The business of the end of the world: why AI companies cannot afford to be afraid

Research series from Laboratorio del Futuro

I. The arithmetic of the abyss

There is a sentence Dario Amodei spoke on a podcast in February 2026 that ought to be engraved on the door of every boardroom in Silicon Valley. He was talking about the decision of how many data centers to buy, how much computing capacity to reserve for the years ahead. And he said: "If I buy a trillion dollars of compute a year for 2027 because I think I'm going to make a trillion dollars in revenue, and I actually make 800 billion, there is no force on Earth, no hedge on Earth that could keep me from going bankrupt".

The sentence is extraordinary for several reasons. First, because it comes from the CEO of a company that, by its own reckoning, may be building the most dangerous system humanity has ever created. Second, because he says it without irony: he is not describing a remote hypothesis, but the everyday risk of his business. And third, because it reveals something the earlier articles in this series have been hinting at and that now becomes explicit: AI companies cannot afford the luxury of being afraid. Not because they don't feel fear, but because fear, in their cost structure, is an expense they cannot take on.

This article is about that: the political economy of fear. About why the very companies that warn of an existential risk are trapped in a logic that keeps them from slowing down. And about why that trap is not a moral flaw in their leaders, but a structural property of the industry they have built.

II. The number that changes everything

Let's start with the numbers, because they are what make the trap intelligible.

OpenAI, according to financial documents reviewed by analysts, projects annual losses of roughly 14 billion dollars in 2026, almost triple its losses two years earlier. Its infrastructure commitments exceed a trillion dollars through 2035. Anthropic, for its part, could burn close to 20 billion before reaching profitability. And the most pessimistic projections suggest OpenAI could pile up 85 billion in losses before the business becomes sustainable.

These are not the figures of a company failing for lack of demand. They are the figures of a company growing at an extraordinary pace and, at the same time, spending more than it takes in. The problem is not that nobody wants its products. The problem is that every new user, every query, every token processed costs more than the user pays.

The industry has called this the problem of "diseconomies of scale". In the twentieth-century internet business model, adding a user was almost free: the marginal cost tended toward zero, and scale was the road to profitability. In AI, the marginal cost of every query is real, tangible and rising. An executive at a robotics company put it bluntly: his engineers burn between 300 and 500 dollars in compute in a single afternoon, more than their daily salary.

The result is a paradox no business textbook had foreseen: the product works so well that it ruins whoever sells it. The more it is used, the more money is lost. And the more money is lost, the more the company has to grow to justify the investments that allow it to keep growing.

III. Amodei's timing trap

Amodei has explained this trap with a clarity that deserves to be quoted at length, because it is the key to the whole article.

His logic goes like this. Anthropic is growing at roughly ten times a year. In early 2026, it was bringing in about 10 billion on an annualized basis. If that rate holds, by the end of 2027 it will be bringing in a trillion. To meet that demand, it has to reserve data centers one or two years in advance. In other words: it has to decide now how much capacity to buy for two years from now.

If it buys the capacity needed for a trillion in revenue and the revenue arrives, it wins. If it buys and revenue stalls at 800 billion, it goes bankrupt. If it buys less and revenue reaches a trillion, it loses customers for lack of capacity. And if the growth rate is not ten times but five, it also goes bankrupt.

"If I'm off by one year on that growth rate," Amodei said, "I go bankrupt".

This is the trap. It is not a moral trap. It is an arithmetic trap. The uncertainty about when the revenue will arrive is so great that any investment decision is an existential bet. And the only way to reduce the risk is for every competitor to slow down at the same pace. But nobody can verify that the others are slowing down. And whoever slows down first loses.

Amodei criticized his competitors for what he called "YOLOing" on spending: "I get the impression that some of the other companies haven't written down the spreadsheet, that they don't really understand the risks they're taking. They're just doing stuff because it sounds cool". The criticism is fair. But it is also revealing: Amodei, who has written the spreadsheet, has reached the same conclusion as those who haven't. The difference is not one of prudence. It is one of style. Everyone is buying compute. Everyone is betting. The only difference is that some do it with a spreadsheet in front of them and others without one.

IV. Debt as anesthesia

There is a second layer to this trap that makes fear even harder to sustain: financing.

Big Tech no longer funds its AI investments out of the profits of its traditional businesses. According to Bloomberg, the four largest AI companies have turned this year to debt issuance and special purpose vehicles to finance 745 billion dollars in infrastructure investment. S&P Global estimates that the six largest U.S. hyperscalers will spend more than seven trillion dollars on data centers and other AI-related investments between 2025 and 2030, and will have negative free cash flow at least until the following year.

S&P Global analyst Mark Habib was explicit: "It's safe to assume there's more debt out there than what we're seeing". And he added: "If it's not reported, then it's not part of the hyperscalers' disclosure. And if we're not rating the transactions where this liability can build up, then the fact is we might not have visibility".

Debt is the anesthesia of fear. As long as the money keeps flowing, the risk is deferred. But anesthesia has a limit. CoreWeave, the Nvidia-backed cloud services company, has become the symbol of that limit: its bonds maturing in 2032 traded below 80% of face value, with yields above 15%, and its stock fell 15% in a single day. Investors, who for years had accepted the narrative of infinite growth, began asking when the profits would arrive.

The answer is that nobody knows. Not even Amodei, who has built his career on the prediction that AI will transform the world. When asked when the trillions in revenue would start to arrive, he said: "I don't think it's guaranteed that it's immediate. It could be one year, it could be two, I could even stretch it to five, although I'm skeptical of that".

Five years. In an industry where growth is measured in months and valuations depend on the promise of an imminent future, five years is an eternity.

