Is AI a Bubble Like the Dot-Com Era? What History Says
By Stephen Manning · @StephenCFU on X · · 14 min read
Is AI a bubble like the dot-com era? Probably in part, and the tech can still win. Amara's Law, how AI differs from 2000, and the signs that show where we are.
Is AI a bubble like the dot-com era? Probably in part, and that is not the same as saying AI is a fad. The dot-com crash ruined the investors who paid for the internet's buildout too early, and the internet still went on to deliver more than the hype promised, just years later than promised. My bet is that AI follows the same rhythm: adoption that feels slower than we want, and still spreads faster than the internet did.
The real question is whether all the money pouring into artificial intelligence, hundreds of billions of it, will come back out as earnings. "Bubble or no bubble" is the wrong way to ask it. What matters is where we are on the curve, and which parts of the buildout keep earning through a shakeout if one comes. Below is the pattern that keeps repeating, where AI differs from 2000, the signs I watch, and how I position without having to call the timing.
What is Amara's Law?
The most useful idea I know for investing in any new technology is named after Roy Amara, a futurist who led the Institute for the Future, and it is almost embarrassingly simple.
"We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run."Amara's Law
Both halves are true at the same time, which is what makes it so hard to trade. Early on, the hype gets ahead of the reality. Everyone drags the far-off endpoint into next year. Then reality disappoints, the "where is the return on all this money" complaints start, and people announce the whole thing was overhyped. That moment of disappointment is usually the point where the long run is being most badly underestimated, because the technology then spends the next decade quietly working its way into everything, arriving bigger than the hype promised but years later than it promised.
Has this happened before?
Many times. The internet is the obvious example. The Nasdaq Composite peaked in March 2000 and lost about 78 percent of its value by October 2002, and plenty of people declared the internet a bubble. They were right about the timing and dead wrong about the destination. Amazon and Google survived the crash, and the next twenty years of cloud computing and mobile were built on the wreckage.
This is not an internet quirk. It keeps happening.
Britain's railway mania reached its height in 1846, when Parliament passed 263 acts setting up new railway companies. The Bank of England had already raised interest rates in late 1845, and the speculation collapsed. Many middle-class families who had put their savings into the new companies lost everything. Yet the projects authorized between 1844 and 1846 added more than 6,000 miles of track, and the practical lines among them became part of the network that reorganized the British economy.
Electricity is my favorite case, because the lag had a specific cause worth knowing. Factories electrified early and saw little gain, because all most of them did was rip out the steam engine and drop one big electric motor in its place, still driving every machine through the same shafts and belts. The real leap came when factories were redesigned around small motors at each machine, laid out along the flow of the work. The economist Paul David traced how long that took: US productivity growth finally leapt in the 1920s, roughly four decades after electricity was commercialized.
GPS spent years as a mostly military system, and the US government deliberately degraded its accuracy for civilian users until May 2000. It took smartphones to turn it into the invisible layer under ride-hailing, maps and food delivery.
The same three things happen every time. The financial bubble comes early and pops before the technology pays off. The eventual winners turn out to be applications nobody could have named at the start. And the infrastructure everyone overbuilt during the mania becomes the cheap foundation the long run runs on.
One caution on that fiber row. The late-1990s fiber buildout was largely financed with debt, and when the money dried up, carriers such as Global Crossing and WorldCom went bankrupt in 2002. The glass they laid outlived them. Fiber in the ground stays useful for decades, which is exactly why the glut could become the cheap foundation of the next boom. A GPU is a short-lived asset, so an AI overbuild leaves far less durable infrastructure behind. The buildings and the power hookups last. The chips inside them do not.
Will AI spread slower or faster than the internet?
Slower than we want, faster than the internet. That is my honest bet: adoption will feel frustratingly slow for a few years and then compound faster than most people expect. Probably slower than the true believers promise, but faster than the internet took, for one concrete reason. The internet had to build its physical network almost from scratch. AI gets to run on top of cloud, mobile and connectivity that are already in place. In 2025, the International Telecommunication Union estimated that about 6 billion people, roughly three-quarters of the world, were using the internet. The pipes are already in the ground, so the moment a capability actually works, it can spread quickly.
There will probably be a reckoning too. The buildout is running well ahead of near-term demand, which is exactly what the pattern predicts: hundreds of billions in spending against revenue that is real but far smaller. Some companies may not survive the gap between the two. I would watch the leveraged, no-moat, pure-buildout players most closely, the ones financing hardware with debt and planning to sort out demand later. When money gets tight or the mood turns, those tend to go first, and the survivors buy their capacity for cents on the dollar.
