Translator's note: This article by Michael Roberts, an economist on artificial intelligence (AI) and productive artificial intelligence (GenAI) that have been storming, draws attention to an important point: AI does not increase worker productivity and this area is still far from meeting the profitability expectations of the investing capitalists.
Therefore, artificial intelligence, which stands out with the drag of a few Big Technology companies, cannot go beyond being a financial bubble for now. Moreover, for example, if Nvidia stumbles a little, the probability of this balloon bursting is quite high. Another article we planned to translate soon points out that AI is only one thing that currently works: Soft surveillance and control; a gradual and deep political pressure...
The square brackets in the text belong to the translator.
Artificial intelligence that bubbles
Michael Roberts
The Next Recession
July 27, 2025
The Great Seven shares – NVIDIA, Microsoft, Alphabet (Google), Apple, Meta, Tesla and Amazon – currently account for about 35% of the U.S. stock market value and NVIDIA's market value is about 19% of the Amazing Seven. The S&P 500 has never been so concentrated in a single stock, with Nvidia accounting for about 8% of the index.
This is a highly unstable exchange, driven by Nvidia, which is currently at record highs, producing only seven stocks and all the processors that AI [artificial intelligence] companies need to develop their models in particular. If Nvidia's revenue growth weakens, this will create a major downward pressure on the over-valued stock market. As Torsten Slok, chief economist on one of the largest investment institutions, said: “The difference between the IT bubble in the 1990s and the AI bubble today is that the top 10 companies in the S&P 500 are valued more today than they did in the 1990s.”
So, is the giant AI sector a giant bubble funded by imaginary capital that cannot be realized with the incomes of AI leaders and, more importantly, their profits? By the end of this year, Meta, Amazon, Microsoft, Google and Tesla will spend over $560 billion in capital on AI over the past two years, but will generate only $35 billion in revenue. Amazon plans to make a capital expenditure of $105 billion this year, but it will only generate $5 billion in revenue, and it's not profit because revenues are measured before the cost of AI services is deducted. Investments in AI will reach capital expenditure of 332 billion dollars in 2025 for revenue of only 28.7 billion dollars. Investments in huge data centers, which are necessary to train and source AI models, are planned to reach $1 trillion by the end of the decade.
But if any of the Magnificent Seven starts shying away from their spending in relation to their income and profits and reduce their chip purchases, the stock price of Nvidia could fall rapidly and lead others after them.
Will the expected return of this huge capital investment come true? Jim Covello, Goldman Sachs' head of stock research, questions whether companies that plan to invest $1 trillion in artificial intelligence development can get their worth of this money. On the other hand, a partner of venture capital firm Sequoia estimates that technology companies need to generate $600 billion in extra revenue justifying extra capital expenditures in 2024 alone. This figure is about six times the revenue that companies can earn.
Let's take the well-known ChatGPT. Allegedly, it has 500 million weekly active users, but according to the latest count, it only has 15.5 million paid subscribers, which means a return rate of only 3%. According to a survey conducted by Menlo Ventures with 5,000 American adults, the number of people using AI chatbots is increasing, while the number of people paying for the AI service they use is very small and their annual income is about 12 billion dollars.
The situation is even worse when it comes to profits from AI. Annual profit gains from major tech companies have been steady or slowing over the past few quarters, and are expected to slow further in 2025 and 2026.
So while the excitement of AI takes the stock market to new heights, a huge investment of money and resources, astronomical payments to AI trainers, and the construction of huge data centers [there]; but so far no significant revenue has been generated and there is almost no profit. This is a steroid-friendly version of the dot.com bubble.
However, no matter how much ballooning it is, this does not mean that a new “revolutionary” technology will not emerge, which will eventually radically change the productivity limits of large economies, thus bringing in a new era of growth. The Dot.com bubble exploded in a huge drop in the stock market in 2000, but the internet spread to all business sectors and all households and the Magnificent Seven appeared.
Let's give another example from the 19th century. In the 1840s, the Railway Mania occurred while numerous companies raised funds to invest in railway line construction in England. Railway shares soamed and stock prices doubled in 18 months from the beginning of 1843. But after the balloon, the flood came in 1845, many companies went bankrupt and stock prices fell by half. This caused a widespread financial crisis and a decline in production. However, railways were built, transportation costs fell sharply, and consumers' demand for travel has increased dramatically. England experienced an economic boom in the 1850s.
Will the AI bubble lead to financial collapse and crisis following the same path, but will it eventually form the basis of a new growth in productivity? In my previous articles about AI, I conveyed the doubts of experts such as Nobel laureate Daron Acemoglu and others about the benefits of AI in terms of efficiency. In addition, the OECD's recent in-depth report on productivity growth in major economies includes an assessment of the impact of the internet on productivity growth in the last 25 years.
