I’ve been writing here on and off since 2007, and I think we’re living through the biggest tech boom I’ve seen in that time. Most people agree that AI is the most important thing to happen to computing in a generation, and I mostly agree. What I can’t reconcile is the arithmetic. There are two walls in front of this industry, and I’ve come to believe they’re really the same wall.
Wall one: the money#
In 2024, Sequoia’s David Cahn asked what he called AI’s $600B question. By his math, the data center spending needs about $600 billion a year in revenue to pay off, and even with a generous guess at AI revenue, he still found a $500 billion hole.
Then look at the flagship. Ed Zitron reported OpenAI’s 2025 numbers from audited financial documents, which the Financial Times independently verified. OpenAI had $13.07 billion in revenue and $34 billion in costs and expenses, and lost $20.92 billion on operations.
To be fair, the trend is going the right way. Revenue grew about 3.5x from 2024, and costs grew about 2.7x. The standard answer is “give it time.” Maybe. But growth only fixes it if the models keep getting much better and cheaper to run at the same time. Which brings me to the second wall.
Wall two: the math#
This one worries me more. Money problems can be fixed with more money. This one is about how the models work.
Today’s models are next-token predictors. Very good ones. I don’t fully understand how something trained to guess the next word can write working Go for me. But “I don’t understand how it works” is not the same as “it has no ceiling,” and I think the evidence for a ceiling is growing.
Yann LeCun has argued for years that next-token prediction is the wrong foundation. Every token is a sample with some chance of error, so over a long chain the errors compound. A 1% error per step over 100 steps (0.99^100 ≈ 0.37) leaves you only about a 37% chance of a clean run. He left Meta, and his new company, AMI Labs, raised $1.03 billion to build “world models” that learn from reality, not just from language.
The labs’ answer is reasoning models, which spend more compute thinking before they answer. It clearly helps. But when Apple’s researchers tested reasoning models on puzzles of rising difficulty in The Illusion of Thinking, the models’ thinking effort went up and then dropped off even with token budget left, and accuracy collapsed. A follow-up replication found some failures came from puzzles with no solution, but confirmed the models still stumble on Tower of Hanoi at around 8 disks. So I think reasoning may be papering over the architecture, buying real gains at a rising compute cost.
The best argument against me is that the error rate per step keeps dropping, and the same math gets friendly fast. At 0.1% error per step, 100 steps come out clean about 90% of the time. And METR found that the length of tasks AI agents can finish with 50% reliability has been doubling about every 7 months for six years. If that holds, compounding error shrinks on its own. I’m not sure it will hold, but the trend line isn’t on my side yet.
Why the two walls are one wall#
The bull case needs two things. Models keep getting better, and they keep getting cheaper to run. If either one stops, the $600 billion never shows up.
The cheaper part is real, but mostly for yesterday’s capability. A recent paper, The Price of Progress, found that the price of reaching a given benchmark score falls about 5x to 10x per year. But I think customers want the frontier, and the same paper found the price of running frontier models is rising 3x to 18x per year, because the models are bigger and reason longer.
So a given level of capability gets cheap fast, but the bill for the best capability keeps going up. That only pays off if each step up buys a big jump in what the models can do. If next-token prediction is near its ceiling, you spend more compute, on longer reasoning chains, for smaller gains. That’s where the two walls become one. The spending only makes sense with capability growth, and the architecture might not deliver it at a price anyone will pay.
But can’t you just sell shovels?#
The usual advice in a gold rush is to sell picks and shovels instead of mining. In this rush that’s Nvidia, and it has worked spectacularly. It isn’t wrong.
But shovel sellers only do well while miners keep buying. In the dot-com bust, Cisco and Lucent were hit hard when the telecom companies buying their gear collapsed. Nvidia’s revenue depends heavily on labs and cloud companies spending on the belief that the gold is real. If the capability jumps don’t arrive, the first thing cut is next year’s shovel order. So selling shovels is the same bet on whether the gold is real.
Where I land#
The best counterargument I know comes from Carlota Perez. Bubbles still build things. The railway mania and the dot-com fiber glut wiped out investors but left behind rails and fiber that powered the next era. By that logic, today’s data centers are our railroads, and it barely matters if the companies paying for them go bust. Paul Kedrosky lays out this view for AI spending and calls it “a compelling argument,” then writes, “There are at least two problems with this view, however.” I think the Perez view is mostly right, but it also admits most of these companies go bust.
So here’s my bet, for what it’s worth from someone who’s wrong about as often as anyone. The compute gets built and mostly stays. Most of the “AGI is imminent” companies don’t survive in their current form. The winners treat today’s models as a useful but limited tool, not a religion. I think that means three groups. First, B2B software companies putting AI into a narrow workflow they can check, with a human in the loop. Second, teams betting on smaller, specialized models instead of one all-knowing model. Third, whoever is working on the next architecture instead of only scaling this one. That last group matters because of the innovator’s dilemma. The incumbents are too committed to what’s working right now.
I came back to this blog partly to force myself to think these things through in public instead of in the shower. So tell me where I’m wrong. Will better models close the gap between the spending and the revenue, or are we going to keep paying more and more for smaller and smaller gains?
Update (September 28, 2026): Since I wrote this, the second-quarter numbers came out, and they cut both ways. Anthropic said it made $11.6 billion in revenue and its first operating profit, $559 million, which is the best evidence yet against my first wall. Those figures are self-reported and it didn’t say how it calculated the profit. OpenAI went the other way. Per the WSJ, its revenue grew about 18% to $6.7 billion while its operating loss grew about 32% to $12.3 billion (SiliconANGLE has both). So at least one lab has found a way to make the math work, for now. I’m watching whether that holds as the models get bigger.
Reply by Email