How is AI different from the dot-com bubble?
In one important way it looks better. Many of the dot-com high-flyers went public without ever making a profit, some with barely any revenue, and they depended on the next funding round to stay alive. The leaders of the AI buildout are among the most profitable companies in the world. Alphabet alone reported $39.7 billion of operating income in the first quarter of 2026. That is a real cushion, for as long as the spending keeps coming out of profits rather than borrowed money.
In another way this market is more exposed than it was in 2000, and it is concentration. By the end of 2025, the 10 largest companies made up nearly 41 percent of the S&P 500, against a peak of about 27 percent during 2000, according to RBC Wealth Management. So if a few of those leaders stumble during a shakeout, the whole index feels it, including people who never bought an AI stock on purpose and simply own an index fund.
How can you tell where we are on the curve?
Three signs tell you roughly where you are, and I pay more attention to them than to any price chart.
The first is the gap between spending and revenue. When it narrows, adoption is catching up, and that is healthy. When it widens while the financing gets more aggressive, you are late in the overbuild, and the wreck is getting closer. For a sense of how wide that gap is, Bain & Company estimated in September 2025 that AI would need about $2 trillion in annual revenue by 2030 to fund the computing power needed to meet expected demand, and that even counting the savings from AI, the world would still be about $800 billion short. That is Bain's estimate, not mine, but it shows how far the revenue side still has to travel.
The second is the quality of that financing. Early in a buildout, companies fund it out of cash flow. Late in one, they start reaching for debt and clever off-balance-sheet structures. Borrowing is how the fiber carriers paid for their buildout in the late 1990s, and when you see the same shift, the shakeout is usually not far behind. The AI leaders started this cycle paying for the buildout out of operating cash, which is the clearest difference from 2000, so watch for the turn. In October 2025, Meta moved its Hyperion data center campus in Louisiana, about $27 billion of planned development, into a joint venture 80 percent owned by funds managed by Blue Owl Capital and partly financed with bonds sold to outside investors. Meta kept 20 percent, agreed to lease the campus and guaranteed part of its value for the first 16 years, yet because Meta does not control the venture, the campus stays off Meta's balance sheet. One deal does not make a wreck, but that is the kind of structure this sign asks you to count.
The third is how long the chips are supposed to last. Watch the useful life companies assign to their servers and AI chips on the books, and compare it with how fast new generations replace them. If the hardware wears out economically before it is fully depreciated, earnings look better than they really are, and the write-downs show up later. The big spenders do not even agree with each other. Effective January 2025, Amazon cut the useful life of a subset of its servers and networking equipment from six years to five, citing the faster pace of technology development, particularly in artificial intelligence, only a year after stretching it from five to six. That same January, Meta extended the useful life of certain servers and network assets to 5.5 years, a change it expected to reduce its 2025 depreciation expense by about $2.9 billion. Same kind of hardware, opposite assumptions, and the choice moves reported profit by billions.
What should you read in a big tech earnings report?
Those signs show up every quarter in the earnings reports of the big spenders, and the headline profit number is usually the least useful line in them. Here is what I actually read, because it applies to a lot more than one company.
Start by finding out where the headline came from. A big jump in earnings per share can come from paper gains on private-company stakes rather than from money customers paid. In its first-quarter 2026 report, Alphabet's earnings per share rose 82 percent to $5.11. Its own footnote shows that gains on equity securities, mostly unrealized gains on stakes in companies that do not trade publicly, added $2.35 of that. Same company, very different story, and the only thing that changed was whether you looked at the engine or at the accounting bolted onto it. That habit alone will save you from half the bad conclusions people draw at earnings time.
Next, read the cloud line and the backlog. That same Alphabet report showed Google Cloud revenue up 63 percent to $20.0 billion, and Sundar Pichai said the cloud backlog of signed commitments had nearly doubled in a quarter to over $460 billion. A backlog like that is the closest thing we have to proof that AI demand is turning into paid revenue rather than press releases, which is the revenue side of the first sign.
Then read the spending. In the same quarter Alphabet's capital spending reached about $35.7 billion and its free cash flow fell to about $10.1 billion, from about $24.6 billion the quarter before. The same release shows Alphabet also raised $31.1 billion that quarter by selling bonds, so read the financing line as well as the capex line. The whole bull-bear fight is over whether spending like that is conviction or recklessness. What tips me one way or the other is whether management ties the spending to signed backlog. Money spent against real commitments defends itself. Money spent against a hope does not.
