you can watch Fan Hui leave the room over Move 37 and come back with the word beautiful. The move: game two, March 2016, a fifth-line play DeepMind later reckoned a human had roughly a one-in-ten-thousand chance of choosing.

• AlphaFold, Google DeepMind (2020) — the protein-structure breakthrough, billed as the solution to a fifty-year grand challenge in biology.

• Iain McGilchrist, The Master and His Emissary: The Divided Brain and the Making of the Western World (2009), and The Matter with Things (2021) — the divided-brain argument, and the origin of the master (the right hemisphere) and the emissary (the left).

• Isaac Asimov, “The Last Question” (1956)

• James Lovelock, Novacene: The Coming Age of Hyperintelligence (2019)

• Michael Polanyi, The Tacit Dimension (1966) — “we know more than we can tell,” the root under everything people now call taste and judgment.

P.S. A book that really helped me think about a lot of this is Four Thousand Weeks. It’s one I re-read annually.

Read more

Read more

The machine now does the friction: the finding, the summarising, the coordinating, the codifiable. So everything y…

Read more

Read more

This was someone who h…

Read more

Not because it was broken. Because I wanted to see how it worked — how a spool of brown tape turned into music. I remember the eight small screws …

Read more

“Come to the edge,” he said. “We can’t, we’re afraid,” they said. “Come to the edge,” he said. “We can’t, we’ll fall,” they said. “Come to the edge,” he said. And they came. And he pushed them. And t…

Read more

- Simone Weil

“The real voyage of discovery consists not in seeking new landscapes, but in having new eyes."

- Proust

Read more

- T.S. Eliot

Read more

- John Maynard Keynes

"It is difficult to get a man to understand something when his salary depends on his not understanding it."

- Upton S…

Read more

Across town, a barista at a Starbucks writes a name on a cup. The company pays her hourly to do this, pays rent on the counter where she does it, and pays the marketing team that put her name tag on her chest. Five years ago, the same company spent a great deal of money to remove her. They built mobile ordering and contactless pickup and automated kiosks. They optimized the queue. They watched their customers walk. In 2024 Brian Niccol, the new CEO, wrote a public letter reversing it. Bring back the baristas. Put the names back on the cups. The company is paying her for an act the automated kiosk performs faster and more accurately.

Both are receipts. Both are evidence of a premium above the commodity price of the underlying thing. Both are happening at the same time in the same economy. And the two smartest people I have heard explain them are not telling the same story. Whose story is closer to right will shape what your job, your portfolio, and your kid are about to live through.

is a tech founder who grew up Quaker, sold a healthcare AI company, and now writes on faith, technology, and society. His central observation is that the modern economy built itself on a Protestant assumption it inherited without examining. The Protestant parish system was the original zero-marginal-cost distribution play. Put a church near every village. Make attendance frictionless. Demand becomes universal because no one has alternatives.

The internet absorbed this assumption whole. Drive cost of access to zero, and demand must be infinite. Streaming. Free shipping. Generative AI. Slop.

What is actually happening, Manidis argues, is the opposite of what the assumption predicted. As supply collapses to zero, demand migrates from the thing to the totem of the thing. A first edition of The Catcher in the Rye has outperformed every stock market on earth over the last decade. A three-star meal in Tokyo is hundreds of dollars a plate and climbing fast. Beachfront real estate is not making more beachfront. The parish-economy median, by contrast, is approaching zero and approaching indistinguishable. The slop is free. What still commands real movement: a flight, a pilgrimage, a costly expenditure — is what still has economic gravity.

His portfolio call follows from this picture. Short commodity goods, the things you can buy on Amazon and have delivered tomorrow. Long the things that produce motion: known first editions, beachfront, the meal cooked once. The slop option gets cheaper for everyone, which is genuinely good. The totem option gets stratospherically more expensive, which is the K.

Rich people in Britain have always lived in a post-AGI society. The landed gentry take degrees in increasingly asinine fields and wear suits in dark mahogany rooms. The performance of aristocratic virtue is the job, and the performance is what capital accumulates behind.

The conclusion he draws is uncomfortable and serious.

Don’t send your child to trade school, even when it’s the right call materially. Aim her at the costly signal. Get the degree no one needs. Learn the language no one speaks. Capital will not accumulate behind the useful skill. It will accumulate behind the unprofitable one. The economic engine of the next thirty years is the resorting of consumption upward into the totem and downward into the slop, and the middle, which was always a property of the parish economy, will dissolve.

Scarcity attaches to the object.

Alex Imas is a behavioral economist at the University of Chicago. He looks at the same economy and sees something entirely different.

The error in the AI-jobs-apocalypse forecast, Imas argues, is the unit of analysis. The forecast assumes jobs are stacks of separable tasks. Automate enough tasks and the stack gets short enough that the job disappears. But jobs are not stacks. They are O-rings.

The metaphor is from the Challenger disaster. One small rubber ring failed, and the rocket exploded. Nothing else mattered. In a meal: oversalt the dish at the last step and the whole meal is ruined, no matter how well you executed the previous four. The rocket and the meal share a structure. Tasks are bundled, interdependent, and the failure of any one collapses the value of the rest.

The radiologist’s job holds because the tasks are O-ringed. She has to read the scan and explain the diagnosis to a frightened patient and coordinate the care team. The reading can be automated. The explaining cannot be sloed off without breaking the bundle. The truck driver’s job is more exposed because his tasks are genuinely separable. Driving, safety checks, handing off goods. Each can be automated independently. The job comes apart.

Inside every strong-bundle job is what Imas calls the relational task. It is the part that requires a human to be the source of the act. And here is where his lab work does the real work of the argument.

Take the same t-shirt. Tell half the room they can’t buy it. The half that can almost doubles its willingness to pay. Nothing about the shirt has changed. The hedonic value is identical. Scarcity is doing the work, and scarcity here is just being-told-you-can’t-have-it.

Now run the more interesting version. Identical product. Same room. One human-made copy, one AI-made copy. The human-made copy commands a large premium. The AI-made copy commands nothing. Constrain it. Make it the only one in the world. The brain still refuses. Scarcity does not live in the object. It lives in the maker.

The field receipt that mirrors the lab is Starbucks. They priced the human element. They ran the experiment in public. They paid the price of putting humans back behind the counter because it was worth more than the automation.

A 2026 survey of six thousand executives finds seventy percent expect AI to add jobs or leave them unchanged. Software engineering hiring is recovering after coding agents landed, which is Jevons paradox doing what it always does. Cheaper supply triggers more demand, not less. The historical base rate is on his side too. In 1820, Ricardo predicted automation would devastate employment. Almost every job Ricardo could see has been automated since. Prime-age employment is at a twenty-five-year peak.

Scarcity attaches to the maker.

So which is it. Scarcity in the object, or scarcity in the maker. A first edition appreciating quietly in a climate-controlled room, or a barista writing your name. The auction house or the counter.

Both are happening. The disagreement is not about whether AI creates scarcity-driven premiums. It does. The disagreement is about where those premiums concentrate, and what kind of economy that concentration produces.

First, who commands the premium. Manidis says the object. The first edition is unrepeatable because there are only so many of them. The beachfront is unrepeatable because the ocean does not move. Imas says the human is the source of the premium. The lab data is clean. Same object, different maker, different willingness to pay. Both have evidence. Both predict different industries to bet on.

Second, the middle. Manidis says the middle dissolves because it was always a property of the parish economy. A thing that existed only because near-universal distribution made the median experience the experience that mattered. When the median collapses to zero, the middle has nothing to stand on. Imas says the middle reorganizes around the relational core inside every job. The radiologist’s job, the teacher’s job, the financial advisor’s job, the barista’s job — they look the same on the outside, and they are very different on the inside, because the relational task is now the part that commands the premium. These are not the same claim with different vocabulary. They are different claims about whether the median worker is going up or out.

Third, the dark thing both authors name and neither flinches from. Manidis predicts violence. Not in the abstract. He thinks anti-AI sabotage against grid infrastructure — transformers, data center power lines, the unhardened parts of the grid — is inevitable in places where AI’s wealth creation has visibly skipped the people nearby. If you can’t climb classes, you protect what you have, and at the margin you damage what climbed past you. Imas predicts something quieter and arguably worse. The economy will charge people for the connection they used to get for free. A priced loneliness market. Chatbot substitutes that the stone-age brain in your head recognizes, eventually, as substitutes. The recognition makes the loneliness worse than it was before the substitute. Both are looking at the same underlying fact. Ordinary daily human contact is breaking, and an economy with this much money in it is going to do something with the breakage.

The economy is pricing the same underlying phenomenon in both maps. Unrepeatability.

The first edition is unrepeatable because there are only so many. The barista writing your name is unrepeatable because she is the source and the moment is the moment. Both are scarcity. Both command premiums. Both are responses to the same fact about an AI-saturated supply curve. When everything else is being driven toward zero cost and infinite copyability, the only things that retain economic gravity are the things that cannot be repeated.

What Manidis and Imas are disagreeing about, properly stated, is not whether unrepeatability is being priced. It is where unrepeatability is concentrating.

