Yes, AI is a bubble.
And what it means for you.
After letter #100, a reader wrote to me with one question.
They'd been reading about the AI boom and wanted to know if it's heading for a bust. Like the dot-com crash. Like 2008. Like 1929.
My answer is yes.
AI is a bubble. And I think that's mostly good news for you.
I know how that sounds from someone who builds AI systems every day and teaches people to use them. So let me show you the math, why I think it can't close, and why the crash, whenever it comes, matters much less for your work than for the people who paid for the party.
(One thing before we start: none of this is investment advice. I'm talking about your work, not your portfolio.)
The math doesn't close
This year, the four biggest tech companies plan to spend somewhere around $700 billion on AI infrastructure. That's almost double last year. Data centers, power, and above all chips.
Two Stanford economists, Jared Bernstein and Ryan Cummings, did the math. To justify what they're spending, the biggest players need somewhere between $2.4 and $3 trillion a year in new AI revenue by 2030. That's five to six times what they're making today.
That's only half the story. A big part of that money buys chips, and chips lose their value in about five years. Railway track carried trains for a century. These chips have to pay for themselves quickly or not at all.
(I didn't know about this until recently.)
So for the money to come back, somebody has to pay a toll on AI. A big toll, for a long time.
That's the bet. And I don't think the toll can hold.
Why the toll won't hold
Here are the three reasons.
First, intelligence keeps getting cheaper. The investors at a16z tracked the price of a fixed level of AI capability. The level GPT-3 had when it launched cost $60 per million tokens in late 2021. Three years later, you could buy the same level for 6 cents. A thousand times cheaper. Their estimate: for the same capability, the price falls about ten times every year.
Now try to repay a five-year investment on a product whose price drops ten times a year.
Second, the frontier doesn't stay private. Epoch AI, a research group that tracks this closely, measured how far the best open-weight models trail the best closed ones. About four months. Four months after a lab ships something extraordinary, a free version that's almost as good is downloadable, often from a Chinese lab, and it runs on hardware you control.
I've watched several of my clients make exactly that move this year: take a workflow off a premium model, run it on an open one, and cut the bill dramatically. (Letter 100.)
The pink line is the best open-weight model you could download at each point in time. In early 2023 it was more than a year behind the best closed model. Today it's about four months. Source: Epoch AI, CC BY.
A moat that lasts one season isn't much of a moat.
Third, there's no stickiness. This is the one I keep coming back to. Social media, marketplaces and delivery apps all get more valuable the more people use them. Leaving costs you your friends, your sellers, your history. AI has none of that. You can switch models in seconds, and your "memory" (if you're well trained in using AI) increasingly lives in files on your own laptop (or drive or Notion...).
Put the three together and you get my definition of this bubble: a commodity priced like a monopoly.
The valuations assume someone will own AI the way someone owns search or social. The economics say intelligence is turning into something closer to electricity or bandwidth: essential, everywhere, and cheaper every year.
The labs know this. When you can't build a moat with the product, you try to build it somewhere else. I see three ways out:
Become the bank: regulated, compliant, "the ONLY safe option" (Letter 100).
Become the memory: devices that go everywhere with you and collect your context, so that leaving costs you your history.
Become the utility: own the data centers and sell the infrastructure itself.
Each can work for somebody. All three are harder businesses than the one the valuations assume.
The best argument against me
Let me give you the strongest case on the other side, because it's a good one.
It goes like this. Demand is exploding. An agent reads, plans, tries, fails, retries and checks, and every one of those steps burns tokens. So even if the price per token collapses, we'll use so many more tokens that revenue grows anyway.
Economists have a name for this. In 1865, William Stanley Jevons noticed that as steam engines burned coal more efficiently, Britain ended up burning far more coal. Cheaper energy, more energy used.
I think that's right about usage. I see it with my own clients: as agents arrived, the AI bills went up.
But exploding demand doesn't tell you who gets paid. If a dozen models can do the job, the volume goes to whoever is cheapest, and the cheapest keeps getting cheaper. Electricity demand grew for a century, and that never turned every power company into a gold mine.
Huge demand and a commodity price can live together. That combination is great for the people using the thing, and brutal for the people who paid to build it at monopoly prices.
A bubble doesn't mean a scam
Now, the part most people get wrong about bubbles.
A bubble says the price is wrong. It says nothing about whether the thing is useful.
Some of the most useful things we have came out of bubbles.
In the 1840s, Britain went mad for railways. In 1846 alone, Parliament authorised new lines totalling around 9,500 miles. By 1847, railway investment had reached roughly 7 percent of the country's national income. Then the crash came. Railway shares kept falling until 1850, and a lot of people lost a lot of money.
