ESSAY
Materialization —
Making Things in the Age of AI
Yushiro Kato / September 2026
The Sunday With No Color
It was a Sunday. I spent the morning calling paint shops.
The day before, a sheet-metal part we had delivered to a customer's plant came back defective, and I had been thoroughly chewed out. The problem was the paint. But no paint shop is open on a Sunday. I called dozens. The one shop that finally picked up didn't have the color. I asked them to mix it on the spot, and we barely made it.
Apologize. Find the cause. Drive the replacement part over yourself. That was the entire set of things I could do that day.
I was twenty-six, the year I started the company.
Not long before, I had been a consultant advising large manufacturers. Cost reduction, design overhauls, subsidiary consolidation. I built decks and said clever-sounding things to executives. I thought I understood manufacturing. That confidence did not survive my first few days on an actual shop floor.
In my own company, I read drawings and visited job shops one at a time. One of my cofounders spent months working unpaid in a small factory to learn the work from the inside. I have visited well over a thousand plants myself. I ground on quotes and fought quality, cost, and delivery late into the night. I stared at drawings to find a single cent of cost. I bowed in front of defect reports. I apologized on the phone for late deliveries.
Once, a simple bracket arrived covered in burrs. On the drawing it was nothing: a piece of sheet metal bent into an L. I filed the burrs off one at a time until my hands stopped working. Just deburring. But by the time I had lost the feeling in my fingertips, I understood a little of what it means to make something.
Five years later, in 2022, we had somehow become a company that could take on the hardest things anyone makes — core components for semiconductor production equipment, water treatment complexes larger than a baseball stadium, an entire machine tool of the kind that sits at the heart of a factory. Several hundred thousand distinct products a year, from one-offs to medium volume.
Making things is not glamorous intellectual work. It is an accumulation of physics and dirt and reality. The 3D CAD can be correct and the steel still won't bend that way. Skip the surface prep and the paint comes off. Weld aluminum and the distortion is brutal. I learned all of it with my hands, not my head.
Which is exactly why I welcome the fact that AI is making intellectual work faster. AI writes mathematics. It predicted two hundred million protein structures. The world's intelligence has been extended. That is real, and I have no interest in talking it down.
But I keep coming back to the same question.
After all this intelligence, how much has the physical world we live in actually changed?
Honestly: almost not at all. The cars on the road, the commute, the equipment in the hospital, the factory down the street — all of it looks like it did ten years ago.
The reason is simple. AI getting smarter and the physical world changing are not the same event. To change reality, something physical has to be made. This is obvious.
And it is the thing most easily forgotten in the current boom.
I know in my body how little of "making a physical thing" AI moves on its own.
The World AI Can't Solve
A plane falls out of the sky.
There are a few hundred people on it. They are holding smartphones. A movie is playing in the cabin. On the ground, controllers are tracking every flight on earth in real time. Our intelligence has come that far. And then fatigue in a single piece of metal passes one threshold, the aircraft obeys physics, and people die easily.
Challenger, prepared over more than a decade, came apart in 1986 because of one small part: an O-ring.
Knowledge now looks capable of solving everything. So why is it that when a plane goes down, or two cars hit each other, people still die so easily?
Precisely because AI is passing the limits of intelligence, we are going to run into a more primitive limit. Physical limits.
- 1.19 million people die in road accidents every year.
- 7 million die early from air pollution.
- Half the world, more than 4 billion people, has no way to travel freely by car.
- The bridges, water systems, and power grids built in the twentieth century are reaching end of life across the developed world at the same time.
- More than 90% of leading-edge semiconductors come from one island. Manufacturing everywhere is rate-limited by geopolitics.
None of this is solved by the number of papers AI writes. Not by inference speed, not by lines of code. It doesn't go away until matter, energy, space — physical reality itself — changes.
And it isn't only damage.
On the other side of the wall there is release. Longer lives and extended bodies from advances in drugs and medical devices. Energy and compute cheap enough to be in anyone's hands. Tokyo to New York in forty minutes. Access to space as a physical place. Eyes on the whole earth in real time, so that no disaster is ever missed again.
Every one of them gets closer if we can make things, and stays far away if we can't.
I call these physical limits. Split into the damage side and the release side, they come to fourteen clusters. They are limits that physically exist at this moment. Papers alone don't solve them. Code alone doesn't either. They get solved only after something is designed, prototyped, tested, produced at volume, shipped, installed, operated, and maintained. Priced out, my rough estimate is about $126 trillion a year — roughly 1.15 times world GDP. It is an order-of-magnitude claim, so look at the exponent rather than the digits.
This is not a new problem created by AI. As long as humans live in a physical world, it has always been there. If anything, the further AI goes past human intelligence, the more this primitive limit moves to the front as the bottleneck.