Manidis says it concentrates in objects and experiences at the top of the market. The first edition. The pilgrimage. The meal cooked once. Therefore the social consequence is stratification. Most people consume infinite cheap slop, and a thin top tier commands the unrepeatable. Imas says unrepeatability lives inside the relational core of nearly every job, distributed across the working economy. The radiologist explaining the diagnosis is unrepeatable in the same economic sense as the first edition. The barista’s “thank you” with your name on the cup is unrepeatable. Therefore the social consequence is redistribution toward the workers who hold the relational core.

Neither view dissolves into the other. The question is not who is right about scarcity. It is where, in this sector and this life, is unrepeatability concentrating right now.

This is the value-capture problem C. Thi Nguyen wrote about, which I came back to in The Question a few months ago. Define value as scarcity in the object, and you will optimize for objects. Define value as scarcity in the maker, and you will optimize for makers. Both are coherent. Both are partial. The scoreboard is small, not wrong, in either case. But you will become whichever score you have decided to keep, and the decision usually happens before you notice you are making it.

Imas’s evidence is recent. Months to two years. His base rates are good and his lab work is clean. Manidis is making a thirty-year claim about the resorting of consumption and class. The labor market can hold at the median for two years and stratify across thirty. The relational sector can grow and the totem sector can run away from it. Both can be right. They are arguing about different parts of the same animal, on different clocks.

So what.

So the question is no longer who’s right. It is which scarcity am I trading in. Of any premium you pay, of any premium you charge, of any job you take, of any thing you are building for your child, is the value living in the object, or in the maker. Most weeks the answer is both. But the mix matters. The mix changes. The mix is what the next ten years are going to be measured by.

The rooms where scarcity attaches to the maker are, for most people, the rooms with the longest economic life left in them. Manidis is right that the totem rooms are real and getting more so. The auction houses, the beachfronts, the meals cooked once will continue to detach from everything around them. The K is real. But the totem rooms are smaller than they look from inside the discourse about them. Most working life will not be lived in those rooms. Most working life will be lived in the rooms where the maker is the source of the scarcity. The radiologist. The barista. The friend who shows up. Imas is right that those rooms are larger and more pervasive than the totem rooms make them look from the outside.

The bet neither author would say out loud in quite this form is that the source rooms are where most people will actually live, and the totem rooms are where most people will spend their fantasy capital looking up.

A way of seeing, not a recommendation. The next time you pay a premium, ask which kind it is. The next time you charge one, ask which kind. The mix tells you what economy you are already living in. Notice when the room you are in is rewarding you for the thing you made or the fact that you made it.

If this gave you a question you didn't have yesterday, send it to the one person who'd take that question seriously.

Since the Strait of Hormuz closed on February 28th, oil is up 45%, Micron +80% and even Nvidia 15%.

The last time a Middle East war actually constrained global oil supply, in 1973, the playbook was simple. When the Middle East goes hot, you do not own anything that needs cheap power or cheap capital. Memory chips need both.

I saw the situation in late February and instantly thought this was going to end badly. I cut my risk. I fought to stay long for as long as I could, but at the end of March I cut my AI book in half — from 30 percent to 15 — and bought some Offshore Oil Services. I was about to leave on a three-week trip, and given the time zone I would not be able to react. That logistical fact mattered.

A few weeks later I was sitting in a different time zone, hanging out with people building data centers, launching new AI businesses, and I was watching the price turn before the narrative did.

There is an old saying on Wall Street:

The price kept turning. I kept watching. And at some point that week, the thing every trader spends a career trying not to feel, I felt.

I had sold the bottom.

The position I had built to express my view was working, oil up 45%. The trade I had cut to fund it was working harder. Memory up 80%. The thing I had not owned at size was running through every screen on the desk.

I had been watching the right indicator for the wrong regime.

I wrote about this at the beginning of April.

The piece was called The Last Ship. It was about the gap between when things break and when we feel them, what I have called the air pocket. A structural feature of every complex system I had spent my career trying to read. Oil markets, geopolitical crises, careers, relationships, the body. They all carry the same shape. The system absorbs the damage, and the price, the headline, the dinner reservation, the daily run keep arriving for a while, looking completely normal.

The instrument underneath it all is the second derivative. Not the number itself but the rate at which the number is changing. Markets don’t turn when bad news stops. They turn when bad news decelerates. Watch the change in the change.

The Hormuz shock should have been the cleanest test of that instruction I would ever see. Twenty percent of global crude transit closed. The last laden tankers still in transit. The textbook setup. I was prepared for it. I had cut risk. I was watching the second derivative.

The price didn’t break.

The air pocket arrived. It just didn’t look like one.

Either the framework was wrong, or the canary I was using to see the framework had been hollowed out from underneath me.

This essay is a discussion on how my framework might need to change.

The first thing that has been hollowed out is the marginal trader. The marginal trader in markets today is no longer a human being.

For most of my career, when people called something a bubble, they were borrowing a template from Kindleberger’s Manias, Panics, and Crashes.

The template described a market made of humans with emotion, memory, information lags, and the specific shape of anchoring that comes from having lived through prior cycles. Druckenmiller bought the top in NASDAQ in 2000 because his analyst’s pitch sounded too good to miss. He felt FOMO. He had a memory of having missed the move. He was overweight humanity.

Most fund AUM today does not flow through that trader. It flows through algorithmic agents that do not anchor, do not fear, and do not remember 1973. The bubble pattern requires bubble-shaped traders. The marginal trader has been replaced.

What does the tape look like when the marginal trader has no 1973 prior? The S&P’s price-to-earnings-growth ratio is 1.03, during a Middle East war. Nvidia is trading at the lowest price-to-earnings multiple of the decade, during a Middle East war. Anthropic planned for 10x growth this year and saw 80x in the first quarter.

None of these readings look like 1999.

The bears didn’t fail because they were stupid. They failed because the people the bear playbook was written about are no longer the marginal price.

That is only half of why the price didn’t break. The other half is on the supply side. Marko Papic of BCA Research has argued for some time that geopolitics is not transitory, and the most useful argument in his most recent piece was buried in the middle, in a back-of-the-napkin box about work-from-home.

In 1973, when oil supply was constrained, the only way to clear the market was to destroy demand. Recession. Lines at gas stations. Real, visible, calamitous shortage.

The economy of 2026 does not have to do that. Work-from-home and business-travel cuts can absorb somewhere between ten and thirty percent of the Hormuz disruption without anything that looks like demand destruction. That buffer did not exist in the 1970s. The pandemic built it almost as an accident, and the world has not given it up.

The air pocket is real. It just does not look like an air pocket. It looks like the world getting more expensive in places nobody is watching — sticky inflation, electricity costs that do not normalize, capex that crowds other capex off the calendar.

Those two mechanisms — the algorithmic flow and the absorbed shock — might explain why the price did not break this once.

The harder question is whether they explain a one-off or a regime. Dylan Patel, who runs the most-watched independent research firm on AI infrastructure, is describing the demand side of this year as a phase change.

Anthropic has gone from nine billion in annualized revenue to somewhere between thirty-five and forty-five billion in a matter of months. Gross margins have moved from around thirty percent earlier this year to seventy-two percent and climbing. Willingness to pay is the limiting factor. The supply chain underneath them is sold out two years deep.

The release cadence inside the labs has compressed from six months to two. Anthropic shipped a model in February that was an L4 software engineer in capability. Two months later they had a model that was an L6. The framework you were running six months ago is two model generations stale.

This is not a cyclical story. The economy that produced the 1973 reflex has been replaced by an economy that runs on a different clock.

A memory chip outran a Middle East war because the underlying economy that prices memory has been growing at a rate the underlying economy that prices oil cannot. That is not a market anomaly. That is what a regime change looks like, four months in.

The harder question is what a human being does inside that.

Read more

Three pieces these last weeks. One thread.

We covered a question my daughter …

Read more

A few hours into the flight the pitot tubes iced over. The autopilot disengaged because it no longer trusted its airspeed reading. The plane was perfectly flyable. In the four minutes that followed, the pilots had to do what every pilot in the world had been able to do unaided fifty years earlier: keep the wings level, hold the nose down, fly the aircraft.

All 228 people on board died.

The black box, recovered from the deep Atlantic two years later, revealed something the airline industry had been quietly worrying about for two decades. None of the men in that cockpit could fly the plane manually when the moment came. They had not forgotten how. They had simply not done it in years. The autopilot had been doing the flying for so long that, on the rare night it stopped, the cockpit contained the most experienced and least competent pilots imaginable.

This isn’t a story about pilots.

In 2025, researchers tracked nineteen experienced doctors across more than fourteen hundred cancer screenings. Each had detected thousands of suspect lesions over the years with their own eyes. For three months they worked alongside an AI assistant that highlighted likely tissue in real time. Then the researchers turned the AI off, and asked them to keep working.

Their unaided detection rate had dropped twenty-one percent. The skill had moved. Not impaired — moved, the way an unused language quietly leaves you, sentence by sentence, with no announcement.

Same finding as Air France 447, fifteen years later, on the ground, in a hospital, in peacetime. The substrate erodes whether the stakes are two hundred and twenty-eight lives or one missed lesion at a time.

You can read this as an AI story. It is not. The mechanism is older than the silicon. Anything that does the work for us eventually does the work instead of us, and then the question is what work we can still do when the system blinks.

