Why There Is No AI Bubble

Let me concede the surface case up front. Valuations are stretched. The capex numbers are astronomical — hundreds of billions of dollars a year into data centers, chips, and power. Everyone with a Bloomberg terminal and a memory of 1999 is drawing the comparison, and the comparison draws itself. I'm not going to argue that nothing is overpriced, because I don't know, and neither does anyone else.

Here's what I am going to argue: the bubble debate is the wrong debate. Whether or not some AI assets are overpriced, the thing being built is real, and it will still be standing whatever prices do. The bubble-callers aren't wrong that there's froth. They're wrong about what froth means.

Two Definitions, and Why Only One Matters

When an economist says "bubble," they usually mean something specific: asset prices detached from fundamentals, sustained by the expectation of selling to someone at a higher price. By that yardstick, are parts of AI a bubble? Maybe. Some funding rounds will look silly in hindsight. Some companies are priced for a decade of flawless execution. That's been true of every technology buildout in history, including the ones nobody calls bubbles.

But notice what that definition measures: it tells you about investors. It tells you whether the people who bought at today's prices will get their money back. It tells you nothing about the question that actually matters, which is about the economy: if prices fell tomorrow, what would still be standing?

A hollow buildout leaves nothing. When the tide went out on Beanie Babies, there was nothing underneath — no capacity, no infrastructure, no compounding productivity. That's what "the investments will be for naught" has to mean for the bears to be right. And it's the claim that falls apart the moment you inventory what the money is actually buying.

The Two-Ledgers Lesson of the Dot-Com Crash

The bears' favorite analogy is their weakest evidence, and it's worth slowing down to see why.

The dot-com crash was the canonical bubble. The Nasdaq fell roughly 78% from its March 2000 peak to its October 2002 trough. Hundreds of internet companies died. Real people were genuinely ruined. If you bought the index at the top, you waited fifteen years to break even. Every terrible thing the bears predict for AI actually happened to the internet.

And the economy kept everything that mattered. The fiber laid during the boom — much of it dark for years, because the buildout ran far ahead of demand — got lit as demand caught up, and carried the next two decades of the internet economy. The protocols survived. The engineers trained during the boom built the next generation of companies. And Amazon — a public dot-com stock that fell from over $100 to around $6 in the crash — became one of the most valuable companies on earth.

The lesson isn't "crashes are fine." The lesson is that investor returns and economic capacity are different ledgers. The crash told you who overpaid. It didn't un-lay the fiber. Shareholders of specific companies lost money while the economy kept every dollar of the capacity they financed. The bubble-callers keep reading one ledger and drawing conclusions about the other.

So here's the uncomfortable arithmetic for the "for naught" crowd: if AI's worst case is the internet's worst case — an 80% drawdown followed by the infrastructure underwriting twenty years of growth — then the argument they're making was already tested once, at full scale, and it failed.

What the Money Is Actually Buying

Walk the capital stack from most durable to least, because this is where a skeptical reader should push hardest, and the answer holds up.

Power and land. A large share of AI capex isn't chips at all — it's power generation, grid interconnects, transmission, land, and data center shells. These are multi-decade assets that are useful under literally every future. If AI demand somehow evaporated, a grid interconnect doesn't care what the electrons are for. The United States needed this buildout anyway; AI is just the first customer willing to pay for it at speed.

Facilities and models. Cooling systems, specialized data center fit-outs, and the trained models themselves are more demand-dependent — a frontier model loses commercial value as better ones ship. But the know-how compounds. Every training run teaches the organization that ran it how to do the next one better, and that recipe knowledge doesn't depreciate with the weights.

People. The least discussed asset and possibly the most durable: an entire generation of engineers, researchers, and operators who now know how to build with this technology. Human capital doesn't show up on anyone's balance sheet, and it's the asset that made the post-dot-com internet possible.

The silicon. This is where the bears think they have their kill shot, so let's take it head-on. The implicit model is "GPUs are worthless in three years" — buy an accelerator today, write it off before the decade's half over, repeat until the music stops. The observable record says otherwise. Google moved its server depreciation to six years. Meta went to five and a half. A100s that shipped in 2020 remain commercially rented for inference and fine-tuning today, five-plus years on. The pattern is a cascade, not a cliff: frontier training migrates to the newest chips, and older generations absorb inference, fine-tuning, and batch workloads — work that gets more abundant, not less, as models improve.

