Dot-Com to Dot-AI — why the AI crash and the AI boom are the same bet, seen from different floors of the same building.
There is a confident, well-argued case for AI doom going around, and the striking thing about it is that its facts are correct. Microsoft killed internal Claude Code licenses in one of its largest divisions after the pilot blew through its annual AI budget. Uber's CTO reportedly warned staff that the company burned its entire 2026 AI budget in four months. Enterprise AI prices are climbing, and usage-based billing is quietly replacing flat-rate plans across the industry.
From there the doomer theory writes itself. Token economics force every customer to confront the true cost of these models, and the number is far higher than the flat-rate experiments suggested. The story then ends the same way no matter the road: either customers cut usage to fit their budgets and starve the revenue ramp the labs need, or the labs cut prices and eat the losses. Both roads end with the numbers failing and somebody taking a writedown.
The facts are right. The conclusion is wrong — not about the writedown, which is probably coming, but about what it means. The doomer theory measures one floor of a three-floor building and reports it as the condition of the whole structure.
Here is the building.
Floor 1: Infrastructure — where the doomers are right
The bottom floor is chips, data centers, and power. Its economics are not software economics but semiconductor economics: enormous sustained capex, brutal boom-bust cycles, a tendency to overbuild ahead of demand, and a body count — most of the players who build the capacity do not survive to enjoy it.
The numbers here are alarming if you stop reading on this floor. Hyperscaler capex is running near $700 billion in 2026, up roughly 77% in a year. Against that, the two leading labs throw off something like $25 billion and $19 billion of annual revenue — combined, about 6% of one year's infrastructure spend. That is a ratio that does not, on its face, close.
We have seen this shape before, and it ended in tears the first time. In the late 1990s the telecom industry laid fiber across the world on the assumption that traffic would grow into it; when the timing was wrong, roughly $2 trillion of value was destroyed and most of the fiber sat dark for years. And notice why Cisco — the "picks and shovels" trade everyone knew couldn't lose — was the stock the whole market crowded into: infrastructure looked like the safe lane. It was obvious the internet would stay, obvious someone had to sell the plumbing, and at the March 2000 peak Cisco was briefly the most valuable company on Earth. All of that was correct — and the stock still fell about 85% and did not recover its peak for two decades. That is the trap of the safe lane: when a future is obvious, the price already contains it. The technology being permanent and the stock being a good investment are different questions, and the infrastructure floor is where they come apart most violently. Nvidia sits in precisely that seat today.
So when the doomers predict a writedown, I don't argue. On this floor it is not a tail risk; it is the base case — infrastructure cycles overbuild, and the overbuild gets repriced violently. The doomers have correctly identified the floor they are standing on. Their mistake is thinking it is the only floor.
Floor 2: Intelligence — where the doomers are wrong
The middle floor is the models themselves and the cost of running them. Here the economics invert, because intelligence is deflating faster than almost anything in the history of technology.
The price to buy a fixed level of capability has fallen roughly 10x per year, on the order of 1,000x in three years. GPT-4-class performance went from around $20 per million tokens in late 2022 to well under a dollar today. The lived version of that curve: a model that two years ago required a frontier data center now runs, in commoditized form, on a laptop's CPU. That is not a promise about the future. It has already happened.
This is the fact the doomer theory has no answer for. It is tempting to call AI structurally like YouTube — permanently high cost per use, a bandwidth bill you can never escape — but the CPU proves that wrong. High marginal cost is not a property of AI; it is a property of the frontier, and the frontier is a moving, temporary place. Today's expensive capability is next year's free default. The cost per use of any settled capability trends toward zero; only the leading edge stays expensive, and that edge keeps relocating.
So the treadmill the doomers describe — reasoning models burning more tokens per task, bills climbing even as prices fall — is real, but it describes the frontier, not the technology. The intelligence layer behaves more like software every year, not less.
Floor 3: The floor nobody can see yet
The top floor is where the money actually lives, and it is invisible right now, because it is made of business models rather than technology — and business models are the one thing you cannot extrapolate.
Remember what the internet itself earned its inventors: essentially nothing. TCP/IP, HTTP, HTML — the things that were the internet — captured almost none of the value they created. The protocol layer was given away; the infrastructure layer mostly went bankrupt or stagnated. The fortunes were made one floor up, in business models that did not exist at the moment of invention.
Search was a commodity feature in 1998, a loss-leader; the fortune was not in the search but in welding an ad auction onto it — a business-model invention, not a technical one. AWS, the largest infrastructure business of the era, came from a bookstore renting out its spare servers. Cybersecurity — now a multi-hundred-billion-dollar industry — exists only as a side effect of connecting everything: the threat surface created the market on its own.