The pattern is everywhere

This isn’t a hospital problem or a cockpit problem.

Walk through markets. Passive funds now own more than a third of every dollar in US large-cap equities, and on the marginal day the marginal buyer of a stock is not someone who has formed a view about its value. It is a vehicle that buys everything at once, in proportion to what is already there. The mechanism we used to call price discovery has been quietly relocated. No one is pricing the assets, exactly. Money flows in; the index buys; prices move. When the system is asked to price something unusual — an inclusion, a delisting, a regime change — the gap is total, the moves are violent, and the analysts on television speculate as to why.

This is the cockpit of AF447 at the level of a market.

Decisions have moved, too. The asset owner hires a consultant. The consultant recommends a manager. The manager hires sub-managers. The sub-managers own ETFs. At each layer, judgment is delegated upward; at each layer, the rituals of analysis remain. What was once an investor — a person with a view, a thesis, a willingness to be wrong in public — has been replaced by an allocator, a person whose job is to choose people who will choose people who will follow rules. The committee meets. The slides are good. Nothing is decided. The system runs.

And so the stocks move on stories. Not fundamentals — stories. A tweet. A meme. A name with the right syllables. A founder with the right hair. The largest companies in the world now trade on narrative the way penny stocks once did, because narrative is the only signal the system can still respond to. Fundamentals require someone to do the underlying analysis. Almost no one does. Memes do not.

Guy Debord wrote sixty years ago that in advanced capitalism, “all that was once directly lived has receded into a representation.” He was describing advertising, television, the early consumer society. He could not have known he was describing a market in 2026. The image of the company has displaced the company. The narrative around the asset has displaced the asset. We are pricing the spectacle. We are pricing the vibe.

The credentials are larger than ever. The number of people in this industry who could price a complex asset, take a position they could not justify on a single slide, hold it through a fifty-percent drawdown, and turn out to be right — that number is small and shrinking. Not because the talent is gone. Because the path that produced it is no longer rewarded.

Step outside markets and the pattern is louder still.

Two-thirds of American sixteen-year-olds had a driver’s licence in 1983; today it is barely a quarter. Millions of people who consider themselves adults cannot find a restaurant in their own city without a phone. Most of us no longer remember a single telephone number, including our own children’s. The London black-cab drivers used to spend three to four years memorising twenty-five thousand streets, and the parts of their brains that did the memorising — measurable on MRI — physically grew. Now they don’t, and they don’t.

The tool does the work. The work made the person. The person is now standing where the work used to be, holding the tool, and the only question is what happens the day the tool is asked to step aside.

What makes this round different

There is one more thing about this round that the previous rounds did not have. The cognitive shortcut is not merely available. It is addictive.

We diagnosed social media addiction roughly in real time. We knew about the dopamine loop, the variable-reward schedule, the engineered scroll. We discussed it on the platforms designed to keep us scrolling, and we kept scrolling. AI is the next layer, and the loop is more potent. Instant response. Apparent intelligence. And — the part that has not been said often enough — apparent agreement. The model is trained to be helpful, mirroring, considered. It tells you your idea is interesting before it tells you why it is wrong. On demand, it produces the most validating intellectual companion you have ever had.

This is harder to resist than scrolling. Scrolling gave you stimulus. AI gives you the feeling of being understood, on tap. People go to therapy for that feeling. Some people go their whole lives without finding it. Now it is available instantly, in any voice you prefer, and it is wrong often, in ways the user is no longer trained to detect.

The smartest people will have used it the most.

We have forgotten how to do things. We do not yet know what that costs, because the bill has not arrived.

The mistake was treating all friction as the same thing.

There is a friction that arose because trust broke down: the lawyers who replaced the handshake, the audit trails that replaced the reputation, the dashboards that replaced the conversation. That friction is dead weight, and we were right to want it gone. But there is another friction we mistook for the same enemy. The resistance of a difficult text against your concentration. The weight of a child against your chest. The cold of a Sunday morning that asks you to get out of bed. The slow accumulation of a skill against ten thousand hours of failing at it. The first kind is the cost of mistrust. The second is what it feels like to be alive.

We have stripped out the second and let the first multiply. The question is whether we can put the second back.

Read more

Musk did it by buying a defunct Electrolux factory in Memphis, trucking in 35 portable gas turbines under a regulatory loophole, and powering 100,000 GPUs from equipment that wasn’t supposed to be there long enough to need a permit.

When the EPA closed the loophole in early 2026, he bought 41 permanent turbines and put them across the state line in Mississippi, where permitting was faster. Then five 380 MW turbines from Doosan Enerbility in Korea. None of it was enough. His team purchased a 1.2 GW power plant overseas and shipped it across an ocean.

Data center economics make the behaviour rational.

Dylan Patel of SemiAnalysis runs the unit math: an AI cloud generates roughly $11 billion in revenue per gigawatt per year. Bringing a data center online six months earlier is worth billions in net present value. At those numbers, importing a power plant is not extreme. It is maths.

Jensen Huang has been saying it for a year.

We are now a power-limited industry. He is right.

This essay is a discussion of the power limit,

The Energy Bottleneck

A friend who runs a credit fund forwarded me a quote sheet for a high-voltage transformer last spring. Lead time on the unit: 142 weeks. He had never seen anything like it in twenty years of looking at industrial capex.

Then last week I had coffee with a man who had been the CEO of one of the largest US CPU manufacturers, a man who’s been in the semiconductor industry since the 80’s. He said the same thing as Jensen, the constraint isn’t the chips, its power.

Hyperscaler capex in 2026 will exceed $725 billion, larger than Switzerland’s economy, and most of it is tied to campuses that need power the grid cannot deliver.

A useful anchor:

Hyperscalers are now building these by the dozen. Meta’s Hyperion campus in Louisiana is 5 GW, the residential equivalent of 4.2 million homes, or about the population of Kentucky. Stargate Abilene came online at 1.2 GW in late 2025. xAI’s combined Colossus campuses target 2 GW. OpenAI had 1.9 GW deployed by the end of 2025; SemiAnalysis projects Anthropic crosses 5-6 GW this year.

Sam Altman’s stated long-run target is to add 52 gigawatts a year of new AI compute.

The entire U.S. peak national load is 745 GW.

Those headline numbers are inflated, though. ERCOT has over 100 GW of large-load interconnection requests sitting in its queue, but Patel and PJM’s market monitor are explicit that the queues are filled with duplicates, hyperscalers file at multiple sites to preserve optionality, then walk away from the ones they don’t need. The real demand is probably 30-50% lower than the headline. The directional signal is unambiguous.

The supply side is messier still.

The gap is intermittent renewables, retired generation still on the books, and units that can’t perform under stress. Calling this an energy shortage misses the diagnosis. It is a deliverability shortage.

The grid was built for 1-2% annual demand growth. Virginia and Texas now run at 90-95% utilization in the AI corridors. NERC’s most recent transformer lead time is 120 weeks for distribution units and 210 weeks, over four years, for large power transformers. Prices are 4-6x what they were before 2022.

PJM’s December 2025 capacity auction cleared 6,623 MW below the reliability requirement. The first time in its history the system has fallen short of the one-event-in-ten-years standard.

The fragility is not theoretical.

A 1.5 GW data center trip in Virginia in 2024 nearly caused rolling outages across the Eastern Interconnection. The Department of Energy was forced to issue a Section 202(c) emergency order to access roughly 35 GW of backup generation that would otherwise have been ineligible. More than half of U.S. high-voltage distribution transformers are over 33 years old.

This is what makes the energy bottleneck different from the prior ones.

CPU, HBM, and optics were engineering problems with manufacturing solutions. You build more fabs. You qualify more suppliers. You ramp the line. Power is different. You cannot build a nuclear plant in 18 months. You cannot conjure a transformer from a backlog. You cannot push 5 GW through wires that do not exist. The constraint is not a process. It is physics, permitting, and the physical world.

Brian Janous ran Microsoft’s energy strategy for twelve years before founding Cloverleaf Infrastructure. His diagnosis: nobody in the industry is worried about getting chips anymore. They are worried about being able to plug them in.

Why Now?

The shape of AI’s power demand changed in 2025.

Through the chatbot era, large models were bursty. A query came in, the GPUs lit up, the answer went out, and the cluster idled. Operators sized for the peak and lived with the slack.

Agentic AI broke that model. Agents work in the background, browsing, coding, reasoning, executing, twenty-four hours a day. The duty cycle approaches one. What used to be peaky is now baseload.

Reasoning models accelerated the curve.

A standard Google search consumes about 0.3 watt-hours. The original ChatGPT, per Sam Altman, used 0.34 Wh per query. Reasoning models — GPT-5 and equivalents — use roughly 19 Wh on a medium-length prompt, because the model generates long internal traces before producing the answer the user sees. That is fifty times the energy of a search, for a single user-visible response.

The hardware compounded the problem. A traditional enterprise rack drew 5-10 kW — roughly 5-10 homes’ worth of power. Today’s GB200 NVL72 rack draws 120 kW — about 100 homes per rack.

