Пузырь AI: почему шампанское на Сен-Барте может скоро закончиться

Every time Silicon Valley starts talking about a new era where the old laws of economics no longer apply, somewhere in the shadows an experienced investor begins checking the reliability of his parachute. Right now we're observing a classic picture: while OpenAI spends billions training new models and startup unicorns multiply like mushrooms after rain, skeptics begin asking uncomfortable questions about profitability. The irony is that the history of technological manias is cyclical.

If you look at the British "canal mania" of the 1790s, you see the same symptoms we see today. Back then people believed a network of waterways would connect the world and create endless wealth. The canals were built, but most investors went bankrupt, while the cream was skimmed by those who came to the ruins of their ambitions.

Today's artificial intelligence industry is heterogeneous, and it's critically important to understand this. We can conditionally divide it into four layers. The first is "shovel sellers" or chip manufacturers like Nvidia. They've already made their billions and feel great. The second layer is cloud giants that rent out computing power. The third is creators of foundational models, such as OpenAI or Google. And the fourth is applied startups that build their products on top of others' APIs. The problem is that most of the money right now is burning at the third and fourth levels, where infrastructure costs often exceed revenue from users. This is a classic growth trap, where each new customer brings not profit but additional losses.

The situation is complicated by the existence of two parallel realities in the world of AI. The first reality is the corporate sector, where algorithms quietly and efficiently optimize logistics, find errors in code, and automate boring accounting. Here economics works, and companies are willing to pay for real efficiency. The second reality is "magical" consumer AI. These are chatbots that compose poetry and video generators requiring colossal computing power. This is where the biggest bubble is being inflated. Users are accustomed to free or very cheap tools, while the cost of one complex query remains high. If venture capital dries up before these models become orders of magnitude more efficient, we're in for a loud crash.

Investors in this race can also be divided into several camps. There are "true believers" convinced of the imminent arrival of AGI and willing to subsidize any losses for a bright future. There are those driven by pure fear of missing out—they jump into any round where the name contains the letters A and I.

And there are pragmatists who understand: when OpenAI or another major player faces a liquidity crisis, the market will begin a great fire sale of assets. It is at this moment that capital will be redistributed from dreamers to those who know how to count money. The bankruptcy of a major player won't kill the technology, but it will forever change the rules of the game, forcing the industry to mature.

What does this mean for us? Most likely we'll see a series of acquisitions and market consolidation. Small "wrappers" over GPT-4 will disappear because they have no foundation of their own. Those will survive who are embedded in real business processes, not just entertaining the public. Technological optimism is wonderful, but it doesn't replace positive cash flow. When the champagne in Saint-Barthélemy runs out and investors demand profit reports, we'll find out who was actually building the future and who was just blowing a beautiful bubble. The history of the dot-com crash taught us that Amazon and Google were born precisely from the ruins of a burst bubble. It will be the same with artificial intelligence.

Bottom line: Should we expect a crash tomorrow? Unlikely. But the era of "free money" for AI startups without a business model is officially coming to an end. Who will be the new Amazon in the world of neural networks?