Companies that have invested $800 billion in building data centers for artificial intelligence risk not recouping these funds. The market is confusing two fundamentally different products: mass-market AI services, on which it is extremely difficult to make money, and scientific breakthroughs, which cannot be kept in one pair of hands.

This contradiction is precisely what underpins the key argument for revising valuations of AI infrastructure. Proponents of large-scale investment in computing power often demonstrate a superficial understanding of the industry's economics, accumulating hundreds of billions of dollars in debt "on vibes."

What data centers earn money on

Data center capacity is spent on two tasks. The first is training the model. The second is processing user requests once the model is ready. It is inference that brings in the main revenue: millions of people use it, and demand seems endless.

But this part has a structural problem. Users arrive in waves, and server load fluctuates. To keep everything from breaking during peak hours, capacity has to be built with a large reserve. As a result, there are now more data centers than are actually needed.

With training there is no such problem: the load is steady, and the needs are known in advance. Anthropic, OpenAI, and other developers promise to load surplus capacity with training new models. But their goal is not a convenient service for work, but breakthroughs: new medicines, cyberweapons, self-improving models. When Anthropic talks about a "threat to humanity," it is not about job losses, but a hint at the proximity of major discoveries.

Why breakthroughs don't make money

The main weakness of this plan is that any model is quickly copied. A small model can be trained on the answers of a large one, and it will adopt its skills. In addition, many models are released publicly. As a result, even the most advanced development becomes obsolete within a few months. A patent protects an invention, but nothing protects a model.

The analogy with pharmaceuticals is telling. A company spent billions on a drug for a rare disease, and two months later a cheap generic appeared. It will not be able to recoup the money, unless it keeps finding such drugs one after another endlessly. It is precisely this scenario that the $800 billion is betting on.

Hence two questions for AI investors:

  • how much ordinary AI services actually cost and how much the path to breakthroughs costs;
  • whether the investments account for the risk that the model will lose value before it pays for itself.

The conclusion is grim. Ordinary AI services are needed by everyone, but competition drives prices down, and there is not much money to be made on them. Breakthrough discoveries cannot be sold en masse. And if they do happen, no private company will be able to hold onto them without the state's help. That is why Anthropic keeps pressing on the risks theme, although it has not shown that it is close to such discoveries. The question remains who will ultimately lose the money.

Cryptalist analysis: The AI infrastructure market is pricing in the simultaneous realization of two mutually exclusive scenarios — mass monetization of services and retention of a technological advantage. Both assumptions are structurally vulnerable: competition and open models quickly reduce the value of developments to zero. For institutional investors, this is a signal to reassess the risks of debt financing for data centers and to take a cautious approach to assets tied to the payback of AI capacity.