The more I dig into the economics, the harder it is to see AI data centers as a good business, and they’re now my leading candidate for what pops the AI bubble in the next 6 to 12 months. The concern isn’t that AI stops improving or that demand vanishes. It’s that spending has raced far ahead of proven revenue while the assets being financed lose value unusually fast. OpenAI reportedly burns roughly $60 billion a year on compute against about $13 billion in revenue, and the five biggest cloud providers are on track to spend something like $725 billion on AI capex in 2026 alone. Those numbers require more than strong AI adoption. They require an enormous new profit pool to emerge quickly enough to support the assets already being built.
The clearest tell that AI data centers may not be a great business is that the two companies built entirely around them (CoreWeave and Nebius) are still spending heavily to scale. CoreWeave booked a $1.17 billion net loss for 2025 and a further $740 million loss in Q1 2026 alone. Nebius posted a $100 million adjusted net loss in Q1: its reported “profit” coming almost entirely from a one-time gain on revaluing equity investments rather than from the business itself. Which makes you wonder what SpaceX sees that the rest of us do not, given that it now wants to run the same capital-hungry, low-return model in orbit.
The reality is that renting GPU capacity is a capital-heavy, low-margin commodity service facing falling prices as chips improve, models get more efficient, and competition expands. The owners of the buildings and the chips don’t capture the value AI software creates. They build the roads but don’t collect the tolls. Layer on decades-long debt against two-to-three-year hardware, off-balance-sheet structures, junk-rated balance sheets, and circular deals where chipmakers fund the startups that buy their chips, and you get a structure that only holds together if spending keeps climbing. In addition, much of the apparent demand is also concentrated among a small number of heavily funded but unprofitable AI companies. AI usage can grow rapidly while the companies supplying its infrastructure still earn poor returns.
That’s what makes the next phase dangerous. A correction doesn’t require AI to fail. It just needs the spending to stop rising. One hyperscaler trimming GPU orders by 20 to 30 percent would cascade through chipmakers, power developers, construction firms, and the debt-laden neo-clouds. Write-downs then tighten credit, which cuts spending, which drives more write-downs. The Fed has already flagged AI as a top systemic risk, and history is full of world-changing technologies, from railroads to fiber, that still wiped out the investors who funded them. AI can transform the economy and still destroy much of the capital poured into building it.
Related content:
Thanks for being a regular reader — consider becoming a paid supporter 🙏
