Railroads and dark fiber ran this exact playbook. Both times the technology won and the builders died. Here is why the AI build-out looks identical, and why your RAM got more expensive because of it.
I build on this stuff for a living. My company runs on cloud infrastructure. We pay for AI tokens every month. And a few weeks ago I looked at a quote to add memory to a machine and did a double take. The price had nearly doubled in a year.
That RAM quote is where this story starts. Because the reason memory got expensive is the same reason people keep asking if AI is a bubble. And to answer that honestly, you cannot just look at AI. You have to look at the two times we did almost exactly this before.
The pattern is older than the internet
Infrastructure bubbles all rhyme. Someone spots a genuinely world-changing technology. Money floods in. And then comes the move that defines every one of these episodes: companies build capacity far ahead of real demand, betting that demand will show up to fill it.
Sometimes it does. Usually it does not, at least not on the schedule the spending assumed. The gap between capacity built and capacity used is where the money dies.
We have run this experiment twice at national scale. First with railroads. Then with fiber. We are now running it a third time, with data centers.
Railroads: the original overbuild
Britain went first. During the Railway Mania of the 1840s, Parliament passed 263 acts in the single year of 1846 to set up new railway companies, for routes totalling around 9,500 miles. A lot of that track was planned on the assumption that traffic would arrive later. Much of it never earned back what it cost.
America did it bigger. Between 1866 and 1873, roughly 35,000 miles of new track were laid across the country, a lot of it financed on optimism and shaky debt. Then the bill came due. In the Panic of 1873, 89 of the country’s 364 railroads went bankrupt, and around 18,000 businesses failed in two years.
We did not even learn from it. Twenty years later, the Panic of 1893 hit for the same reason: too much railroad, financed too loosely. By mid-1894 a quarter of all US railroads had failed, more than 40,000 miles of them.
The technology was real. Railroads did reshape the world. But being right about the technology did not save the people who overbuilt it.
Fiber: capacity nobody could use
Fast forward a century. During the late-1990s telecom boom, companies laid more than 80 million miles of fiber optic cable across the US. The pitch was a number WorldCom kept repeating: internet traffic was doubling every 100 days. It was not. Real traffic was roughly doubling once a year, which is fast, but nowhere near the story the spending was built on.
So what happened to all that glass? Barely any of it got used. By some estimates less than 5% of the fiber laid in the boom was ever lit. Even four years after the bubble burst, 85% to 95% of it was still sitting dark. The industry literally invented a name for capacity built and never used: dark fiber.
The companies that laid it mostly did not survive to enjoy it. Global Crossing raised billions to wire the planet, then filed for bankruptcy in January 2002 with $12.4 billion of debt and thousands of miles of unused cable, after its chairman had quietly sold more than $700 million of his own stock. 360networks completed a $900 million IPO and was bankrupt about fourteen months later. WorldCom went down in July 2002 as the largest bankruptcy in US history at the time, $107 billion in assets, held up by roughly $11 billion of accounting fraud. Corning, which made the actual glass, fell from nearly $100 a share in 2000 to about $1 in 2002. All in, telecom stocks lost more than $2 trillion in value.
And the purest version of the madness was next door, in dot-coms. Pets.com went public in February 2000 having booked $619,000 of revenue while spending $11.8 million on advertising. It sold products for about a third of what it paid for them. It went from IPO to switching off the lights in 268 days, its stock from $11 to 19 cents, roughly $300 million burned. It is a punchline now. At the time it was a stock people lined up to buy.
Same shape as the railroads. Real technology. Genuine future. Ruinous timing.
AI: same script, much bigger budget
Now look at what is happening today.
The five biggest US cloud and AI players, Microsoft, Alphabet, Amazon, Meta and Oracle, have signalled combined capital spending of roughly $660 to $690 billion for 2026, close to double what they spent in 2025. Across the largest data center operators globally, the spend is heading toward $750 billion in a single year, and about three quarters of that, roughly $450 billion, is tied directly to AI: the chips, the servers, the buildings.
Each of these data centers costs billions. They are being built now, at full speed, for demand that is expected later. That is the tell. It is the railroad move and the fiber move, in concrete and silicon.
So here is the fair question. Is the money actually coming in to justify it?
Look at the two names everyone points to. OpenAI is running at roughly a $24 billion revenue rate as of early 2026, and still projecting a $14 billion loss for the year, with no profit expected before 2029 or 2030, while preparing to ask investors to value it above a trillion dollars. Anthropic has grown fast too, to around a $30 billion annual run rate.
Add up the frontier labs and you are somewhere in the range of $50 to $60 billion of annual revenue. Set that against roughly $450 billion of AI infrastructure spending in a single year. The revenue is real and growing quickly. It is also nowhere near what the build-out needs to break even.
