AI Is a Bubble. So Was the Internet.

Is AI a bubble like the dot-com era? The honest answer, as of mid-2026, is that parts of it are, and pretending otherwise is as lazy as the doom takes it is reacting to. The AI bubble debate has collapsed into two useless camps: this is 2000 all over again, run, versus these are profitable companies, relax. Both skip the specifics, and the specifics are where the answer lives.

Here is the title spelled out, because it is the whole argument: AI is a bubble in places, and the internet was a bubble too, and the internet still changed everything. Those facts coexist. The better question is not the binary "is it a bubble," which is unanswerable, but "which parts are bubble, which are real, and what survives the deflation." That you can act on.

What "bubble" actually means

A bubble is not just high prices. As INSEAD's faculty put it, a bubble exists when prices exceed what future fundamentals can realistically deliver, not merely when they exceed today's. That distinction is the crux, because the AI thesis can be entirely correct and the market can still be mispriced. Two separate claims.

The scale is real either way. Hyperscaler AI capex was roughly $400 billion in 2025 and is tracking near $700 billion in 2026, according to Goldman Sachs estimates, and AI spending drove an estimated 42% of US GDP growth in the first quarter of 2026. When one category of corporate spending moves the national growth number, the question stops being academic.

The signal that should worry you: circular financing

If you want the one thing that rhymes most with the dot-com bust, it is this. Nvidia agreed to invest up to $100 billion in OpenAI, which OpenAI then spends on data centers filled with Nvidia chips. Bernstein analyst Stacy Rasgon flagged the obvious problem the day it was announced: this fuels circular concerns. By some estimates there is over $800 billion in interlocking deals across the chip, cloud, and model layers, a figure worth treating as a rough order of magnitude rather than a precise count.

This is the structure that took down Global Crossing and WorldCom. Suppliers financed their own customers, demand looked organic, and revenue looked robust right up until the dollars stopped circling. The test, again from INSEAD, is whether real end-user demand is being generated or whether money is mostly moving in a loop. The difference between a flywheel and a house of cards is exactly that, and from the outside the two look identical for a long time.

There is a sober number on the demand side too. Sequoia's David Cahn has argued the industry needs to fill a roughly $600 billion annual revenue gap to justify the capex it has committed, and through 2026 that gap has been widening, not closing. The people running these companies dismiss the fear, which is exactly what you would expect them to do, and it settles nothing.

The signal that should reassure you: real money, real demand

Now the other side, with the same rigor. This is where the post has to put a number on the bull case, not just an adjective.

AI cloud revenue is large and growing fast. On Q1 2026 earnings, Microsoft confirmed its AI business passed a $37 billion annual run rate, up 123% year over year, with AWS and Google Cloud also growing at AI-driven double digits. At the model layer, OpenAI ended 2025 near $20 billion in annual recurring revenue, roughly tripling year over year, and Anthropic was around a $9 billion run rate. The hyperscalers report being supply-constrained, not demand-constrained. That is real cash from customers outside the financing loop, and it is the strongest argument that the demand is not a mirage.

The funding structure also differs from 2000. The companies making the largest bets are profitable incumbents funding capex largely from earnings rather than debt, which removes the systemic-leverage fuse that turns corrections into catastrophes. Aggregate valuations are less extreme too, by one estimate around 22 times forward earnings versus 27 in March 2000, though concentration in the top names is higher now, so the risk is shaped differently rather than simply smaller.

Compare the infrastructure leaders. Cisco peaked near $550 billion in 2000 on a valuation built mostly on expectation, then fell about 80%. Nvidia's valuation is backed by actual earnings and demand that is, for now, real. There is also a physical brake the dot-com fiber boom never hit: power and grid capacity, which caps how fast the build-out can run ahead of demand. That is an inference, not a guarantee, but it is a real constraint 1999 lacked.

The dot-com lesson everyone gets wrong

People remember the dot-com crash as proof the internet was overhyped. That is backwards. The thesis was right. The internet did reorganize the entire economy, exactly as the bulls promised.

And roughly $5 trillion in value was still destroyed. The Nasdaq fell 78% from its March 2000 peak to October 2002. Pets.com, Webvan, and Boo.com died. Amazon fell more than 90% and nearly ran out of cash, then became Amazon. Only about half of dot-com companies survived to 2004. All of that happened at once. That is the part the "it's different this time" crowd skips: being right about the technology is not protection against a brutal repricing of the stocks.

So the lesson is not that AI is fake. It is that "the technology is real" and "the valuations are unsustainable" can both be true, and the second can wipe you out while the first slowly comes good for someone else. That is the title, restated: a bubble and a revolution are frequently the same event.

What this means if you're building AI products

This is the part that matters if you are an operator or founder rather than a trader, and it is where most bubble takes are useless to you. You cannot time the pop. Nobody rang a bell in March 2000. What you can control is whether you are built like a survivor or a casualty.

The survivors of 2000 had three things the casualties did not: real revenue from real customers, unit economics that actually worked, and enough discipline to outlast the funding winter. The casualties ran on narrative and someone else's capital.

Translate that to now. Charge real money and watch whether people actually pay, because that is your proof you sit outside the circular loop rather than inside it. Make your unit economics work at today's token prices, not the subsidized prices that exist only while compute is sold below cost to win the land grab. And do not build a company whose survival depends on the next round closing or the model staying cheap. If a 30% market correction would end you, you do not have a product, you have a position in someone else's bet.

There is a quieter risk too, separate from valuations. MIT's research found that about 95% of enterprise generative AI pilots produced no measurable profit, with integration, not model quality, as the cause. And as of mid-2026, most of the pure-play AI companies remain unprofitable even as their revenue grows fast. Even if the market never crashes, most AI products are failing on their own merits right now. That failure should occupy you more than the macro one.

The takeaway

Is AI a bubble? In parts, yes, by the only definition that means anything. Is the technology real? Also yes. The dot-com era already proved those two answers coexist, and that the coexistence is what makes the repricing dangerous: the truth of the thesis lulls you into ignoring the math on the price.

Stop asking whether it pops and when. Ask whether your product makes money at honest prices, with demand you can prove is real. Build that, and the bubble question stops being existential and becomes someone else's problem.