V. The circularity of optimism

There is a third layer that makes fear almost impossible to sustain: the circularity of the investments.

Goldman Sachs noted in a recent analysis that Big Tech recorded roughly 150 billion dollars in unrealized gains on investments in private companies in the second quarter of 2026, equivalent to 12% of S&P 500 earnings. They are "marks", not cash. But they count as earnings.

The circularity works like this: the hyperscalers invest in AI labs. The labs use that money to buy compute from the hyperscalers. And the stakes are revalued with every new funding round. It is a cycle that feeds itself, and it works as long as the rounds keep going up. But a mark can be reversed as quickly as it was created. And 12% of the earnings of the most important index in the world depends on those rounds continuing to go well.

The Financial Times described it with a metaphor that sums up the entire article: "The AI trade wobbled significantly... it was a kind of dress rehearsal for what the bursting of the bubble could look like". The likely cause will not be cold economics, nor a blown fuse in the financial wiring, but vibes. The perception that something can go wrong. And that perception, once it sets in, cannot be uninstalled with spreadsheets.

VI. What fear does to safety

This brings us to the point where economics meets safety, and where the previous article in this series takes on its true meaning.

If AI companies are trapped in a logic that demands they grow at any cost, how much room do they have to invest in safety? The honest answer is: very little. Safety does not generate revenue. It does not accelerate growth. It does not justify valuations. It is a cost, and in an industry where every dollar counts toward survival, costs are the first thing to be cut when growth slows down.

The White House accord we analyzed in the previous article set out four layers of controls and audits. But it did not require the results to be made public, or the auditors to have the power to halt a deployment, or the companies to disclose their incidents. It was, in essence, a voluntary self-regulation agreement in an industry that, according to its own CEOs, is on the brink of bankruptcy if growth slows down even a little.

The question that arises is unavoidable: if growth slows down, what gets cut first? The investment in safety, or the investment in capacity? The answer, in a company competing to survive, is obvious. Safety is a luxury only those with a margin can afford. And in this industry, the margin does not exist.

VII. Closing: fear as a cost

Amodei said that if he is off by one year in his growth forecast, he goes bankrupt. The sentence is a confession. Not of incompetence, but of powerlessness. The CEO of one of the most important companies in the world, in one of the most transformative technologies in history, is saying in public that he has no margin for error. That his company cannot afford the luxury of slowing down, or of investing less, or of being wrong.

And if he cannot afford the luxury of slowing down, he cannot afford the luxury of being afraid either. Because fear, in his logic, is a cost he cannot take on. Fear leads to caution. Caution leads to slowing down. And slowing down, in an industry that grows tenfold a year or dies, is death.

This is the trap the previous article in the series described from the outside and that this one describes from the inside. It is not that AI companies are evil. It is that they are trapped in a structure that demands they grow at any cost and leaves them no room for prudence. The fear their leaders express in public is sincere. But it is a fear they cannot afford to act on.

The broken mirror of this series reflects, in this article, something more than AI. It reflects an industry that has built a machine it cannot stop, not because it does not want to, but because stopping it would mean being crushed by those who don't. And it reflects a society that has decided, without saying so, that growth matters more than safety.

Amodei said that if he is off by one year, he goes bankrupt. What he did not say is that if he is not off, he can go bankrupt too. Only, in that case, the bankruptcy would be of something else.

Bibliography

• Investing.com (2026). "AI Cash Burn Rates: The Cost of Building the Future". (Cash burn projections for Anthropic and OpenAI; OpenAI's annual losses in 2026)
• Data Center Dynamics (2026). "Anthropic CEO on AI data center spend: 'If you're off by a couple years, it can be ruinous'". (Amodei's remarks on the risk of bankruptcy from forecasting errors; his criticism of competitors' "YOLOing")
• Fortune (2026). "Anthropic CEO Dario Amodei explains his spending caution, warning if AI growth forecasts are off by just a year, 'then you go bankrupt'". (The detailed explanation of the timing trap; the spreadsheet logic; the criticism of rivals)
• 证券时报 (2026). "用户越多亏得越多 AI行业求解'规模不经济'难题". (The problem of diseconomies of scale; the example of engineers burning 300-500 dollars in an afternoon; the drop in the token spending index)
• Bloomberg (2026). "AI大手、なぜ借り入れ急増？ データセンター投資が変わる財務戦略". (Big Tech's shift in financing strategy toward debt)
• S&P Global / CNBC TV18 (2026). "Top hyperscalers likely to have more AI-linked debt than now known, S&P Global warns". (The warning about hidden debt; interconnected financing structures; the estimated 7 trillion in spending through 2030)
• CoreWeave / 观察者网 (2026). "CoreWeave拟将保险业万亿美元资金池引入AI算力融资体系". (The fall in CoreWeave's bonds; the yield above 15%; the debt rating)
• eWeek (2026). "OpenAI's $14B Loss in 2026 Won't Slow IPO Push". (OpenAI's projected losses; the trillion-dollar infrastructure commitments)
• Exame (2026). "OpenAI pode acumular US$ 85 bilhões em prejuízo na corrida da IA". (Projections of accumulated losses; spending on compute infrastructure)
• Investing.com Australia (2026). "AI Capex and Earnings: What Happens When Growth Slows". (Big Tech's unrealized gains; the circularity of investments; the 12% of S&P 500 earnings)
• FT中文网 (2026). "The dress rehearsal for an AI reckoning". (The "dress rehearsal" for the bubble bursting; the role of "vibes" in the fall; the Broadcom CEO's warning about the "casino")