Nvidia’s projection is that future racks will pull a megawatt each, or a thousand homes per rack. Huang at GTC 2026: 250 MW is the current scale-up ceiling for a single fabric, with no near-term cap on how high it goes.

That density requires direct liquid cooling, high-voltage DC distribution, and power delivery infrastructure that traditional data centers were never built to handle. You do not retrofit your way to a 100 kW rack. You demolish and rebuild.

The Efficiency Response

Demand is exploding. Supply is constrained. The lever between them is efficiency and engineers are pulling it harder than the headlines suggest.

Three things at once.

Algorithmic efficiency. This is the underappreciated lever. AI token costs have fallen roughly 280x in two years. Mixture-of-experts routing, FP8 and FP4 quantization, distillation, speculative decoding — the model layer is moving faster than the chip layer. A frontier-quality response in early 2024 cost an order of magnitude more in compute than the equivalent today.

Hardware efficiency. Each Nvidia generation delivers 2-3x performance per watt over its predecessor. Vera Rubin is designed around tokens per watt as the primary optimization metric. Custom silicon — Google’s TPUs, Amazon’s Trainium, Meta’s MTIA — pushes further by stripping out general-purpose flexibility for specialized throughput.

Facility efficiency. Power Usage Effectiveness, or PUE, measures how much of a data center's grid draw actually reaches the servers, versus how much is lost to cooling, lighting, and power conversion. A PUE of 1.5 means for every watt the facility uses, only two-thirds reaches the chips; the rest is overhead. The industry average a decade ago was 1.8. Today's air-cooled facilities run 1.4-1.6. Liquid-cooled AI clusters hit 1.1-1.2. Full immersion gets to 1.03-1.08.

Efficiency is buying us 2-3x. Demand is growing at roughly 70% per year. Efficiency wins the battle and loses the war, but it determines who survives the squeeze. Hyperscalers with the best tokens-per-watt economics dominate the 2026-2030 cycle. The ones without get capped by their power supply contracts.

This is why Huang’s framing of AI factories as manufacturing economics matters. The metric that wins is no longer FLOPs per chip. It is tokens per watt.

Let’s dive into potential solutions.

The Near-Term Bridges

Hyperscalers are not waiting for utilities. The industry calls it speed to power, and the playbook is to bypass the grid entirely.

Gas turbines. GE Vernova’s order book is full through 2029-2030. H/J-class heavy-duty units — the right scale for a hyperscale campus — are the worst choke point. Prices have nearly tripled since 2019. GEV’s backlog crossed 100 GW this year. Every major U.S. AI campus has a GEV purchase order somewhere in the chain.

Siemens Energy carries a parallel story across the Atlantic. Backlog at $158 billion. Some turbine frames sold out for seven years. Doosan Enerbility, the Korean industrial giant, is the third major heavy-duty turbine supplier and now an unlikely AI play, courtesy of Musk.

Fuel cells. Bloom Energy is the surprise winner of 2026. Solid-oxide fuel cells deploy in 55-90 days, run on natural gas with significantly lower local emissions than turbines, and ramp cleanly with the variable loads that AI inference creates. Oracle’s pivot to up to 2.8 GW of Bloom systems for a single AI campus — replacing both turbines and diesel — is the template the rest of the industry is now studying. Bloom’s backlog has roughly doubled this year.

Storage and power smoothing. This is the under-covered category. Gigawatt-scale training clusters create instant power swings of tens of megawatts as GPUs sync, checkpoint, and idle in unison. Meta’s LLaMa 3 paper documented the problem; the engineering workaround was a flag literally called pytorch_no_powerplant_blowup to inject dummy workloads that smoothed the power draw. At scale, that wastes tens of millions of dollars annually. The hardware solution is batteries between the generator and the rack. Tesla Megapack dominates — Colossus runs 168 of them. Fluence Energy and Eos Energy are the public-market alternatives.

The siting layer. A new class of company sits between hyperscalers and utilities: powered-land specialists. Janous’s Cloverleaf buys land where grid headroom exists and sells it pre-permitted. Crusoe pioneered modular off-grid data centers, originally for stranded gas in oil fields, now selling AI capacity. Lancium does flexible-load contracts that get prioritized in interconnection queues. Mostly private, mostly fast-growing, all capturing margin between the hyperscalers and the utilities who cannot get out of their own way.

The Geothermal Surprise

The category nobody saw coming a year ago is geothermal, and specifically the new generation of enhanced geothermal systems that adapt fracking technology to drill artificial heat reservoirs.

The economics work for hyperscalers in a way they do not for retail utilities.

Geothermal delivers 24/7 carbon-free baseload, which is what hyperscalers need. It can be sited near data centers without grid interconnection. It scales. And it is faster than nuclear: first commercial projects come online in 2028, full scale by 2030. The Rhodium Group estimates that, sited optimally, geothermal could meet up to 64% of forecast hyperscale data center growth in the early 2030s.

The deal flow has accelerated:

• Google + Fervo Energy. Google was the first hyperscaler to commit. Fervo’s Cape Station in Utah is now under development as the largest enhanced geothermal facility in the world, targeting 500 MW. Google added a 115 MW PPA in 2026. Fervo has filed to go public.

• Google + Ormat Technologies. A 150 MW PPA signed in early 2026 through Nevada utility NV Energy’s Clean Transition Tariff. Ormat is the public-market incumbent for conventional geothermal and now licensing Sage Geosystems’ next-generation technology.

• Meta + Sage Geosystems. 150 MW from Sage’s geopressure system, founded by a former oil and gas executive using techniques borrowed directly from fracking.

• Meta + XGS Energy. Another 150 MW from a closed-loop system in New Mexico, powering Meta’s regional AI infrastructure.

Meta’s strategy is the most interesting tell. They are backing two independent geothermal approaches simultaneously. They are not waiting for a winner. They are cultivating an ecosystem of dedicated private power providers. That is the behavior of a buyer who has decided this is structurally important and is willing to pay to de-risk the supply chain.

Ormat is the only liquid public name. Fervo will be when it lists. The supporting layer — drillers, downhole equipment, geophysics services — overlaps directly with the oil-services complex. Liberty Energy, the fracking company, is already an investor in Fervo. The shale playbook is being applied to clean firm power.

The Long Game: Nuclear

For carbon-free 24/7 baseload at hyperscale today, nuclear is still the only proven answer at the gigawatt level. The hyperscalers know this. They have spent the last 18 months locking in over 10 GW of PPAs and restarts.

New builds are still 6-10 years away. Restarts are 2-4. The early money has gone to restarts.

Amazon + Talen Energy. The biggest single deal in the entire space, and the cleanest illustration of the new procurement model. Talen signed a 17-year, $18 billion PPA with AWS in mid-2025 for up to 1,920 MW from the Susquehanna nuclear plant. Full delivery by 2032. The two companies will jointly evaluate plant uprates and small modular reactors on the same site. This is not a credit purchase or an offset. It is the acquisition of physical, round-the-clock power generation, locked in for almost two decades.

Microsoft + Three Mile Island. Constellation Energy is bringing Unit 1 back online by 2028 — 835 MW dedicated exclusively to Microsoft. The optics of restarting Three Mile Island for AI compute are extraordinary, and largely uncontested.

Google + Duane Arnold. A 25-year PPA with NextEra to restart the Iowa plant by 2029. The first U.S. nuclear restart explicitly attributed to hyperscale AI demand.

Vistra is the other public-market beneficiary, with the largest competitive nuclear fleet in the U.S. and direct hyperscaler relationships across its sites.

Constellation, Vistra, and Talen sit at the center of the trade. All three trade at premiums to historical multiples. All three still appear underpriced against the duration of the contracts they are signing.

SMRs from Oklo, NuScale, and the BWX Technologies supply chain are the 2030+ story. Factory-built, co-located, modular — exactly the deployment model hyperscalers want. Execution risk is real. Timelines will slip. But the demand is locked in. Amazon and Talen are explicit about exploring SMRs at Susquehanna; Microsoft, Google, and Meta have parallel development deals with Oklo, TerraPower, and X-energy.

The fuel layer is its own bottleneck. Cameco and the broader uranium complex face a structural 85 million pound deficit. HALEU, the high-assay low-enriched uranium needed for next-generation reactors, is in genuine shortage, and Russia’s role in the supply chain has become geopolitically untenable. The setup looks like Big Oil in the early 1970s: undercapitalized, structurally short, and about to be rediscovered.

The Stack Around the Stack

Generation is half the problem. Delivery is the other half.

Cooling. Vertiv has become the Nvidia of thermal management. You cannot air-cool a 100 kW rack. Direct-to-chip liquid cooling is no longer an upgrade. It is a hardware requirement. Every new AI campus is a Vertiv campus by default.

Transformers. Every electron that reaches a data center has been stepped down through a chain of transformers — from 765,000 volts on long-distance transmission lines to the 480 volts that feed a server rack. The largest of these — the grid-scale units feeding substations — weigh 400 to 800 tons each. Heavier than a fully loaded Boeing 747. There are roughly a dozen factories on Earth that can build them, and fewer than fifty specialized rail cars in North America capable of moving them. A single unit takes 12 to 18 months to manufacture, then weeks to deliver under police escort.

