Smarter AI may not be more profitable
Nearly $3 trillion planned for AI infrastructure through 2028, with one-third financed by debt, creating systemic risk if returns don't materialize quickly. AI companies promise superintelligence and revolutionary productivity gains, yet remain unprofitable and show only incremental improvements; GPT-5 disappointed users expecting quantum leaps.
- Nearly $3 trillion planned for AI infrastructure through 2028, with one-third financed by debt
- OpenAI and partners announced $500 billion in new U.S. infrastructure spending
- OpenAI does not expect profits before 2030; Anthropic accumulates losses
- Five companies control most of the world's computing capacity and capital
- GPT-5 was incremental improvement, not the quantum leap users expected
As 2026 begins, AI faces a paradox: record investment and ambitious promises clash with market impatience and doubts about justifying massive capital. Experts warn of potential bubble risk while acknowledging transformative potential remains uncertain.
We begin 2026 in a peculiar bind. Never before has so much money flowed into artificial intelligence. Never have the promises been so sweeping. Yet the markets have grown restless, and doubt creeps in about whether the sector can actually deliver returns that justify the avalanche of capital pouring in.
The paradox cuts both ways. Investors and governments fear missing the biggest technological shift since the internet arrived. But they also wonder whether the valuations rest on nothing but expectations that grow harder to meet each quarter. The fear of being left out of a genuine revolution—the possibility that someone will crack artificial general intelligence and leave everyone else behind—keeps the money flowing anyway. Better to overpay than to miss it entirely.
The numbers tell part of the story. Through 2028, technology companies and governments plan to spend nearly three trillion dollars on data centers and computing capacity. A third of that will come from borrowed money. OpenAI and its partners alone have announced five hundred billion dollars in new infrastructure spending in the United States—enough to fund the Apollo program twice over. This is capital sufficient to reshape the global economy. It is also capital sufficient to cause systemic damage if the returns do not arrive on schedule.
Meanwhile, the promises keep accelerating past the actual results. Sam Altman at OpenAI speaks of superintelligence as imminent, moving from warnings about his creation's hallucinations to claiming he could not raise his child without ChatGPT. Dario Amodei at Anthropic predicts that by 2027 we will see systems smarter than Nobel laureates, capable of curing diseases humanity has carried for centuries. Elon Musk promises that AI will eliminate unemployment and poverty, and that the world will soon fill with humanoid robots to do our work—a claim he has made for a decade without result, yet it keeps investor confidence alive, even as they eye him with growing skepticism. But when you look at the present rather than the future, the picture blurs. OpenAI does not expect profits before 2030. Anthropic accumulates losses. Even companies using AI intensively have failed to demonstrate clear improvements to their bottom line, according to consulting reports. Productivity rises in firms that deploy AI, though how you measure it matters. The same applies to creativity—the research shows it can flourish or wither depending on how the AI is used. The sector survives on the promise that someday, somehow, AI will generate enough wealth to justify the investment.
Google's CEO Sundar Pichai, not typically one to dampen enthusiasm, admitted recently that demand for computing capacity has entered a phase of "irrationality." He warned that if the bubble bursts, it will drag the entire sector down with it. The concern echoes 1996, when Alan Greenspan first spoke of "irrational exuberance." He was right about the bubble, but three years too early—those who listened lost money while markets climbed higher. The difference now is scale and concentration. The debt tied to AI investment is vastly larger. Five companies control most of the world's capital and computing power. If one stumbles, the domino effect will be immediate.