And the demand everyone is counting on, enterprise adoption, is not showing up on schedule. An MIT report last year, looking at 300 deployments and 150-plus executive interviews, found that despite $30 to $40 billion of enterprise spending on generative AI, about 95% of organisations were seeing no measurable return. Only 5% were getting real value. Everyone is building for the enterprise wave. The enterprise wave, so far, is mostly stalled pilots.
This is why your laptop got more expensive
Here is the part that reaches people who never touch a data center.
Three companies, Samsung, SK Hynix and Micron, make over 95% of the world’s DRAM, the memory in your laptop and your phone. AI accelerators need a special, more expensive kind of memory called HBM, and it is far more profitable to make. So those three quietly shifted capacity toward HBM for the AI build-out and away from the ordinary memory the rest of us buy.
The result is a squeeze. Data centers now consume an estimated 70% of the memory chips produced worldwide. Prices have jumped hard. Samsung pushed a 32GB DDR5 module from about $149 to $239, a 60% jump. Contract prices for DDR5 have more than doubled. Samsung and SK Hynix are warning the shortage runs into 2027 and beyond, with customers reserving supply years ahead.
That is the bubble touching your wallet. You did not buy an AI product. But the AI build-out bid up the memory in the device in your pocket, and you are paying for it anyway.
So who actually won last time?
This is the twist worth sitting with.
When the fiber bubble burst, all that dark fiber did not vanish. It got sold off cheap. And the companies that scooped it up were not the ones that laid it. Google, Amazon and Facebook bought or leased that surplus capacity for a fraction of what it cost to build, and used it as the backbone for search, cloud and video. The infrastructure outlived the companies that funded it, and the winners were the ones who showed up after the crash with cash and a use for it.
That is the honest hopeful side. Demand for fiber did eventually catch up. It just took more than a decade, and it arrived long after the original investors were wiped out. The glut was real. The future was also real. Both things were true.
But this time the index is not made of Pets.coms
Here is the fairest objection to all of this, and it is a strong one.
The companies driving today’s build-out are nothing like Global Crossing or Pets.com. The fiber and dot-com bubbles floated on businesses with thin profits, sometimes almost no revenue, that needed a constant drip of fresh capital just to breathe. When the capital stopped, they were gone in months.
The names leading the AI spend are the opposite. Microsoft, Alphabet, Amazon, Meta, Nvidia and Apple are some of the most profitable companies that have ever existed. The so-called Magnificent Seven throw off something like 70% of the entire economic profit of the S&P 500. They are funding most of this build-out from the cash their existing businesses already generate, not from speculators who can vanish overnight. Cisco traded at 200 times earnings in 2000 on a far shakier story than any of these carry now.
So no, Google is not going to evaporate the way 360networks did. These giants can be wrong about AI for years and still be standing. That is a real difference, and it matters.
But look closely at what that argument actually says. It says the incumbents survive. It does not say the spending pays off. A profitable company can pour hundreds of billions into capacity that demand never fills, take the writedown, and walk away intact. The build is still an overbuild. The money is still gone. It just does not take the company down with it.
And the speculative layer has not disappeared. It has moved. Instead of Pets.com, this cycle has AI pure-plays and neocloud GPU landlords raising on the promise of the future, some of them buying chips with debt against contracts that only hold if the boom holds. If there is a Global Crossing hiding in this story, it is probably not Google. It is one of the names you only started hearing two years ago.
So, are we in a bubble?
Put it plainly. On the evidence, yes. Capacity is being built years ahead of the revenue and the adoption that would justify it, increasingly on debt, on a demand story that has not yet arrived. That is the exact pattern that broke railroads and telecom.
But a bubble does not mean everyone dies. The last two times, the technology was real and the future did arrive. It just arrived for different people than the ones who paid for it, on a slower clock than the spending assumed. The builders ate the loss. The infrastructure, and the payoff, went to whoever was still standing afterward with cash and a use for it.
The technology is not the bubble. The technology is probably the most important thing many of us will build on in our careers. The bubble is the belief that demand will arrive exactly when the spending needs it to. And “this time it is actually useful” is the line every bubble tells, right before it proves that being right and being early are two separate ways to go broke.
So no, I am not betting against AI. I am building on it. But when I sign off on that doubled memory quote, I remember how the last two of these ended.
Railroads were real. The people who overbuilt them went bankrupt, and someone else ran the trains. Fiber was real. Corning went from $100 to $1, and Google bought the cable for scrap. Both times, the future showed up exactly as promised. It just showed up for different people than the ones who paid for it.
That is the part nobody spending $450 billion this year wants to hear. The winners of the AI era may not be the names on today’s invoices. They may be whoever is standing there with cash and a use for it after the correction, buying the future at the price of scrap.
Twice is a coincidence. We are about to find out if three is a pattern.
Top comments (0)