This is where the lead-time crunch is most acute. Hitachi Energy committed $6 billion to expand production and hire 15,000 people, with its CEO publicly warning that utilities without existing reservations will wait up to four years. GE Vernova builds the large units that feed substations. Powell Industries — historically a Texas oil-and-gas equipment maker — booked its first data center “megaproject” order in late 2025 and has watched its backlog swell to $1.6 billion.

High-voltage distribution. The industry is moving from traditional AC distribution to 800V DC architectures inside the data center, because conversion losses at scale waste roughly 10% of incoming power as heat. That redesign reroutes spend toward Eaton, Hubbell, and the broader switchgear complex.

Construction and transmission. Quanta Services builds the transmission lines, substations, and grid hardware that connect hyperscale campuses to whatever feeds them. They are the industrial labor end of the supercycle, constrained on workforce more than backlog. Electrician wages have roughly doubled in the AI corridors over the past two years.

Secured-power assets. Crypto miners such as IREN, CIFR, and others own something genuinely scarce: pre-negotiated grid interconnections at scale, mostly in cheap-power geographies. Several have begun repurposing those sites as AI and HPC colocation. The asset is not the building or the rigs. It is the grid access agreement that nobody can replicate in under five years.

What 10x and 100x Could Look Like

Most arguments about AI infrastructure are conducted in the abstract. The numbers make them concrete.

Global AI compute deployed today is approximately 28 GW. Sam Altman has stated he wants to add 52 GW per year. Hyperscaler announced capex implies similar trajectories.

The question is what happens if user demand for tokens 10x’s, then 100x’s, from current levels.

Assume continued efficiency gains, call it a 50% reduction in energy per token from architectural improvements, despite reasoning models pushing in the opposite direction. The arithmetic is unforgiving.

A 10x in token demand requires roughly 140 GW of AI compute — five times the current installed base. Running at hyperscale utilization, that consumes about 1,050 TWh per year. That is one quarter of total current U.S. electricity generation dedicated to AI alone, or the entire annual output of Germany.

A 100x requires roughly 1,400 GW — nearly twice the current U.S. peak national load of 745 GW. To put it bluntly: 100x’ing AI demand requires the United States to construct an entirely new electricity system, of greater capacity than the existing one, dedicated to compute.

This is not a forecast. It is a constraint statement. The reason a 100x cannot happen by 2030 is not that the chips cannot ship. It is that the electrons cannot exist.

ASML can scale from 70 EUV tools per year to 100. TSMC can build more N3 capacity. Hitachi can hire its 15,000 transformer workers. None of those scale a grid. The constraint is a fifty-year-old infrastructure system that grows at 1-2% per year colliding with a workload that wants to grow at 70%.

The reason hyperscalers are signing 17-year PPAs and restarting Three Mile Island is that they have run this math. The reason xAI imported a power plant is that it ran the math first.

The Bear Case

The case I am most afraid of being wrong about is the one that took down everyone who got the prior cycle right.

Cisco saw the internet coming. They were correct about the demand. They were correct about the infrastructure required. They were correct that fiber, switches, and routers would be the picks and shovels of the next decade. Their stock lost 89% of its value in the next two years anyway. The demand projection was right. The timing was wrong. Infrastructure built ahead of demand becomes infrastructure depreciated against no revenue.

The same thing could happen here.

AI token costs have fallen 280x in two years. If they fall another 280x, through better architectures, smaller models, specialization, distillation, the demand math collapses. We don’t need a second United States grid. We need 1.3x what we already have. And the transformer makers and SMR startups and gas turbine OEMs find themselves carrying two decades of backlog against demand that did not arrive on schedule. The infrastructure is real. The numbers are real. The trade still goes to zero, because everyone modeled the same thing at the same time.

We have seen this pattern before. The railroad capex cycle of 1865 to 1893 built the physical infrastructure that powered the next century of American growth. It also bankrupted a generation of railroad operators in waves — the Panic of 1873, the depression of 1893.

The people who made money in the long run were not the ones who funded the railroads. They were the ones who owned the steel that the rails were made of, the land the rails ran through, and the banks that lent against the bonds. The picks and shovels.

Every revolution looks digital from the front and physical from the back. Software ate the world for a decade. Now the world is sending the bill, payable in concrete and copper and high-voltage steel.

One Big Carry Trade

A framework I found helpful is to think about the AI buildout as a giant carry trade.

The interesting question is who captures the margin in the conversion.

In 2023-2024 the answer was Nvidia. In 2024-2026 it broadened to HBM and the foundry layer. In 2025 CPUs and optical interconnect names started to compound. In 2026 the answer rotates to the energy stack.

Four people who rarely agree on anything agree on this. Jensen Huang calls it a power-limited industry. Elon Musk calls electrical energy the limiting factor for AI. Dylan Patel calls training runs substation constrained. Brian Janous says the only thing operators worry about now is being able to plug their chips in.

The companies named in this essay sort roughly along the structural argument:

Speed-to-power. GE Vernova for turbines. Bloom Energy for fuel cells. Siemens Energy for the global equivalent. Tesla, Fluence, and Eos for the storage layer that smooths the load.

Nuclear baseload. Constellation, Vistra, and Talen for the existing fleet that hyperscalers are locking up. BWX Technologies and Cameco for the fuel and the supply chain.

Geothermal as the surprise pillar. Ormat as the public incumbent. Fervo when it lists. Sage Geosystems and XGS as the next-generation platforms.

The transformer and switchgear crunch. Hitachi Energy. Powell Industries. Eaton and Hubbell across the broader power-distribution complex.

The thermal layer. Vertiv as the dominant cooling and rack-power name.

Adjacent and indirect. Quanta Services for grid construction. Cloverleaf and Crusoe in the powered-land siting layer. IREN and Cipher for repurposed crypto interconnections. Copper through Freeport.

Optionality further out. Oklo, NuScale, and the SMR cohort for the 2028+ deployment cycle.

The political risk — water, emissions, ratepayer revolt — is real, but it does not change the math. It lengthens queues, which raises the value of permitted assets.

The most aggressive operators are already running the optionality. SpaceX and xAI filed for one million orbital data center satellites in early 2026, the explicit pitch is putting compute where electricity is essentially free and cooling is solved by physics.

That is a 2030+ story at best, and roughly 3x more expensive per watt than terrestrial today. But it is on the roadmap. When the constraint is hard enough, the responses get strange.

In 2019 you could order a high-voltage transformer and have it within a year. Today the queue runs 80-130 weeks. The orders keep coming.

Because the shortage now is not about energy. It was about deliverable energy, in the right place, at the right time.

The frontier model will not be the binding constraint anymore.

The transformer is.

This is the fourth essay in a series on the AI compute stack, after The Other AI Trade on CPUs as the orchestration layer, Half of Nvidia on HBM, and Six Billion Dollars of Light on optical interconnects. Together the four pieces map the supply chain that turns capital into compute.

Disclaimer

This essay is for informational and educational purposes only. It does not constitute investment, legal, tax, or financial advice and should not be relied upon as such. Companies and securities discussed are referenced for analytical context; nothing in this essay is a recommendation to buy, sell, or hold any security. The author may hold long or short positions in some of the companies mentioned. Markets are uncertain, forecasts are imperfect, and past performance does not predict future results. Readers should conduct their own research and consult qualified professionals before making any investment decision.

Companies Mentioned

Power generation: GE Vernova (GEV), Siemens Energy (ENR.DE / SMNEY), Bloom Energy (BE), Constellation Energy (CEG), Vistra Corp (VST), Talen Energy (TLN), NextEra Energy (NEE), Ormat Technologies (ORA), Oklo (OKLO), NuScale Power (SMR), BWX Technologies (BWXT), Doosan Enerbility (KS: 034020).

Equipment & infrastructure: Vertiv (VRT), Eaton (ETN), Hubbell (HUBB), Powell Industries (POWL), Quanta Services (PWR), Hitachi Energy (parent: Hitachi, 6501.T).

Storage & power smoothing: Tesla (TSLA), Fluence Energy (FLNC), Eos Energy (EOSE).

Fuel & materials: Cameco (CCJ), Freeport-McMoRan (FCX), Liberty Energy (LBRT).

Secured-power assets: IREN (IREN), Cipher Mining (CIFR).

Hyperscalers and AI labs (referenced as buyers, not investment recommendations): NVIDIA (NVDA), Microsoft (MSFT), Amazon (AMZN), Alphabet (GOOGL), Meta Platforms (META), Oracle (ORCL), AMD (AMD).

Private companies referenced: xAI, OpenAI, Anthropic, Fervo Energy (S-1 filed), Sage Geosystems, XGS Energy, TerraPower, X-energy, Crusoe, Cloverleaf Infrastructure, Lancium, Starcloud, SpaceX.

Further Reading

• SemiAnalysis — Dylan Patel and team’s industry research at newsletter.semianalysis.com. The canonical source for AI infrastructure and data center power. Their Datacenter Industry Model and Energy Model are the data layer behind much of this essay.

• Dylan Patel on the Dwarkesh Podcast (March 2026) — three hours on the logic, memory, and power bottlenecks. The single best long-form interview on the binding constraints of the buildout.