Everything else, from model launch delays to researchers leaving for rival labs, is worth watching but rarely moves a quarter's earnings. Cloud and spending will.
How I am positioning, near and long
Here is the move, and it is the reason I sleep fine not knowing the timing. You do not have to predict when any of this pays off. You have to own the things that keep earning across whatever timing actually shows up, the same test I use when I ask whether a holding would still pay me if the market closed for five years.
Own the rails, not the train.
For the near term, the most dependable winners are the ones who get paid through the whole buildout no matter which app or model wins. That is the physical bottleneck: the companies that make the chips, the machines that make the chips, the custom silicon, and the memory. Everybody building AI has to buy from them, full stop. This is the layer that sells shovels in the gold rush, and it is the least dependent on guessing which application ends up mattering.
Sitting right next to it is the constraint almost nobody talks about until they trip over it: power. AI data centers eat staggering amounts of electricity, and you cannot magic electricity into existence. New generation and grid capacity come online slowly, gated by permits and construction crews. So the companies that generate and move power benefit directly, and the constraint itself acts as a brake on how fast the whole thing can overbuild. That is why I spent a good part of 2026 moving weight toward power in my own book.
For the long term, the winners are the applications built on top that we genuinely cannot name yet, plus the survivors who scoop up the cheap infrastructure after a crash. Nobody can name the next Uber-on-GPS today. What you can do is own the layer every future application will have to run on.
The winners nobody files under "AI"
This is the part I find most interesting, and it gets almost no airtime.
Everyone argues about the chip makers and the model labs. Meanwhile there is a quieter, sturdier way to win from AI, and it belongs to companies that already dominate their markets and simply use AI as a tool to widen the moat they already have. What makes this so appealing is that the benefit does not require the company to win a new market, launch a product, or outspend anybody on capital. It just makes a business that already prints money a little more efficient and a little stickier.
Take a warehouse membership retailer. Its real moat is not the pallets of paper towels, it is the renewal rate on the memberships, and the fee income that drops almost straight to profit. AI does not change what that business is. It makes the company better at spotting which members are about to drift away so it can pull them back, and it sharpens the brutally efficient buying operation that lets the retailer beat everyone on price in the first place.
Or take a consumer-products giant with dozens of category-leading brands. The moat there is brand and scale and shelf space, and AI mostly shows up as lower cost, smarter marketing and a leaner supply chain, which lands as margin rather than as some new revenue line. That is about as low-risk as a benefit gets.
A cloud-and-logistics conglomerate gets it from both ends at once. Its cloud arm sells the compute AI runs on, and AI simultaneously tightens its own enormous logistics and inventory operation. Two payoffs from one technology, and only one of them makes headlines.
The strongest version of this whole idea is the proprietary data moat. AI is only as good as the data feeding it, and companies sitting on unique, hard-to-copy data watch that data become more valuable without lifting a finger. Financial data. Legal and professional data. Medical and scientific records. These firms do not have to spend to capture the upside, because they already own the asset, and AI just repriced it higher.
Tying it together
Here is the shape of a portfolio built for all of this, and notice that none of it asks you to call the timing.
That mix does not care much when AI arrives. If it comes faster than expected, the infrastructure and the moats compound faster. If it disappoints for a few years first, which Amara's Law says it probably will, those same holdings keep earning while the overextended players get flushed, and you are already positioned for the long run when it finally shows up.
One last discipline. Watch that spending-versus-revenue gap and the quality of the financing, and treat a widening gap funded by fresh debt as your cue to get more defensive. A falling stock price on its own does not tell you much. It is just weather. The financing going bad is the thing that turns an overbuild into a wreck, and that is the sign worth acting on.
Every earnings season from the big spenders is the next data point. Read the cloud line, the backlog and the way management explains the spending, and mostly ignore the headline.
I write about patterns like this in the CFU newsletter, which goes to 2,300+ investors. You can join it here.
Cash Flow University
Educational writing about market patterns, not investment advice, not a recommendation to buy or sell anything, and not tailored to your situation. Every company and pattern here is an illustration of an idea. Do your own work, and talk to a licensed advisor before you make any move. Assume I hold positions in some of what I have described.
Filed under: Market Analysis