The smarter intelligence gets, the wider the gap grows between it and the progress of matter.
It is fair to say humanity's bottleneck is shifting to physical limits.
That is where the question starts. Who crosses this wall, and how?
Thirty Years of Asymmetry
Two plain facts.
In 1995, taking a car from concept to volume production took about four years. Thirty years later, it still takes about four years. There has been enormous improvement and effort. The fundamental lead time has not moved.
Here is the other fact. Thirty years ago, standing up an e-commerce site — database, server, payments, UI, testing — took years. Today, with off-the-shelf SDKs, the cloud, and AI, a working version is up in under an hour.
That is the asymmetry between Bit and Atom. Bits copy for nothing, old code can be reused, and AI can generate them in bulk. Atoms do not exist unless they are physically made. Steel takes the time the press takes. Paint takes the time it takes to dry. An engine takes the time its tests take.
Over thirty years we made the world of Bits more than a hundred times faster. The world of Atoms moved at roughly the same speed in most industries. In the AI era, the gap widens further. Even if AI makes thinking ten thousand times faster and produces ten thousand times the theory, if a car still takes four years, that upside gets diluted in the physical world. Solving physical limits means solving this asymmetry.
Why are Atoms so slow? We have named three properties that make them different. The simplest one is this: the physical world has no Ctrl+Z.
Software can be broken. Ship a bug, roll it back. Press Ctrl+Z and you are one move earlier. So you can move fast, break things, and fix them fast. The physical world has no Ctrl+Z. Drill the hole in the wrong place and that metal is gone. Assemble a product wrong, and if it fails, someone dies. So the people who make matter carry weight on every single move and have no choice but to be careful. Check, test, record, check again. That inability to undo — what we call fixity — is one of the large structural reasons Atoms are slow.
What Astro Boy Got Right
I keep thinking about a comic I read as a child.
In 1952, Osamu Tezuka began drawing Mighty Atom, known in the United States as Astro Boy. Seventy years ago, when Japan had not yet climbed out of postwar poverty. Neither Tezuka nor his readers treated it as a forecast. It was a good boys' comic. Nobody thought it was prophecy.
Seventy years on, the distribution of what came true is strange.
The Bit predictions almost all landed. An artificial intelligence you can talk to. Instant translation of a foreign language. Meeting face to face at a distance. A network spanning the world. Things that would have looked like magic then are in our palms. Even twenty years ago, a real translation machine was a fantasy. Now it is ordinary.
The Atom predictions almost all did not. Flying cars are barely in testing. Fusion, an endless clean source of energy, has been proven in principle and is still in research. Humanoids that move like people in any environment have just reached the doorway of production. Cities on the moon and on Mars are still "the future," the same as in 1952.
To someone in 1952, these would all have looked equally impossible. A talking robot and a city on the moon were equally dreams. Seventy years later, the information dreams came true one after another, and the material dreams were left roughly where they were. That asymmetry is the deepest feature of the world we live in.
But the material dreams are not unrealized because they are physically impossible. In most cases the principle has been understood for a very long time. What is missing is always the time it takes to finish the thing and get it into society.
Take the airplane. Everyone knows the Wright brothers: 1903, first powered flight. Fewer know when the principle arrived. George Cayley separated lift, drag, thrust, and weight, drew the shape of the modern aircraft — fixed wing, fuselage, tail — and published On Aerial Navigation in 1809. From the principle to a single machine leaving the ground took roughly a hundred years. From there to safe, cheap air travel anyone could buy, several more decades. Not once was the bottleneck "we don't understand how flight works." It was how to control it safely, how to build a light enough engine, how to produce it. All of the time sat on the side of turning it into matter.
There is no shortage of examples. The CT scanner: 54 years from the mathematics — the Radon transform, 1917 — to a machine that could image a human body. The laser: 43 years from Einstein predicting stimulated emission in 1917 to the first working device in 1960. Electricity itself: 51 years from Faraday's discovery of induction in 1831 to a power station lighting a street in 1882. In every case, close to a human lifetime disappears between understanding a thing and being able to make it.
Flying cars, fusion, humanoids: the same. The principle is visible. Fusion crossed net energy gain in 2022, for the first time anywhere. And still it does not reach society, because the time to turn it into matter is enormous.
So there is only one thing worth asking. Why is making things this slow? And is that slowness really unchangeable?
The Pandemic as a Mirror
In the spring of 2020, masks disappeared from stores everywhere.
Sanitizer went. Protective gear went. Not enough ventilators. Not enough test kits. Vaccines could be designed; the capacity to produce them could not keep up. The richest countries on earth could not make enough of a few layers of non-woven fabric for themselves.