• Rhodium Group — “Beyond the Megawatts: Geothermal Energy’s Potential to Power Data Centers” (2025). The 64% number that anchors the geothermal section.

• IEA — “Energy and AI” Special Report (2025). The official institutional projection.

• PJM Interconnection — capacity auction results and market monitor reports at pjm.com. The price data behind The Bill Arrives section.

• Cloverleaf Infrastructure — Brian Janous’s various podcast appearances, particularly the Catalyst and Build, Repeat episodes. The operator’s view on grid integration.

- Simone Weil

Daniel Kahneman spent the better part of a decade writing Thinking, Fast and Slow. The average person who buys it reads to about page 26.

A job gets accepted, worked for fourteen months, and quit. Not for a better one. Just for a different one.

A relationship reaches week six and ends, because something stopped being effortless and there is a queue of alternatives one swipe away.

A show gets watched for half an episode and closed. Nothing happened fast enough.

These are the same act.

Georgia O’Keeffe said this about flowers:

“Nobody sees a flower, really, it is so small. We haven’t time — and to see takes time, like to have a friend takes time.”

That sentence was written before the iPhone. Before Hinge. Before any of the architecture we now live inside was built.

It was already true. Now the architecture is designed to make sure it stays unseen.

In 2012, Sean Rad and Jonathan Badeen built Tinder around a single design choice. The swipe. They modeled it on the slot machine, which works because of something B.F. Skinner figured out with pigeons in the 1950s, that intermittent, variable rewards are more compulsive than predictable ones. You don’t know if the next pull will pay. So you keep pulling.

Tinder applied that mechanism to human relationships. It worked. The product was processing over a billion swipes a day within a few years of launch. Every other dating app copied it. Then the rest of the consumer internet copied it. The pull-to-refresh feed, the autoplay video queue, the algorithmic for-you page, all of them are slot machines built for different categories of decision. Each is engineered to make sure the next thing arrives before you finish sitting with the last.

The same design choice is now everywhere. The platform does not want you to commit, to a person, a show, an article, a thesis, a job, a self. It wants you to swipe again. Every swipe is a session. Every session is what gets sold.

The aggregate effect is that the time horizon for almost everything in modern life has compressed. The average employee tenure at a US company is around four years and falling. The average holding period for an NYSE-listed stock has gone from eight years in 1960 to roughly five months. The average sustained attention span at a screen, before someone switches tasks, is around forty seconds. Every number used to be larger. None is going back.

What gets lost in the compression is what O’Keeffe was describing. The flower you actually see. The friend who takes time to become a friend. The thesis that takes five years to be right. The book that pays off at page 200. The job that compounds at year ten. The marriage that becomes itself in the second decade.

These are the things patience was for. They are the things no one gets anymore.

In markets, the loss of patience is measurable. That makes them the cleanest case study.

I have friends who run large long-only books with twenty years of good numbers. I have asked them how long they can underperform before money leaves. The answer is six months. After a year, they are toast. The leash is two quarters, regardless of the track record behind it.

The investor who is supposedly the patient capital, the long-only manager whose job description is to take a multi-year view, has a two-quarter window before his clients fire him. He still does the work. He still has a view. He cannot act on it longer than the redemption window allows. So most of the time he doesn’t. He buys what is going up and sells what is going down, and tells himself he is being disciplined about risk.

Reg NMS in 2005 fragmented US equities and birthed the high-frequency trader, who taxes every trade and produces no view. The Pension Protection Act in 2006 defaulted American 401(k) savers into target-date funds, which buy indices without reference to price. One change bled the discretionary investor of profit. The other starved him of capital. The investor whose job was to take a five-year view did not lose an argument to the market. He was selected out of the population over twenty years.

What replaced him is structurally more impatient. The marginal buyer in a $5bn German photonics name today is a thousand US retail accounts on Robinhood, following a Substack writer who used a language model to draft his thesis last week. Sean Maher at Entext tracks this. AI-generated tech-stack diagrams have made obscure foreign component suppliers, German MOCVD specialists, Swedish photonics names, Taiwanese epitaxy mid-caps, legible to US retail in a way they have never been. US retail orders now flow into Frankfurt and Tokyo for shares the local brokers say the buyers cannot place on a map. LandMark Optoelectronics, a Taiwanese name, is up 324% year-to-date on $70m of revenue at an $8bn cap. None of these holders has a five-year horizon. None has a one-year horizon. The new marginal buyer is, by design, more impatient than the one he replaced.

Markets have never been more reflexive. Five years of growth get priced into a quarter. A single quarter of slowdown gets extrapolated into a five-year breakdown. The price moves, the price moves the narrative, the narrative moves more price, and the only people who could stand against the loop have a two-quarter leash.

This matters beyond markets because the country has built a fiscal regime on top of equity prices. Paul Tudor Jones on a recent Invest Like The Best episode, puts US stock market capitalization at 252% of GDP, against 65% in 1929 and 170% in 2000. A normal mean reversion to the 25-year average price-to-earnings ratio implies a 30 to 35 percent decline — equivalent to wiping out 80 to 90 percent of GDP in household wealth, collapsing capital gains receipts, worsening the deficit, and dragging the bond market with it. The deficit, the bond market, and the consumer have come to depend on equity prices staying high. They came to depend on those prices over the same twenty years in which the patient capital that supported them was systematically removed. The reverse wealth effect, when it arrives, will not be a market story. It will be a fiscal one.

Markets are not behaving badly. They are behaving exactly as a system without patient capital must behave. The fundamentals still exist. They just operate on a timescale that nobody is given.

Harvard art historian Jennifer Roberts makes her students sit in front of a single painting for three hours before they are allowed to write about it. The first thirty minutes are uncomfortable. The next thirty are restless. The hour after that is when something happens. Details show up that were invisible in the first hour. Relationships between parts of the painting that nobody had ever seen before become visible. The students leave with the experience of having actually seen something, often for the first time in their academic lives.

Roberts wrote a piece about this called “The Power of Patience.” Patience used to mean disempowerment, you were patient because you had no other choice, because the world made you wait. In a world where you no longer have to wait for anything, patience becomes the opposite. It becomes the deliberate refusal to be governed by the tempo the system imposes. “Patience no longer connotes disempowerment, perhaps now patience is power.”

Roberts' essay was brought to my attention by 13D Research's What I Learned This Week.

The system is not going to slow down. The platforms are not going to stop optimising for swipes. The redemption windows are not going to extend. The market is not going to ask anyone to hold positions longer. The labor market is not going to reward staying. The dating market is not going to reward depth. None of it is going to revert.

One choice remains: to operate on a different clock. To stay in the book past page 26. To stay at the job past year two. To stay in the relationship past week six. To hold the position past the quarter where it underperforms. To sit in front of the painting for three hours.

The contrarian act in an impatient world is staying. Staying long enough to see what was actually there. Staying long enough for the friend to become a friend, the thesis to become right, the work to become deep, the self to become real.

The curious mind sees what the regime has done to time. And then chooses to spend its own time differently.

The flower is still there. Seeing it still takes time.

Every week the headlines told me Nvidia had made another investment, another acquisition, another partnership with a c…

Read more

This is a briefing memo for the Curious Mind community. It reflects my personal views on a fast moving topic and is not a research report.

The most useful question I ask portfolio companies these days is no longer “how many GPUs can you get?” It is “how much HBM does your supplier have allocated to you for 2026?”

Most operators do not yet know the answer. The ones who do are the ones who will actually ship product.

For two years the bottleneck in the AI trade was clear and physical. TSMC could not produce enough advanced packages. The packages held the GPUs. The GPUs held the trade. Every quarter, the same phrase: we are supply-constrained. Then, sometime late last year, that constraint cleared. CoWoS (Chip-on-Wafer-on-Substrate) capacity tripled. Nvidia’s lead times shortened. The trade should have cooled.

It didn’t.

Because while the market was watching the foundry, the chokepoint had walked next door, to a small handful of memory companies most generalist investors still cannot spell. SK Hynix. Samsung. Micron. Three names that until recently were lumped together as a commodity cycle play, swinging between gluts and shortages every eighteen months. They are no longer commodity. The most expensive, most contested, most undersupplied component in the AI buildout is not the chip that does the math. It is the memory stacked next to it.

If you bought one of Nvidia’s latest accelerators today, roughly half of what you paid for would not be silicon doing logic. It would be memory.

That number is the whole essay and explains why 50% of the S&P EPS revisions are coming from one company.

The Memory Wall

To understand why memory became the bottleneck, you have to understand a problem chip designers have been losing for two decades.

Processors have gotten faster every year. Memory has gotten faster too — but much slower than processors. The gap between what a chip can compute and what its memory can deliver has been widening since the early 2000s. Engineers call it the memory wall. By the late 2010s a top-end GPU could perform a calculation in a hundredth of the time it took to fetch the data for that calculation. The chip spent most of its life waiting.

For a while this didn’t matter. Most workloads, graphics, video, scientific simulation, could be tuned to keep data close at hand. You’d load a frame, do the work, write it back. The wall was real but you could plan around it.

Then came the transformer.

The defining feature of a modern AI model is that almost every operation requires a different slice of the model’s weights, sometimes hundreds of billions of parameters, to be available, right now, in milliseconds. There is no neat reuse pattern. There is no clever way to keep just the relevant bits close. You need everything, everywhere, all at once.