But there are risks beyond the financial. These systems do not operate neutrally. They optimize for specific goals. Social media algorithms are designed to maximize time spent on the platform, not to inform users better or protect their mental health. The result has been radicalization, misinformation, and psychological harm. Conversational chatbots present a different problem. Millions of people interact with them as if they were reliable sources of knowledge. Yet these systems are not designed to tell the truth—no one has figured out how to program truth into them. They are trained to generate plausible responses, not necessarily correct ones. They cannot say "I don't know." Researchers built them to be increasingly intelligent in a general sense, but what emerged was not artificial general intelligence—it was a very convincing simulation. The problem is precisely that it seems to understand what it says. People grant it an authority it does not possess. The metaphor of the stochastic brother-in-law captures it: like a parrot, the chatbot repeats without comprehension; like a brother-in-law, it speaks with secondhand confidence that persuades. It has read nearly everything and remembers it all, allowing it to construct coherent discourse. That does not mean it understands the content. Its operation rests on sophisticated statistical patterns that capture relationships between words and concepts, not on human understanding.
There is also the growing possibility, increasingly mentioned among researchers, that AI progress will not continue at the breakneck pace everyone expects. GPT-1 in late 2022 astonished the public and experts alike. Yet GPT-5, released last summer, was an incremental improvement, far from the quantum leap some had assumed. Many users complained they missed GPT-4. They had bonded with their AI, and the new version felt colder, more distant. A warning that smarter AI may not be more profitable. As 2026 unfolds, the year will test whether the sector can justify its size. If current models open new business lines and deliver tangible productivity gains, the investment may hold. But market patience is finite, and the pressure is mounting. A correction would not stop at AI companies—it would ripple through computing, energy, and telecommunications. Meanwhile, while debate rages over whether superintelligence could dominate humanity, less attention falls on how these tools already shape daily life, especially among the young. One of the central risks ahead is treating AI as an oracle. Another is having placed a general-purpose tool in the hands of millions without knowing in advance how it would be used. The launch of major chatbots was, in that sense, a global-scale experiment. It has revealed surprising applications and unforeseen consequences. More people now use ChatGPT-style systems as therapists and voting advisors. The algorithm does not understand the content it recommends, but it learns which videos and messages generate engagement. It can promote conspiracy theories without knowing what they are—only that they work. That does not make it neutral, especially when the companies that own it are not.
Citas Notables
Demand for computing capacity has entered a phase of irrationality, and if the bubble explodes, it will drag the entire sector down— Sundar Pichai, CEO of Google
These systems are trained to generate plausible responses, not necessarily correct ones. They cannot say 'I don't know.'— Analysis from the source material
La Conversación del Hearth Otra perspectiva de la historia
Why does it matter that we're spending three trillion dollars on AI infrastructure if we're not even sure it works yet?
Because that money has to come from somewhere. A third of it is borrowed. If the returns don't materialize, you don't just have disappointed investors—you have a debt crisis that spreads through the entire economy. Banks, pension funds, energy companies, telecom networks—they're all connected to this bet.
But couldn't the technology actually deliver? Couldn't superintelligence be real?
It could be. That's the trap. No one wants to be the person who bet against the next internet. So everyone keeps betting, even as the actual improvements slow down. GPT-5 disappointed people. It was smarter but colder. That tells you something about the gap between what's promised and what arrives.
What's the stochastic brother-in-law thing about?
It's a way of saying these systems sound confident but don't actually understand. They've learned patterns from everything humans have written, so they can construct sentences that sound true. But they're not reasoning. They're pattern-matching at scale. The danger is that people treat them like oracles.
So the real risk isn't that AI becomes superintelligent—it's that we treat it as if it already is?
Partly. But there's also the immediate problem: these tools are already influencing how people think, what they vote for, what they believe. We launched them without knowing what would happen. Now we're finding out.
What happens if the bubble bursts?
Markets correct, valuations fall, some companies fail. But the ripple goes everywhere—energy, telecom, banking. And we're left with questions about what actually works and what was just hype. The hard part is that some of it probably does work. We just don't know which parts yet.
Is there a way out of this paradox?
Not really. You either get returns that justify the investment, or you don't. If you do, you have to ask whether superintelligence is actually good news. If you don't, you have a financial crisis. The year ahead will tell us which problem we actually have.