Fighter Jets, Pizza,
and the Fuel Bill.

It started the way most good conversations do: obliquely, in a group chat nominally about portfolios. Paul opened with AI and the end of the world, which is a reasonable place to start on a Wednesday afternoon if you are a Popular Investor with time on your hands and a low tolerance for unchallenged narratives.

The models work. The people building them say they are frightened of what they are building, and some of them have paid enough for that claim to be taken seriously. None of that is what the market is pricing. The market is pricing data centres, power contracts, and who owns the stack — which is a more boring story, and the one that still moves the money.

A hype machine does not volunteer extinction footnotes. It volunteers user growth numbers, demo reels, and a story about how everything changes now. Choosing to warn about civilisational risk when you are trying to price a public offering is the structural equivalent of standing up at your own wedding to announce a credible probability the marriage ends badly. Altman did something like that. He delayed the listing. He kept saying it. The cost of that position is observable and measurable, which is what makes it different from a tweet.

The test leaks at the edges because the cast does not share the same cost structure. Posting a terrifying AI forecast to damage a competitor costs a few minutes and an account in good standing. Delaying a listing costs real optionality every quarter — missed windows, agitated investors, governance questions with no clean answers. Both actions produce fear in the public discourse. Only one of them is evidence of a belief that had to pay something to exist.

Jensen's anti-doom talk is expensive too, just in the opposite direction. He sells the fuel and needs the buildout to continue. When two people pay real money to argue opposite cases, the argument is probably real.

Eight dollars for five throwaway comments on a trading platform. Write them yourself in three minutes. The tool saved ninety seconds and billed eight dollars for the privilege. At that unit rate you are not buying productivity — you are subsidising a research project that has not yet found what it is selling, at prices set before it knew what things cost to run.

The technology is not the problem. The technology genuinely does impressive things. The problem is that impressive things performed at fighter-jet infrastructure cost, for pizza-delivery-level tasks, produce the kind of unit economics that accountants mark in red and product managers quietly retire before the next board meeting.

The people who built the bomb became the public's primary interpreters of the bomb. Congress did not have its own physicists. The press relied on the people it was supposed to be covering. The race logic — if we slow down and they don't, we lose — was accepted as axiom rather than argued as proposition. Governments signed off on timetables set by the laboratories, because nobody else could read the timetables. All of that maps cleanly onto 2026.

Where the analogy runs out of road: the bomb was not a consumer subscription. Nobody cancelled after month three because the cost-per-explosion exceeded the entertainment value. AI is software — copyable, priceable, cancellable, refundable — and its economic reckoning runs in the same quarter as the existential one. Sora's revenue numbers and the lab safety reports were both published in 2026. The Senate committee and the cap table were reading different documents about the same companies at the same time.

The important correction arrived from Andrius without ceremony. The earlier analysis had used whole-company free cash flow as a proxy for the health of the AI business inside each company. That is the wrong measurement for the question being asked.

Microsoft generates substantial cash flows. Those flows come from Office, Azure, LinkedIn, Windows, and several other businesses that predate artificial intelligence as a serious commercial product. Using Microsoft's free cash flow to argue that its AI investment is working is like using a family's total income to argue that the garage renovation is profitable. The family can clearly afford the renovation. That is a different question from whether the renovation pays for itself.

The circular structure is the more serious concern. A major platform invests in a laboratory. The laboratory spends that money on compute, some of which it buys from the same platform. The platform books the revenue. The laboratory calls this progress. Private valuations rise. The same dollar has appeared in several columns without necessarily representing a paying customer at the end of the chain.

The relevant question for an investor is not whether Microsoft or Amazon or Google has enough cash to continue funding these experiments. They clearly do. The question is: at what point does the AI portion of the business earn its own fuel bill? At what price, for which tasks, paid by which customers who chose to pay because the value exceeded the cost?

Enterprise software and serious infrastructure can carry their own fuel bill. Consumer spectacle still cannot. The jet is extraordinary. The pizza routes keep getting cancelled.

Altman may be genuinely frightened of what he is building. Sora's compute bill still killed Sora. Both cleared the bar of evidence. The listing documents will eventually have to account for both in the same section, under oath, in front of people who have read the Sora revenue numbers.

The AI trade is carrying an industrial buildout story, a civilisational risk story, and a consumer product story that keeps failing on contact with actual prices — all three in the same vehicle, all three priced differently by different parts of the same market. Running only one of those threads gives you a clean narrative and the wrong answer.