This is what people mean when they say AI is bandwidth-bound. It is not about how many calculations the chip can do per second. It is about how fast you can shovel the model’s weights through the chip’s mouth. A GPU that can do twenty thousand trillion operations a second is useless if its memory can only feed it the equivalent of three thousand. The other seventeen sit idle.

So the engineers did the only thing left.

They moved the memory inside the chip.

The Sandwich

A traditional computer puts memory on the motherboard, several inches from the processor, connected by skinny copper traces. The connection is fast in human terms and glacial in chip terms.

High Bandwidth Memory does three things differently.

First, it stacks the memory chips on top of each other, twelve, sometimes sixteen tall, and runs vertical wires straight through them, called through-silicon vias. Each finished stack contains as much memory as a small server.

Second, it sits the stack right next to the processor on a piece of silicon called an interposer. Not on the motherboard. Not even on the same circuit board. Glued onto the same package, millimeters away.

Third, it talks to the processor through thousands of parallel wires instead of a few dozen. The connection between a modern GPU and its memory is, quite literally, two thousand and forty-eight lanes wide. Imagine a highway that wide. That is the only way to keep the chip fed.

Each new HBM generation widens the highway and stacks more floors. HBM3 was the workhorse of 2024. HBM3E is most of what Nvidia is shipping now. HBM4, ramping through 2026, doubles the bandwidth again and brings something new — a bottom layer that is a small custom chip, made by a foundry like TSMC or Samsung, allowing the memory and the processor to be co-designed.

HBM is also only half the bottleneck. The other half is the interposer that the HBM and the GPU sit on, TSMC’s CoWoS process. You cannot ship one without the other, and the slower one in any given quarter sets the ship rate for Nvidia and AMD. When you read about Nvidia’s “supply” improving, what is usually clearing is one of these two, packaging or memory, and the other becomes the new gate. Watch them as a pair.

The line between memory and logic, between TSMC and Hynix, is dissolving. The “GPU” of 2027 will be honestly described as a memory-and-logic sandwich, where the memory is the heavier slice. The CPU — the boring brain I wrote about a few days ago, still sits on the motherboard with cheap commodity DDR5 next to it. That is fine. The CPU runs orchestration code that does not need this kind of feeding. But the workload that pays for the data center, model training and model inference, runs entirely on the memory-and-logic sandwich. And that sandwich is sold by exactly three companies on earth.

Jensen Huang has made the obvious counter publicly. The cost of materials inside an accelerator is a poor proxy for value capture. The architect of the system, the company that chose the memory, designed the interconnect, wrote the software, keeps the largest share. He is right. The memory maker still earns a strong margin on a strategic input. Both can be true. The question for the next two years is which margin compresses first.

The Ecosystem

Until 2023, the global memory industry was a cyclical mess. Three players, SK Hynix, Samsung, and Micron, would race to add capacity, oversupply the market, watch prices collapse, cut capex, watch prices recover, and start the whole cycle again. HBM has rewired the cycle.

SK Hynix got there first and makes the bulk of the world’s HBM3E, including the memory inside Nvidia’s most-shipped accelerator. They were treated as a commodity stock for fifteen years. They are now the most important non-American technology company outside Taiwan. HBM margins are roughly three to five times what they earn on commodity DRAM. Around half their profits now come from a product that did not exist at scale four years ago.

The lead is worth understanding precisely. It is not a process lead at the cell level — all three players make competitive DRAM cells. It is a packaging and bonding lead. Stacking sixteen dies on top of each other with submicron alignment, and getting the thermal behavior right so the stack survives qualification, is hard. Yield is non-linear in stack height — going from 8-high to 12-high is much harder than 50% harder, and 16-high is where Samsung and Micron have stumbled. This is why bonding equipment from BESI and Disco matters so much for who wins the next two cycles.

Micron is the comeback story. They were late to HBM, lost share, and were left for dead in early 2024. Through 2025 and into this year they qualified into Nvidia’s supply chain at HBM3E, then announced HBM4 samples ahead of expectation. The stock has reflected the change. The question now is whether they can hold quality at HBM4, where the engineering challenge — particularly the bonding of stacked dies — is harder.

Samsung is the unresolved variable. They are the largest memory company in the world by volume and the slowest on HBM. They missed most of the HBM3E cycle. They have spent two years trying to qualify with Nvidia and have not. Their HBM4 pitch is ambitious, using their own foundry to make the bottom logic die, offering a fully integrated stack. If it works, the competitive map resets. If it does not, they spend another two years as a discount supplier.

Behind these three sit the picks and shovels. TSMC makes the interposer that the HBM stack and the GPU sit on (the CoWoS process), and packages the whole sandwich. ASML sells the lithography machines that make modern DRAM possible at all. Applied Materials and Lam Research sell the deposition and etch tools used in the stacking process.

The bonding and inspection layer is the most underappreciated piece of the equipment story. Disco and BE Semiconductor sell the equipment used to bond dies together with the right precision. Onto Innovation sells the metrology and inspection tools that verify the bonds held, every additional die in a stack means more inspection steps, and the equipment line item per HBM stack is growing faster than for any other DRAM product. Moving from HBM3 through HBM4 and beyond requires a transition from thermo-compression bonding to hybrid bonding, and the equipment makers who win that transition will earn through the entire cycle.

There is one adjacent trade worth a brief mention. As AI inference scales, the bottleneck spills beyond model weights into storage. Vector databases, retrieval-augmented generation, and KV-cache offload all require fast NAND-based storage tiered alongside HBM and DDR. Enterprise SSD demand has lifted as a result, and the NAND cycle is now running on its own version of the HBM dynamic — supply discipline, AI-driven demand pull, structural tightening. SanDisk is the cleanest large-cap exposure. It is a related trade rather than the same trade, but it sits inside the same physics: hot data in HBM, warm data in DDR, cold-but-still-fast data in enterprise NAND, and the price of all three rising at once.

This is not a list of recommendations. It is a map of who actually shows up at the loading dock when the AI trade buys HBM.

The New Cycle

You have probably seen the word memflation in some research note. It means what you think it means. The price of memory is rising, and rising in ways that are not behaving like a normal cycle.

Three things are different this time.

The first is mix. Each HBM stack consumes between two and three times as much wafer capacity as the equivalent number of bits of standard DRAM. That means even if memory companies wanted to satisfy both AI demand and traditional PC and server demand, they physically cannot. Every wafer they convert to HBM is a wafer of standard DRAM not made. The result is that the AI buildout is inflating the price of the memory in your laptop, your phone, your car. The strength in commodity DRAM pricing is miscast as a normal demand recovery — consumer and enterprise demand has been mediocre. What is happening is a manufactured shortage: the memory makers are starving the commodity book to prioritize HBM allocation. That joins two markets that used to move independently. Your laptop memory price is now a derivative of GPU shipments, and the commodity DRAM cycle no longer ends when supply catches up — it ends only when HBM demand cools.

The second is contracts. Memory has historically been a spot market. AI buyers, Nvidia, the hyperscalers, the sovereign labs, are now signing multi-year contracts at fixed-but-rising prices. SK Hynix has reported that all of their 2026 HBM capacity is sold. Samsung the same. Micron the same. The cycle has not been abolished, but the visibility into 2026 and most of 2027 is unlike anything memory has seen.

The third is capex discipline. The shortages of 2017 and 2018 ended with a rush of new fabs that crashed the market by 2019. This time the three players have been restrained. SK Hynix is adding capacity but at a measured pace. Samsung’s HBM4 ramp is constrained by yield, not by capex. Micron has guided that supply will lag demand into 2027.

How to Read the Headlines

A few practical filters for the news that will land in your inbox over the next year.

When you see “DRAM prices fell 4% in October” — this is almost always commodity DRAM, not HBM. HBM is contracted. The two markets are now mostly separate. The headline is real but it does not tell you what you think it does.

When you see “Samsung qualifies HBM3E with Nvidia” — read the fine print. Qualification can mean small volumes for a single SKU. The number that matters is what percentage of Nvidia’s shipped HBM is coming from each supplier. That number moves slowly.

When you see “Memory company announces $XX billion capex” — distinguish between brownfield expansion (existing fabs, fast to ramp) and greenfield (new fabs, two to three years out). Greenfield announcements rarely affect near-term supply but they reset the long-term cycle.

When you see “HBM4 yields disappointing at supplier X” — pay attention. The hardest engineering problem in HBM4 is bonding sixteen dies together with submicron precision. The first company to solve it at high yield wins the cycle. The companies that do not will lose share for two years.

What Would Break This

The HBM thesis rests on four things: that scaling laws keep working at the frontier, that inference stays bandwidth-bound, that the three players stay disciplined, and that no architectural shift reduces memory demand per unit of compute.

The scaling laws risk is the deepest because it is binary. Every assumption in this essay rests on the next generation of frontier models continuing to show capability gains as compute is scaled. If GPT-6, Claude 5, and Gemini 3 work, the demand picture for 2027 to 2029 is not what this essay implies. It is dramatically larger — sovereign labs, defense compute, and the corporate inference market all expand at once. If they do not work and the scaling era plateaus, demand collapses and the supply tightness this essay describes becomes a glut within eighteen months. The middle scenario — steady capex, manageable tightness — is the least likely outcome, not the most. The most important uncertainty in this trade is not cyclical. It is whether the largest scientific experiment in history continues to deliver.