Even the United States had to decide, in practice, who received a ventilator — which is to say, whose life to save. People died because we could not make things.
I do not know another moment when the inability to make was that exposed. The same structure appeared everywhere in different forms. Masks and sanitizer vanished in Japan and the United States. Dependence on imported pharmaceutical ingredients still has not been resolved. A little later, semiconductors ran short, lines stopped, and a new car you ordered did not arrive for months. A gap in the ability to make traveled down the supply chain until it reached an ordinary person who could not buy a car.
We moved too. In the middle of the pandemic we stood up an organization to concentrate on making medical devices and ventilation equipment. Hundreds and thousands of units a month, to hold up the front line. That was what the world needed then. Only those who could make anything were able to act at all.
That was when I became certain of something. The ability to make is now a matter of national security.
Does that sound overwrought? Go back through history and behind many of the worst decisions states have made, you find the same circumstance: some physical thing was not at hand. When oil, steel, rubber, medicine, or semiconductors are not at hand, people and countries get cornered. Imperial Japan is one example. The reverse holds as well: holding resources and productive capacity redundantly increases the moves available in a crisis. The ability to make things sets, directly, the range of options a country or a society can have.
This is not about one country. The vulnerability of those who cannot make is common to every country and every region.
And who physically makes? Designers, engineers, craftspeople, robots, factories, supply chains. Manufacturing.
Why Materialization Is Slow
Let me define the words I will use from here. I think of the path from an idea to something present in society in three stages: the flash of discovery (Spark), turning it into matter (Materialization), and spreading it through the world (Diffusion). What AI is directly accelerating today is Spark — discovery and ideas. But the middle stage, Materialization, standing an idea up as real matter, is the largest and the most time-consuming.
Professor Takahiro Fujimoto of the University of Tokyo defined manufacturing as transferring design information onto materials. It is a beautiful line. But Materialization is more than that transfer. Turning theory into a design that can actually be built. Turning one unit into a million at the same quality. Making it survive regulation and time, and keep running without breaking. I mean all of it with the single word.
Here is the core of it.
Making things is genuinely slow. That is a fact. But we have the reason wrong. We assume the slowness is a manufacturing problem. Steel is hard. Chemical reactions take time. Assembly is laborious. So it cannot be helped. That is what we think. Break it down, though, and something unexpected shows up.
Take a new car reaching volume production in four years. Four years is 1,461 days. Of those, how many days does it actually take to physically work steel and assemble one car?
The pure working time to build one car adds up to about one or two days. Stamping, welding, painting, assembly, inspection. Add all the net work and that is what you get. Add paint drying and the waiting between steps — the total time for raw material to pass through the plant as one car — and it is three to five days at most.
Which means: out of about four years, pure assembly work inside the automaker accounts for something like 0.1%.
Add everything else that involves physical activity inside a factory — making the parts, testing, prototyping, the manufacturing work at thousands of suppliers — and you still do not come close to 20% of the whole.
So where do the other 80%-plus go?
Almost all of it goes to work off the factory floor. White-collar work. Market research, planning, design, simulation, writing prototype specifications, evaluating test data, revising designs, regulatory work, documentation, supplier negotiation, alignment, quality reviews, designing the production equipment, launch planning. Not the time spent cutting and assembling inside the plant, but the enormous time spent working, coordinating, and deciding ahead of it and around it.
So: most of the slowness in Materialization is not in matter — not the hardness of steel or the rate of a chemical reaction. It is in the knowledge and coordination required to make the thing.
Some physical time cannot be removed. Concrete cures on its own schedule. A reaction proceeds at the rate of the reaction. A clinical trial has to wait the time a human body takes. There are legitimate reasons for the weight of regulation. All of that stays. But it is a small share of four years.
If that is true, the question inverts. If most of the slowness is not matter, then it is, in principle, slowness we built and can rebuild.
Go back to the e-commerce site. What took years thirty years ago now stands up in an hour. That is not a story about people writing code a hundred times faster. Payments, authentication, inventory, search — someone already built each of them properly and left them somewhere anyone can call. Pick a framework and the foundation is already assembled. The person building just brings it over. The site that goes up in an hour is millions of lines somebody else wrote. What got faster was not the writing. It was the amount you no longer have to write.
And this is the part that matters: it was already happening before AI. In an era when generative AI wrote not one line of code, standing up an e-commerce site had already gone from years to days. Software's hundredfold did not come from a smart machine arriving. It came from the accumulation of a system in which you can bring over the answer someone else already produced. AI landed on top of that foundation and made it faster still by combining the pieces well. The source of software's thirty years of speed was not the speed of the process, and not the intelligence of the machine. It was the reuse rate.