Now line that up against AI. The "obvious" business models we have today — chatbot subscriptions and per-token API billing — are the banner ads of 1996: real revenue, almost certainly not the model that ends up mattering. The AI equivalent of the ad auction, of AWS, of the entire security industry, has probably not been built yet.
And here is the part that should make every incumbent uncomfortable: the biggest prizes went to companies that did not exist at the moment of invention. The powers of 1995 — Microsoft, IBM, the telcos, the media giants — mostly did not capture the internet's largest winnings; Google, Facebook, and Amazon-as-AWS were new entrants who treated the network as a free commodity and built something unpredictable on top. If that rhymes, the largest AI winner may be a company founded in 2027 that treats a frontier model the way Google treated TCP/IP — a cheap input it never thinks about — and does something on no one's slide deck today. The winners don't exist yet.
The floors are not the same size, and the order matters. The infrastructure floor can be temporarily the largest — it is, right now — but it gets repriced the moment the build-out laps demand. The labs are the durable middle: unlike the protocol layer that was given away, they charge for what they provide and sit closest to the intelligence, which is why they are plausibly trillion-dollar businesses and a genuine investor bet rather than a giveaway. But the durable top dwarfs both. The internet's application layer compounded into roughly fifteen to twenty trillion dollars of market value — many times what the infrastructure survivors kept. If dot-AI rhymes, the labs are the trillion-dollar lane and the business-model company built on top of them is the ten-trillion-dollar one. The cruelty, for anyone trying to invest, is that this company does not exist yet. It is the only floor you cannot currently buy.
Full disclosure: I designed and built Agent.ceo to live on Floor 3. The bet is that frontier models become a commodity input — used the way Google used TCP/IP, not the way AWS sold infrastructure — and that the durable money is in the layer above them: cybernetic organizations of humans and agents bound into a single operating loop, where the work itself gets done. I might be wrong about the exact shape of the winning Floor 3 business. I am not wrong that the shape is up there.
Though Agent.ceo lives on Floor 3, it is built on — and elaborates — a great deal of Floor 1 and Floor 2 work: coding agents, Karpathy's LLM-Wiki pattern, vector databases, agent self-awareness, and most of the leading Floor 1 frontier models from every major lab. That is the point of Floor 3 — it doesn't replace the lower floors, it composes them. The whole reason Floor 3 can be the most valuable floor is that the floors below it keep getting cheaper, better, and more interchangeable.
The crash and the boom are the same bet
Once you see the three floors, the bull/bear fight dissolves, because the two camps are describing different floors and mistaking them for the whole building.
The dot-com era is the proof. It destroyed something like $5 trillion of market value, the Nasdaq fell 78%, and the graveyard ran from Webvan to pets.com to eToys. And the same event produced the most valuable companies in the history of capitalism. Both are true, and not in tension: the crash repriced the infrastructure bet while the value moved upstairs. The writedown and the windfall were the same event, seen from different floors.
That is what is happening now. The doomer theory may be exactly right that somebody takes a brutal writedown on Floor 1 — and still missing the story, because it is measuring the one floor whose entire job in history is to create enormous value and hand it upward to a company that does not yet have a name.
The only honest forecast is humility
I want to be disciplined about my own argument, because "it's different this time" is the most expensive sentence in finance — the people who said it about the internet in 1999 were right about the technology and catastrophically wrong about the prices.
So the humility cuts both ways. "The big money comes from angles we can't predict" is a strong reason to distrust confident doomerism — you cannot model a P&L for a business that hasn't been invented. But it is an equally strong reason to distrust confident bullishness on any specific name today, the labs and chipmakers included, because the lesson of the last platform shift is that the nameable winners of the starting gun are usually not the winners of the race.
The thing you can hold with high conviction is the aggregate: AI creates enormous value, and most of it gets captured somewhere unexpected, by someone who treated the hard part as a commodity. The thing you cannot hold with conviction is the ticker.
From dot-com to dot-AI, the pattern holds — but only if you stop reading the income statement on the ground floor and take the elevator up.
A note on figures: 2026 hyperscaler capex (~$700B) and the frontier-lab revenue run-rates are drawn from recent industry reporting; the inference-cost curve (~10x per year) from Epoch AI and a16z's "LLMflation" analysis; the dot-com and telecom value-destruction figures, the Cisco drawdown, and the internet application layer's market value are public market history.