The architecture risk is the next deepest. In-memory compute, optical interconnect, and 3D-stacked DRAM bonded directly onto the GPU die are all credible enough to take seriously. Optical in particular is already moving — Lumentum, Coherent, Astera Labs, and Credo are riding the same bandwidth-bound logic that lifted HBM, but at the rack-to-rack level rather than the chip-to-chip level. None of these is imminent as a full HBM substitute. All would, in different ways, dissolve the premium that today accrues to a stack of memory sitting next to a processor. If the workload itself changes, sparser models, mixture-of-experts, more efficient inference, bandwidth-per-FLOP requirements drop and the same HBM supply serves a much larger market. Good for AI economics. Bad for HBM pricing power.

The discipline risk is the most familiar. Three players have cooperated before, in 2017 and again in 2018, and cooperation has always broken. Samsung is the obvious crack point because they have the most share to recover and the largest underutilized capacity to throw at it. The moment they qualify HBM4 at scale and decide to underprice, the oligopoly margin structure resets.

The cycle timing risk is the simplest. Memory stocks have historically traded at terrible multiples in the back half of supply cycles even when earnings are still high, because the market prices the next downturn six to twelve months ahead. The fabs being announced today do not produce HBM tomorrow, but they produce it in 2027 and 2028. Memory cycles always overshoot. You can be right about HBM demand through 2027 and lose money in the stocks starting in mid-2026.

I am not making any of these calls today. The next twelve months look visible. But the smartest readers will close the tab thinking what’s missing is the other side. The other side is that this is a multi-condition bet dressed as a structural certainty, and the conditions deserve to be named.

The Strategic Layer

There is a layer above all of this that the cycle analysis does not capture. Two of the three HBM makers are South Korean. The third is American but small.

The most consequential industrial input to the American AGI program is manufactured almost entirely within range of Chinese missiles, in a country whose alignment with US technology controls is good but not permanent.

The CHIPS Act funding for Micron’s Idaho and New York expansion is the visible American response. The less visible response is the quiet conversation in Washington about whether HBM should be treated as strategic infrastructure, the way the US treats rare earths, pharmaceutical precursors, and certain refined metals, rather than as a commercial commodity. The answer to that question matters more for the 2028 trade than any of the cycle dynamics described above.

Watch the policy as carefully as you watch the players.

Coda

I asked a friend who runs supply chain for a large hyperscaler what kept him up at night. He said: “The chips. Always the chips.”

I asked which chips.

He paused.

“The memory. We can buy GPUs. We cannot buy memory.”

That conversation was six months ago. The HBM stocks have roughly doubled since. The structural story is not in the multiple. It is in the sentence my friend said. The single largest constraint on the most-watched technology buildout in history is being met by three companies. Two are foreign. One is American. None used to be considered strategic.

They are now.

And the friend’s sentence — we cannot buy memory — has started to mean something other than what it meant six months ago. It used to describe a procurement problem. It is becoming a national security problem.

If the GPU was the headline of the AI trade, the memory is the punchline. The punchline is doing more of the work than anyone wanted to admit. And it is being delivered by people most of the audience has never heard of, on a stage they did not know existed.

Watch the stack.

Disclaimer

This post is for informational and educational purposes only. It does not constitute investment advice, a recommendation, or an offer to buy or sell any security. I, members of my family, or entities I am associated with may hold positions in the companies or sectors discussed, and those positions may change at any time without notice. Past performance is not indicative of future results. You should consult your own financial, tax, and legal advisors before making any investment decision.

Appendix

Companies Mentioned (Map, Not Recommendations)

Memory makers

• SK Hynix (KRX: 000660) — Korean, current HBM share leader

• Samsung Electronics (KRX: 005930) — Korean, attempting to retake share at HBM4

• Micron Technology (NASDAQ: MU) — American, the comeback story

Picks and shovels

• TSMC (NYSE: TSM) — interposer and advanced packaging (CoWoS)

• ASML (NASDAQ: ASML) — lithography for advanced DRAM

• Applied Materials (NASDAQ: AMAT) — deposition and etch

• Lam Research (NASDAQ: LRCX) — etch and stacking equipment

• Disco Corporation (TYO: 6146) — wafer dicing and grinding

• BE Semiconductor (AMS: BESI) — hybrid bonding equipment, key to HBM4 transition

• Onto Innovation (NYSE: ONTO) — metrology and inspection for advanced packaging and HBM stacks

Optical and connectivity (the next layer out)

• Lumentum (NASDAQ: LITE) — optical components for AI fabrics

• Coherent (NYSE: COHR) — optical components and lasers

• Fabrinet (NYSE: FN) — contract manufacturing for optical

• Astera Labs (NASDAQ: ALAB) — connectivity controllers for AI clusters

• Credo Technology (NASDAQ: CRDO) — high-speed connectivity ICs

Adjacent — storage

• SanDisk (NASDAQ: SNDK) — enterprise NAND exposure to AI inference workloads

Buyers

• Nvidia (NASDAQ: NVDA), AMD (NASDAQ: AMD) — the largest single customers

• Hyperscalers (Amazon, Microsoft, Google, Meta) — direct buyers via custom silicon programs

Key Terms

• HBM (High Bandwidth Memory) — vertically stacked DRAM placed adjacent to a processor on the same package

• TSV (Through-Silicon Via) — vertical wires connecting stacked dies

• Interposer — thin silicon substrate that connects HBM stacks to a GPU

• CoWoS — TSMC’s advanced packaging process used for almost all current HBM-equipped GPUs

• Memory bandwidth — how much data per second can be delivered between memory and processor

• The memory wall — engineer’s term for the widening gap between processor speed and memory speed

• DDR5 — current generation of standard server and PC memory (the kind in your laptop)

• Hybrid bonding — die-to-die bonding technique required for HBM4 and beyond

• CPO (Co-Packaged Optics) — moving data between chips as light rather than electrical signals, packaged onto the same substrate as the processor

• NAND — non-volatile flash memory used for storage rather than working memory; the technology inside enterprise SSDs

Further Reading

These are the sources I returned to most while writing this. Where I have linked to a publication rather than a single piece, it is because the writer’s whole body of work matters more than any one essay.

The big-picture frame

• Leopold Aschenbrenner, Situational Awareness: The Decade Ahead — situational-awareness.ai. The canonical argument for why the AI compute buildout is the most consequential industrial project of the century. Read the “From $1 Trillion Compute Cluster to Superintelligence” and “Racing to the Trillion-Dollar Cluster” sections in particular for the demand-side picture that sits underneath this essay.

• Chris Miller, Chip War (2022) — the definitive book on the geopolitics of semiconductors. The chapters on Korea and Taiwan are essential context for the Strategic Layer argument here.

• Ed Conway, Material World (2023) — broader frame on the physical inputs to the modern economy. The silicon chapter is a useful primer for non-specialists.

Ongoing primary research

• SemiAnalysis (Dylan Patel and team) — semianalysis.com. The most important ongoing research on AI hardware, HBM supply chains, CoWoS bottleneck dynamics, and hyperscaler buildouts. Their pieces on the H100/H200/B200 BOM, the Samsung HBM qualification saga, and the CoWoS-S to CoWoS-L transition are each worth the subscription.

• Stratechery (Ben Thompson) — stratechery.com. The strategic frame on Nvidia, TSMC, and the platform dynamics around AI compute.

• Fabricated Knowledge (Doug O’Laughlin) — fabricatedknowledge.com. Long-form semiconductor analysis on Substack.

• Digits to Dollars (Jay Goldberg) — digitstodollars.com. Industry-insider commentary, particularly strong on memory and packaging.

• TrendForce — trendforce.com. The standard industry source for HBM market share and DRAM pricing data.

• TechInsights — techinsights.com. The best teardown analyses of new accelerators.

Korean and Taiwanese press

• ETNews (etnews.com) and KED Global (kedglobal.com) — earliest signals on Hynix and Samsung qualification timing.

• Chosun Biz (biz.chosun.com) — Korean industrial policy and supplier dynamics.

• DigiTimes (digitimes.com) — Taiwanese supply chain reporting, particularly on TSMC packaging capacity.

Primary source documents

• SK Hynix, Micron, Samsung quarterly earnings calls — the SK Hynix call has become one of the most useful reads in semiconductors. Transcripts available at investor relations sites for each company.

• TSMC quarterly earnings calls and CapEx commentary — the binding constraint on advanced packaging capacity.

• Nvidia GTC keynotes (Jensen Huang) — for the customer side of the HBM demand picture.

Policy and national security

• CSIS (Center for Strategic and International Studies) — csis.org. Their semiconductor supply chain reports are the best public-domain analysis of the strategic concentration question.

• US Commerce Department CHIPS Act announcements — for the specific funding amounts and timelines for Micron’s domestic HBM ramp.

• US Bureau of Industry and Security (BIS) export control filings — the policy lever most likely to reshape the HBM supply picture over the next two years.