In software, a problem solved once is never solved again. The solution has a name and it sits somewhere another person can take it as is. The physical world is the opposite. That L-shaped bracket I deburred until I lost the feeling in my fingers is being designed today, again, by an enormous number of companies, each as if for the first time. And it is not only across companies. The same thing happens inside a single company. Manufacturing has no equivalent of GitHub or npm where past solutions become broadly reusable packages. Procurement, production engineering, manufacturing, quality assurance, aftermarket: all the same.
This is not a claim about speeding up manufacturing on the factory floor. What is especially slow is everything up to the point where you have settled what to make and how to make it. And there, the answers humanity has spent decades producing have never been turned into assets.
Who Crosses the Wall
Who is the protagonist in solving physical limits?
We need researchers. We need software engineers. We need logistics companies and regulators. But the heaviest role belongs to the people who make things. Designers, welders, production engineers, robots, factories, machines, suppliers. We call this manufacturing. The protagonist in crossing physical limits is manufacturing.
And manufacturing is only the current name for something humans have done since they became human.
We knapped stone into blades. We handled fire and fired clay into vessels. We joined wood into boats, invented the wheel, and walked out of the African savanna. Humans survived, multiplied, and spread across the whole earth not because we thought. Because we made. Thinking, other animals do. But the creature that takes an image in its head and turns it into matter with its hands was only us. We are Homo sapiens, and at the same time Homo faber.
So manufacturing is not a twentieth-century industry. It is not one industrial category among many. It is the means by which we hold the prosperity we have. Medicine, communications, power, cities, food — all of it sits on top of something somebody made. And when the next wall, physical limits, gets crossed, the ones crossing it will be the people who make. Manufacturing is not returning to the center. It has always been at the center. It just has not been treated that way for thirty years. That is what I mean by reclaimed.
Nor is this one country's story. Over the past thirty years the core of making moved from the United States, Europe, and Japan to East and Southeast Asia. China became the largest manufacturing nation. Korea in memory and shipbuilding, Taiwan in leading-edge semiconductors, built positions that do not move. Vietnam, Thailand, India, and Mexico accumulated capability in their own ways. The United States restarted domestic production. Europe began rebuilding its industrial base. Over the next thirty years, the protagonists in solving physical limits can be every country, region, and industry with the capability to make things.
And yet there are not enough people putting themselves into making things. The conversation is dominated by "AI changes society." But who builds the physical foundation AI runs on? Who builds the power plants, runs the semiconductor fabs, produces the robots at volume? And who changes physical reality itself? The largest problem we have to take on in the next thirty years is sitting right here, untouched.
Where I Stand
I am one of the people facing that question.
On that Sunday at twenty-six, all I could do was keep calling, apologize, and drive the replacement part over myself. Nearly ten years later, what I am trying to do has not really changed: make it possible for the people who make things to do it faster, and with more certainty. What changed is the scope — from one company's delivery date to makers everywhere.
At CADDi we set a ten-year vision last year: Accelerate Physical Innovation by 10x. There is no guarantee it works. But if the lead time to make something can be compressed hard, the cycle rate of discovery and making goes up, and the room AI has opened in intelligence finally reaches the physical world. Bits got more than a hundred times faster in thirty years. It should be possible in Atoms.
And what we are building now is not the physical product itself. It is the foundation underneath it — the thing that lets makers everywhere make far faster than they do today.
Back to Two Memories
The summer I was nineteen, I was backpacking alone in Southeast Asia. One night, in a small town in Cambodia, I was surrounded by boys my own age. They offered to take me out in a small boat, and I paid them. On the water they told me they were nineteen as well, and that their families did not own a car. They smoked and laughed. I thought we had become friends. At the end they pressed me for twenty dollars more, after swearing they would not take anything else.
I was not angry. They lived in a world where their families did not eat if there was no money. I was a college student from Tokyo, and having been born on one side of the world, I had far more physical options than they did. Sixteen years later, that summer has not faded.
Ten years after that, on New Year's Eve at twenty-nine, I was in Okinawa. Around the same hours, on the other side of the planet, my closest friend drowned in the sea off Miami and died. He was in the United States for an MBA. He and I had only ever clowned around together, so the idea of him sitting in a business school was funny to me. The last thing that ever passed between us was the message I sent right before he left: An MBA? That's not you at all, lol.
That is what physical limits are. Not statistics. Specific, named, physical absences. AI's progress does not fill those absences directly. But accelerating Materialization can genuinely reduce their number.
If a safe $1,000 car had been in those boys' house, they might have had more options for work. If a cheap, always-on set of eyes and a rescue drone that arrives in seconds had physically existed on that stretch of coast, my friend's outcome might have been different.
And it is not only limits on life and livelihood. There are fourteen large physical limits.
This piece is also my own homework.
Yushiro Kato
September 2026
