This spring, two things were true about Uber in the same week.
Roughly 70% of the code its engineers commit now originates with AI, and Uber burned through its entire 2026 AI budget — billions of dollars — in four months, then started capping what each employee could spend.
Read the first fact and AI is a miracle. Read the second and it's a bubble. Both are true. Most of us can't hold both concepts at the same time. That's the problem. Not the technology.
We have two settings for anything new: this changes everything, or this was over-hyped. Euphoria or dismissal, with almost nothing in between. Right now, the dial is swinging hard towards dismissal (Microsoft quietly winding down internal AI tooling, Gartner officially moving generative AI into what it calls the "trough of disillusionment," every other headline asking whether the whole thing was oversold).
We've been here before. Almost exactly. The bust has a mechanism, and it isn't disappointment.
The Bubble: Here's the part that makes the "bubble" question more than a mood right now, you are not paying the real price of AI. The providers are widely reported to be pricing inference below cost (OpenAI spent about $1.69 for every $1 it earned, almost all of it the cost of running the models). That's not a scandal; it's a strategy. It's the Uber playbook: subsidize the ride to build the habit, raise the price once the habit is locked in. Venture capital is quietly paying part of your AI bill.
The honest counterargument is real, and worth pointing per-token costs are falling fast (roughly $10 to $2.50 per million in a year, with another step-change in chips coming). But cheaper tokens haven't meant cheaper bills. Usage is rising faster than price is falling — an agent burns orders of magnitude more tokens than a chat box — so the invoice keeps climbing even as the per-unit chart drops. Efficiency is real. It just isn't rescuing anyone's unit economics yet.
And unlike Uber, there's no clean endgame. Uber could raise prices once it won the market. AI has four well-funded giants and a wave of open-weight models (nobody wins the field outright, so nobody gets to reprice freely as the prize). The pressure comes from a different clock: the IPO. The moment these companies face public markets, "growth at any cost" becomes "show me the margin," and the subsidy must thin. It already is — When Cursor moved to usage-based billing last summer, developers who'd built their entire workflow on a flat $20 plan watched their bills jump several times over (some reported 20×+). Same work. The subsidy just lifted. (Cursor apologized and refunded, but the direction of travel was set). You don't even have to leave this post for the pattern. The model many of us are building on, Claude's new Fable 5, is included on the Pro, Max, and Team plans at no extra cost… until June 22. On June 23 it comes off, and access runs on usage credits. Two weeks included. Then the meter starts. (Anthropic cites capacity and may fold it back in later — but the buyer's lesson is the same: build now, pay later.)
So a bust is coming, or at least a hard repricing. But here's the move most people miss: that is not a verdict on the technology. A correction in what AI costs is not the same as a correction in what AI does. Confusing the two is exactly the binary reflex this whole piece is about, and the last time we made that mistake, it was called the dot-com crash.
The mirror: Between 1995 and March 2000, the Nasdaq ran up more than 570%. The internet was going to change everything, and the market priced it in. Then it broke. The index fell 78% by October 2002. Around five trillion dollars in value evaporated. More than half of the public dot-com companies were gone within a few years. Pets.com went from a $300M valuation to zero in under a year. Amazon lost roughly 90% of its value, Jeff Bezos opened his 2000 shareholder letter with a single word: "ouch."
By 2002, the smart-sounding take was that the internet had been a fad dressed up as a revolution. That take has aged badly. The twenty years nobody puts on the slide.
Here's what happened after the bust: nothing — and then everything.
The internet didn't die. It went to work, quietly, while the hype was gone. It changed how we communicate, how we buy, how we bank, how we work. The same Amazon that dropped 90% built AWS and became one of the most valuable companies on earth. Google, eBay, Priceline — the survivors of the crash defined the next two decades. And a decade after that, the infrastructure they'd built is what made the entire SaaS boom possible.
The crash didn't disprove the internet. It cleared the field. It separated the companies riding a narrative from the ones quietly building something that worked.
Where AI is: Now look at the AI headlines again, but with that lens.
Microsoft pulling back on a coding tool isn't "AI failed." It's a company discovering that the unit economics need discipline — while it still writes a third of its own code with AI. Uber's budget blowing up isn't disillusionment; it's a rollout that moved faster than its financial controls. Their COO said the honest part out loud: the link between all that AI usage and real customer value isn't there yet.
Not there yet. That's not a eulogy. That's a phase. Gartner still expects the world to spend $2.5 trillion on AI this year. The money isn't leaving. It's being repriced — moving from "spend on everything because it's magic" to "spend on what actually works."
Who wins from here: I'll give you the number that matters most, because I built a post around it early this week: only about 5% of enterprise GenAI pilots deliver measurable impact on the bottom line. The euphoric crowd hears that and panics. The cynics hear it and feel vindicated.
Operators hear it differently. We hear: there's a 5% who figured out the discipline — and the trough is when you join them.
Because this is the pattern the dot-com era proved. The boom rewards everyone, even the companies with no business model. The bust punishes everyone, even the good ones, for a while. But the long climb afterward only rewards the people who kept building.
AI will have its bust. It may already be in it, a brutal repricing, a long and un-glamorous adoption phase, and then a wave of re-imagined deployments we can't fully picture yet — the same way nobody in 2002 pictured the iPhone, the cloud, or the creator economy.
And then it will change everything. Not because it's magic. Not despite being over-hyped. But because that is what real platform shifts do — on a timeline that's always slower than the believers want and far longer than the skeptics expect.
The dial only has two settings. The whole opportunity is in refusing to use it.
I build operational systems for companies navigating exactly this line — adopting AI without betting the company on the hype, or missing the shift by sitting it out. If that's the tightrope you're walking, let's talk.