But by 1850, Britain had around 6,000 new miles of track. That track became the backbone of the country's railways for generations.
The railways went bust. The trains kept running.
The same thing happened a hundred and fifty years later. In the late 1990s, telecom companies buried fibre-optic cable everywhere, convinced internet traffic would pay for all of it. It didn't, not fast enough. The sector lost something like $2 trillion in market value between 2000 and 2002, and by 2002 most estimates put 90 percent or more of that fibre sitting dark, unused.
Then the price of bandwidth collapsed. And a few years later, cheap bandwidth made things possible that nobody had budgeted for. A small startup called YouTube launched in 2005 and started streaming video to the whole world.
The pattern is always the same. Investors pay for the overbuild. Everyone else inherits the infrastructure at a discount.
What the AI crash would look like
I have no idea when it happens. Nobody does. (Anyone who gives you a date is selling something.) It might be a sharp crash, or a long, slow deflation where valuations quietly come down to earth over a few years.
But I'm fairly confident about the shape.
The price of good-enough intelligence keeps falling until it's close to free. Data centers built for a toll that never came get sold off cheap, and someone buys that compute at a discount. Open models keep closing the gap, and more of them run on your laptop or a small box in the office. The very best frontier models will still exist, and they'll still be expensive. But for most of what most people do at work, good enough will be almost free.
That's the "hardware market" scenario from Letter 100. The bubble is how we get there.
In that world, the money goes to whatever can't be copied in four months: chips, energy, data centers, the devices the models run on. That's why I put NVIDIA in a different bucket from the labs. Whoever wins the model race, the models run on someone's hardware.
One honest caveat, though. In the dot-com crash, Cisco sold the routers for the whole boom, and its stock still lost most of its value while the internet kept growing. A good business can still be a painful stock in a crash. (Again: your work, not your portfolio.)
Who gets hurt
The investors who priced AI like a monopoly, obviously.
Then the businesses that are too dependent on someone else's model. If your product is a nice interface on top of a model you don't control, the price of what you resell is falling ten times a year, and your customers will notice.
And then, quietly, you. If you tied yourself to a single vendor.
If all your workflows, prompts, projects and "memory" live inside one product, you've built your house on somebody else's balance sheet. If your company just signed a three-year commitment at today's prices, it locked in a price the market is about to cut.
All of them made the same reasonable bet: that the toll would hold.
The uncomfortable part
The people telling you most loudly that this is not a bubble are, very often, the people whose wealth depends on it not being one. That alone doesn't make them wrong, but it's a good reason to weigh their certainty accordingly. Follow the incentives.
But there's an opposite mistake, and I see it more and more.
Some people hear "bubble" and relax. They decide to wait for the crash, let the hype die down, and learn this stuff later, when it's "real."
That's backwards.
A crash makes the tool cheaper. And a cheaper tool makes the skill matter more. When intelligence costs about as much as bandwidth, every company in the world has access to it, and the only thing still scarce is knowing what to do with it. The person who can tell a good output from a plausible one. The person with the expertise, the judgment and the context to point the machine somewhere useful.
In a nutshell: the people who can make YOU + AI >>> AI.
That's been the thesis of everything I do (this newsletter, my books, courses, bootcamps), and a crash only makes it more true.
What I'd do on Monday
Make sure you could switch models in a day. Keep your context in your own files: your instructions, your examples, your notes on how your work is done. Then actually try your two or three most important workflows on a second model, including an open-weight one. If they break, you've found your lock-in before the market found it for you. If you don't know how to do it, join the AI OS Course waitlist. Access opens in a few weeks. I've honestly put everything I know about using AI properly (for non-technical people) in that course.
Don't lock in today's prices. If you buy AI for a team or a company, prefer short terms and usage-based pricing to long commitments. You're buying something that gets cheaper every year. Let it.
Put your time into what compounds. Your expertise, your judgment, your written-down context. In a world where intelligence is nearly free, those are the only parts of the stack that get more valuable.
Back to the question
So, will AI burst like 1929, like dot-com, like 2008?
My honest answer: probably more like the railways.
A lot of money will disappear. Some famous companies won't survive it. And the infrastructure will stay, get cheaper, and end up carrying things nobody has imagined yet.
When it happens, I want you to be the one using the rails, not the one who paid for them.
Bubbles burn the money. They leave the rails.
AI is only as good as the human operating it.
Have a great weekend.
Stay sharp.
— Charafeddine (CM)
