Most people meet AI through a text box. You type a question, an answer comes back, and everything behind it stays hidden. But that hidden part is huge. It's giant buildings full of computer chips, the power plants that feed them, and the trillions of dollars being lined up to pay for it all. And it's quietly turning into one of the biggest money stories of our time.
Two experts, Dwarkesh Patel and Dylan Patel, had a long talk about this. They started with something small, the price of computer chips. Then the talk kept growing. By the end they were arguing about interest rates, national debt, China, and who ends up with real power in a world where thinking can be made in a factory.
This piece walks through that argument in plain words, and fills in the parts you may not know yet. Keep one thing in mind the whole way. Almost none of this is proven. It's mostly a careful guess about the future. The ideas matter not because they're sure to happen, but because the forces they point at are already moving.
A few words you need first
Four words unlock the rest.
Compute means computer power. It's the ability to do the math that AI runs on. It comes from special chips, most famously the ones made by Nvidia, kept inside data centers. A data center is just a big building packed with computers, fans to keep them cool, and cables.
Training is the slow, costly job of building an AI model. You feed it huge piles of information until it learns. Using the model is what happens every time you ask it a question and it answers. Think of training as writing a schoolbook, and using it as millions of kids later reading that book.
Here's the part that changes how you see all of it. Computer power is now measured in electricity. People talk about how many megawatts a data center uses, and that's just a unit of power, like the ones on your electric bill. So AI isn't only clever code anymore. It's a heavy industry. It needs electricity, land, and water for cooling, just like a steel plant.
Two last words. The big cloud companies are Amazon, Microsoft, and Google. They own most of these buildings. The top AI labs are the companies building the smartest models, like OpenAI, Anthropic, and Google DeepMind.
The loop at the center of it all
The main idea is a loop. Better AI makes more money. That money buys more computer power. More power builds even better AI, which makes even more money, which buys even more power. Round and round.
At first, AI companies just lost money. The argument is that this is flipping. A top model may now earn a few times more from a batch of computer power than that power costs to run. The two experts toss out rough numbers. Running one megawatt of power might cost ten to fifteen million dollars, while a top lab could earn a few times that from it.
If that's true, computer power stops being just a cost. It starts making money for you, the way a factory does. And that flips the smart move. If one dollar of power gives you back four or five, you don't hand the profit to your owners. You buy more power. That's the difference between a normal business that makes money and a machine that keeps feeding itself and growing.
Now make the loop faster. What if the AI helps build the next AI? Today people build better models. Soon the models may help do that work themselves. Then each better model helps make the next one come sooner, and the one after that sooner still. The loop stops crawling and starts to sprint. This idea has a name. People call it AI improving itself. And it's the hinge the whole story turns on, because a loop that speeds up on its own behaves very differently from one that plods along.
Which one AI turns out to be is the question sitting under everything else.
Why the power ends up with a few companies
Follow the loop and you can guess what happens. Whoever makes the most money from a batch of power can pay the most to get it. Say a normal customer earns fifteen million from one megawatt, but a top lab can earn sixty. The lab can happily pay thirty or forty to grab it. And whoever owns the building will sell to whoever pays most. So the power flows to the labs that make the most from it.
One small catch worth noticing. Owning the power doesn't mean owning the building. If Amazon owns the data center but runs Anthropic's work inside it, that power is really Anthropic's. What matters is who's paying, not whose name is on the door.
A few things push the same way and stack up. The huge cost of training gets shared across more and more users, so it gets cheaper per person. The best model earns more per chip. The best model helps build the next one. More users give more information to learn from. More money buys more power.
Now here's where it gets heavy. Picture computer power turning into workers. If AI gets good enough to do real jobs, then every extra chunk of power is like hiring more people who never sleep. And two things are growing fast at once. The chips get better every year, and the models get smarter every year. Put those together and the number of AI "workers" a company can run might grow something like ten times a year. Let that sink in. A single company could end up commanding more working minds than an entire country has people. That's the moment a story about chips becomes a story about power.
But there's a fair argument on the other side, and it's worth just as much space. AI isn't like oil. Ideas leak. Smart tricks spread fast among people who study this. Free, open models keep getting better. Customers can run their own. Big companies tend to get slow and clumsy, which is usually how smaller ones catch up. So this pile-up in one place isn't sure to happen. It only happens if AI breaks the normal ways markets spread out, and it hasn't clearly done that yet.
The strangest idea in the whole talk
Here's the part that sticks with you. Say a customer, like a big trading firm, could make a fortune using a lab's model. Why should the lab sell it to them at all?
If the lab could use that same power itself instead, to make its own models smarter, design its own chips, or start its own business, and earn more that way than the customer would ever pay, then the smart move is to keep the AI for itself. That's not a monopoly built on rules or patents. It's built on a simple fact. The best thing to do with smart AI might be to make even smarter AI.
The same idea flips something people expect. Most think that once AI gets popular, everyday use will be far bigger than training. Maybe not. If using power for research speeds up the next model, and the next model is worth billions, then answering customer questions is the smaller prize. So answering questions becomes a way to pay for training. You serve enough users to bring in cash, then you pour that cash and those chips into building the next model.
If it goes this way, regular people could end up with weaker AI than the labs quietly keep for themselves. That would be a new kind of gap between the powerful and everyone else.
Where does the money actually end up
Now for the hopeful side. The value AI creates is not the same as the money the lab keeps. Say a firm pays a lab ten million and uses the AI to earn fifty. The lab kept ten. The customer kept forty. So even labs that make loads of money might only get a small slice of the total value AI creates. That would keep the money fairly spread out.