The OECD report says: “In the last half century, we have filled offices and pockets with increasingly faster computers, but the increase in labor productivity in developed economies has declined from about 2% annually in the 1990s to 0.8% in the last decade. Even the production per worker of China, which once increased rapidly, has stopped". Research efficiency has decreased. Today, the average scientist produces less groundbreaking ideas per dollar than his colleagues in the 1960s.
Labor productivity growth has been in a downward trend across the OECD since the 1970s and has weakened since the beginning of the century. Productivity in the U.S. increased from the mid-1990s to the mid-2000s, with increased productivity in the production of ICT [Information and Communication Technologies] equipment and the spread of internet-related innovations adopted in industries using ICT, especially in the retail sector. “For the other hand, this recovery was relatively short, and since then, the increase in efficiency has been poor.”
The main factor of increasing labor productivity is investing in new technologies that save labor. But business investments in all countries have slowed down significantly. The OECD makes it clear why: “Despite easy and cheap credit opportunities for firms with access to capital markets, it is in line with historical trends that investments have slowed down, uncertainty and expected profits are in line with historical trends pointing out that expected play a more important role in investment decisions than financial conditions.” In other words, the profitability of capital has decreased, which has reduced the incentive to invest in new technologies.
And so-called “intangible assets” such as software investment could not compensate for the decline in investments in facilities, equipment, etc.: “Despite the rise of intangible assets, total investment has generally been weak since the global financial crisis, which has directly worsened the slowdown in labor productivity.”
Will AI be different? Can companies replace millions of workers in the economy with AI tools, achieve higher efficiency? The problem here is that economic miracles are usually not caused by faster repetition of tasks, but by exploration. So far, AI has increased efficiency rather than creativity. In a survey of more than 7,000 information employees, it was revealed that the weekly email tasks of those who used productive AI intensively decreased by 3.6 hours (31%) while collaborative studies did not change. But when everyone handed over email responses to ChatGPT, inbox volume increased and initial productivity gains disappeared. “The short-term increase in productivity in America in the 1990s teaches us that the gains from new tools, whether it's spreadsheets or artificial intelligence agents, are lost unless they're with groundbreaking innovations.” (OECD).
Major language models [LLM] turn to statistical consensus. A model trained before Galileo would replicate a world-centered universe like a parrot; if it was fed by 19th century texts, the Wright brothers proved that it was impossible for man to fly before achieving success. A review recently published in the journal Nature revealed that although LLMs ease routine scientific work, people are still ahead in decisive insights. Human cognition should be understood as a form of causal reasoning based on theory, rather than the importance AI attaches to information processing and data-driven predictions. AI uses a possibility-based approach to information and is largely backward and imitative, whereas human cognition is forward-looking and can produce real innovations.
The biggest goal of OpenAI and other AI companies is a super intelligently productive AI [GenAI] that can inherit innovation from humans. Until now, this has remained as legendary as a goal as the Holy Grail in literature. The current GenAI can only make gradual explorations, but not fundamental discoveries from scratch like humans.
But OpenAI's guru, Sam Altman, promises that their AI will not only do a single worker's job, they can do all the work: “AI can do an organization's job.” This is the way AI machines take over everything, as they take over the jobs of operating, developing and marketing, in companies (or even in AI companies?) means maximizing profitability by eliminating workers. Therefore, Altman and other AI emperors will continue to expand their data centers and develop more advanced chips, although they have beaten existing models of Chinese AI models such as DeepSeek. Nothing can stop the super-intelligent AI target.
Unfortunately, as MIT Tech explained, many AI models are infamous black boxes, meaning that while a algorithm can produce a useful output, researchers don't know exactly how this happened. This has been the case for years, and AI systems often refute theoretical models based on statistics. In other words, AI trainers don't really know how AI models work. This is the biggest obstacle to reaching the Holy Grail.
Therefore, the AI boom is still just a financial bubble. As one commenter said: “Productive AI doesn't perform the functions it's said to have done before it was sold, and what they can actually do isn't what they can do is that provide business returns, automate the workforce, or do much more than an extension of a cloud software platform. No money, no users, every company loses money, and some companies lose so much money that it's not clear how they survive.”
Meanwhile, the mass construction of data centers consumes an unprecedented level of energy. The International Energy Agency estimates that the electricity consumption of its data centers will double to 945 terawatt-hours by 2030. This is more than the total electrical energy that a country like Japan currently uses. Ireland and the Netherlands have restricted the development of new data centers due to their impact on the electricity grid. While there are large increases in electricity demand in data centers during the training of artificial intelligence models, fluctuations in renewable energy supply threaten the durability and capacity of existing energy systems.
When it comes to productivity and growth results, OECD keeps forecasts cautious. If AI technologies are widespread and successfully implemented, the OECD estimates that global workforce productivity will increase by 2.4% points over the next decade, contributing 4% to world GDP according to current trends. However, if AI does not succeed in reducing the need for human labor and does not spread to all sectors, labor productivity could increase by only 0.8% above the current trend level (from 0.8%) in a decade, and world economic growth will not change. The final decision on this issue has not yet been made.