Efficiency gains cut in favor of longevity, too, which surprises people. As models get more capable per unit of compute, hardware that was marginal for yesterday's frontier becomes ample for tomorrow's routine workload. The chip doesn't have to keep up with the frontier to keep earning. It just has to be worth its electricity.

And that's the honest concession, stated at its actual size: chips draw power, and when power is the binding constraint, the oldest generations retire before their physical lifespan because the compute-per-watt gap makes them uneconomic to run. Amazon shortened depreciation on a subset of its AI servers back to five years. Fine. That's normal capex churn at the margin — an operating decision about which assets to keep running, made by companies managing the largest fleets in history. It is not a hollow buildout, and it is nowhere near "worthless in three years."

The Revenue Is Real — and the Surplus Lands in the Economy

Here's the structural break from 1999 that the analogy-makers skip: the dot-coms had eyeballs and a story. AI has invoices.

Enterprises are paying real money for tokens because the tokens do real work. Developers are paying for tools out of their own pockets — the most honest demand signal in software. And I can speak to this one first-hand, for whatever one consultant's vantage point is worth: the way I build software has changed more in the past three years than in the fifteen before them. Projects that would have been a multi-person, multi-month engagement are now things I ship alone in weeks. That's not a projection on a slide. It's my invoices.

Now, the sophisticated bear has a comeback here, and it's worth taking seriously: real revenue doesn't prove the buildout earns an adequate return. Demand can grow while prices collapse faster than costs. Users can capture most of the benefit while the infrastructure owners fight over thin margins.

To which I'd say: yes — and that's the point. Surplus captured by users is the economic gain. If competition drives the price of intelligence toward its cost and the value lands with the businesses and people building on top of it, that's not the buildout failing. That's the buildout working exactly the way electricity did. The bears' scenario, played all the way out, describes a wealth transfer from infrastructure shareholders to the entire rest of the economy. You can call that a lot of things. "For naught" isn't one of them.

Nowhere Near the Ceiling

The buildout only counts as overbuilding if demand stalls, and every signal points the other way. Each time the cost of running these models has dropped, usage has grown sharply — cheaper intelligence keeps finding new consumption, the way cheaper compute always has. Most software still hasn't been rewritten. Most workflows still haven't been touched. The backlog of things that become economical at each new price point is decades deep, and much of that routine work runs happily on the installed base of older chips.

Capacity built ahead of demand is how big infrastructure waves tend to arrive. The railroads are the honest version of this precedent: massive overbuilding, waves of bankruptcies, fortunes destroyed — and the capacity that industrialized a continent. Both things were true at once. The bankruptcies and the payoff sat on different ledgers, which by now should sound familiar.

What the Bears Get Right

Some startups are wrappers with no moat. Some funding rounds will be case studies in euphoria. Some balance sheets will take real losses, and at some point prices may correct — call it a crash, a correction, a bubble popping, whatever headline writes itself that week. I'm not predicting it and I'm not ruling it out, because this piece doesn't need either. Failed bets are how durable assets find their right owners; that's what the failed bets are for. The equity vanishes. The assets don't.

Whatever the price path, the inventory doesn't move with it. The power plants don't care about the Nasdaq. The models don't unlearn. The engineers don't forget. And the productivity gains already compounding through the economy don't get clawed back because a multiple compressed.

The Wrong Scoreboard

The people waiting for a correction to feel vindicated are watching a scoreboard that doesn't measure the thing being built. When the word "bubble" means anything worth arguing about, it means the tide goes out and there's nothing underneath. The tide can do what it likes. The power plants, the data centers, the models, and the people who know how to use them aren't going anywhere — and the long-term bet is the simple one: a more capable, more efficient economy, running on capacity we're laying down right now.

The dot-com bubble popped. The internet didn't.

If you're trying to figure out what this buildout means for your own business — where AI actually pays off versus where it's still hype — that's the kind of question I work on. Get in touch if it's worth talking through.