There are two more reasons the money might stay spread out. First, when AI gets cheaper, people don't just save money. They use way more of it. Cheaper often means people use so much more that the total spending goes up, not down. Second, AI can't do much alone. An AI that helps design medicine still needs labs and factories and years of testing. An AI that helps run a business still needs the business. So the people who own all that real-world stuff may keep a big share of the reward, because you still can't do anything useful without them.
The money could end up in many places. The lab. The chipmakers. The building owners. The power companies. The customers. And it tends to move around over time instead of settling. There's also a slow chain reaction. If labs are making huge profits, then Nvidia, the memory makers, and the power companies slowly notice they're leaving money on the table, and they raise their prices to grab a share. The money doesn't sit still.
When chips run into the real world
Dwarkesh has the simple hopeful view. If the money is this good, companies will just build more. More chip plants, more power plants, more data centers, until there's enough to go around.
Dylan agrees it'll happen, but says it'll be slow, and this is where knowing the industry helps. Money can move overnight. Factories can't. A modern chip plant takes years to build. The machines that make the best chips need special lenses from one single company in Germany, and that company can't just double its work. Special packaging, fast memory, big power machines, all of it takes a long time to make. So you get massive demand meeting slow supply. And when something is scarce, whoever controls it gets rich until the rest of the world catches up.
Where it turns into a money-and-government problem
Add it all up and the numbers get scary. The talk imagines something like eleven trillion dollars spent on AI buildings by the end of the decade, some of it borrowed, eventually eating a real chunk of the whole US economy. And that size becomes its own brake. A dollar, a worker, or a big machine used for a data center can't also build a house or a factory. So AI ends up fighting everything else for the same limited stuff.
That fight shows up most in the price of borrowing money. If AI projects earn amazing returns, AI companies will offer high interest to borrow, higher than a normal company can. Why would a lender take five percent from a boring old company when an AI project offers nine? So money walks toward AI, and everyone else, families buying homes, small shops, even governments, has to pay more to compete. That could push interest rates up across the whole economy. Not because a central bank is fighting rising prices, but just because AI became a better place to put money. The direction of that makes sense. The size of it is a real guess, because interest rates depend on many other things too.
Higher rates spread in painful ways. Governments that owe a lot have to borrow again at these higher rates, so their bills go up, right when AI may be shrinking the taxes they collect by replacing workers. Countries that don't own any of the AI could get the worst mix. Higher costs to borrow, lower taxes coming in, and workers losing jobs, all at once. The experts compare it to the early 1980s, when US interest rates shot up and set off debt troubles across poorer countries. The country-by-country guesses here are the shakiest part of the whole talk, so take them lightly. But the basic idea holds together.
There's even a surprise for people who own stocks. If AI really pushes growth and interest rates way up, then the usual math for pricing stocks turns against slow, steady, old companies. AI could make the world much richer while making a lot of today's stocks and bonds worth less, because the new AI stuff is so valuable that the old stuff starts to look outdated next to it. Getting richer and having your old things drop in value can happen at the same time.
Rules can cut both ways
You'd think rules would just slow AI down. The sharper point is that rules can slow it and squeeze it into fewer hands at the same time. Ban new data centers, and being short on space helps whoever already has some. Block a model from being shared with the public for six months, and the lab that can still use it in private pulls even further ahead of everyone stuck on the old one. So a rule meant to shrink AI's reach can quietly grow the lead of the top few. The line that really matters is between rules that stop the public from getting a model and rules that stop labs from building better ones in the first place. Only the second kind actually slows things down.
Countries, and why timing is everything
The same idea works for whole countries. Right now the US has far more of the best AI power than China, and US rules block China from buying the best chips. Whether that lead really matters comes down to one thing. How fast does AI improve?
Remember the loop that speeds up on its own. If AI improves fast, and starts helping build better AI, then even a short lead can turn into a lasting one, because the leader pulls away faster than anyone can catch up. But if AI improves slowly, the rules only delay China, which has shown over and over that it can build things at massive scale once it decides to. So the value of the rules depends entirely on how fast AI moves. And here's a twist. Chinese models have stayed surprisingly good using far less computer power, a sign that having more power may matter less than the simple story says.
What it all comes down to
Take away every number and one idea ties it together. Once AI is worth enough, using anything for something other than AI starts to feel costly. A megawatt spent on normal computing can't train a model. A dollar lent for a home can't fund a data center. A question sold to a customer can't be used by the lab's own researchers instead. When one use of a scarce thing becomes hugely valuable, every other use looks expensive, and we all have to keep choosing what to give up.
The two experts hold the two halves of the answer. Dwarkesh shows the push. If smart AI is worth a fortune, the world will bend itself to make as much of it as possible. Dylan shows the drag. Bending the whole world runs straight into money limits, slow factories, angry voters, and plain human pushback. Both can be true at once. Which one wins is what decides the future.
One honest warning before the ending. This whole story is a chain of guesses, and each guess sits on the one before it. When you stack that many maybes on top of each other, the odds that all of them come true drop fast. So hold the big picture, not the exact numbers. The point isn't to predict the future. It's to know which few things to watch as the real numbers come in.
And the deepest point isn't about when super-smart AI shows up. It's this. If thinking becomes something you can build in a factory, and only a few companies own the factories, then who owns those machines becomes one of the biggest questions in money and in politics. That's how a chat about the price of a chip ends up being about debt, power, and who runs things. The concrete being poured today isn't just plumbing for a chatbot. It might be the base for who holds power for the next thirty years.
So the next time an answer pops up in that little text box, the real question isn't how smart the machine is. It's who owns it, and what they'd rather do with it than talk to you.