In this article — choose a section
Trillions of dollars’ worth of infrastructure is being built around artificial intelligence, and its repayment depends on enormous expectations for future profits. At the same time, the technology is making it easier to imitate competence. Between these two narratives stands a cost that is difficult to fit into a sales presentation.
Artificial intelligence can write a convincing business plan. That does not mean, however, that revenue will oblige by delivering on it.
In October 2026 the International Monetary Fund identified disappointment over AI investment and profits as a possible trigger for a market correction. The risk is amplified by debt financing and the concentration of investment. Thus, the debate over the “AI bubble” extends beyond the usual shouting match between technology optimists and professional doomsayers. There are concrete financial dependencies to examine.
A less spectacular calculation must be added to these: the effort required to determine whether what a machine produces deserves to be trusted. This is where financial enthusiasm encounters an unpleasant obstacle. Generating more work and creating more value are different achievements.
The trillions are in the forecasts. The costs have to be paid back…
The scale deserves attention even without dramatic music. In an April 2025 analysis, McKinsey estimated the investment required in data centers through 2030 at approximately €6.05 trillion, of which about €4.69 trillion would go toward infrastructure supporting AI workloads.
This is a projected capital requirement dependent on assumptions about future demand. The sum is neither an account of money already spent nor a guarantee of orders. The authors themselves warn that overbuilding could leave assets without sufficient economic use.
The monetary values in the analysis have been converted using the ECB reference exchange rate for November 12, 2026: €1 to $1.1081, rounded.
To pay for such infrastructure, there must be enough solvent customers. They must use the services over an extended period, at prices that cover costs and generate a return. An impressive demonstration helps make the sale. But electricity, equipment, maintenance and financing still have to be paid for after the applause.
The risk arises when revenue expectations outpace customers’ ability to turn the technology into actual savings or additional income.
When your investor is also your supplier
Particularly revealing is the IMF’s review of 14 publicly announced deals since 2023. It covers approximately €271 billion in equity investments, supplemented by €587 billion in debt, against commitments for future revenue of about €1.26 trillion through services or chip orders.
These figures describe different aspects of interconnected deals. Adding them together as one impressive “total investment” would distort the picture.
The mechanism could look like this: a large technology company invests in an AI developer that buys computing services from it. The financing supports the customer, while its purchases fuel the supplier’s revenue. Such a relationship can have entirely sound business logic.
The vulnerability appears if end demand fails to meet expectations. The investor could then lose both on the value of its stake and on the revenue it expects from the same partner. The IMF explicitly examines this risk.
Money can make an impressive circuit among several companies. At the end of the circuit, however, there must be a customer who pays because they receive a sufficiently useful result.
Productivity faces a difficult task: justifying the promises
Among the academic foundations of the discussion is the working paper by Jessica Wachter and Jonathan Wachter, published by NBER and revised in October 2026. The authors examine what productivity expectations are compatible with enormous infrastructure investments.
In the updated model, calibration against investment implies approximately a threefold increase in productivity in the AI sector. This is a result within an economic model based on specific assumptions. It does not mean that productivity across the entire economy must triple, nor does it represent an established timetable for the bankruptcy of specific companies.
The distinction also matters for how we read the news. The article distributed by Yahoo is based on a publication by MIT Technology Review. The wording “MIT explains” in BeInCrypto’s headline does not turn the journalistic publication into an institutional verdict by the university.
The economic stakes remain serious enough without this additional authority: andinvestments must be followed by corresponding profits. The number of answers generated, by itself, does not prove that this will happen.
Anyone can look competent
In everyday life, we rarely check everything from scratch. We look at whether a document is well argued, whether a project has an active community, and whether the author can show consistent work. Such signs help us decide whom to give our time to.
Generative AI complicates this judgment because it can produce a convincing outward form with little effort from the person presenting it. Polished language loses some of its value as evidence of understanding. We get text that looks ready before we have established whether its content is accurate.
This problem has a specific research basis. In a study by Ranindya Paramita, Christian Kästner and Lori Williams 252 sources from Google and 870 Reddit discussions about trust in software dependencies were analyzed. The authors distinguish deliberate manipulation, behavior that imitates legitimate activity, and inflated metrics resulting from the use of AI, including without malicious intent. The study concerns the software environment; it does not measure “how much of the entire internet is fake”.
The broader public conclusion is more sweeping, but it should be identified as a conclusion: if good presentation becomes easy to imitate, assessing actual quality requires more work. An excellent specialist and a well-prepared imitator can arrive with equally respectable presentations. The difference emerges after the presentation.
I saved myself an hour. Check it for me in two.
Kästner describes this conflict quite directly in his September 2026 publication.. His criticism concerns sending out unprocessed or poorly edited AI output, with the burden of checking it shifted to the recipient. One person’s time savings can become someone else’s additional work.
For an organization, this is a significant accounting blind spot. An employee may report having prepared more documents. Their colleagues may spend more time reading, checking and correcting them. If only the volume produced is measured, the first result looks like success, while the second remains hidden in everyone else’s calendars.
Another research paper involving Kästner, based on interviews with 38 practicing professionals, examines precisely the rising cost of trust in software supply chains. Participants describe the accumulation of checks and controls, the automation of verification, and the transfer of decisions to trusted intermediaries. This is a qualitative study of a specific professional environment, with no claim to provide an overall monetary assessment of the problem.
This is where the two narratives meet. If part of the promised savings is consumed by quality control, the financial return from implementation may prove weaker than expected. This is a possible economic mechanism that deserves to be measured. It does not, on its own, prove that a financial collapse is inevitable.
The sales demonstration usually ends when the answer appears on the screen. The working day of the person responsible for it may begin precisely then.
Useful technology can also be overvalued
There is evidence of real benefits from AI. In the study “Generative AI at Work,” published in the Quarterly Journal of Economics in 2025, data on 5,172 customer service employees show an average 15% increase in cases resolved per hour after the introduction of an AI assistant. The effect varies significantly among employees and is especially beneficial for the less experienced.
This is a measurable result in a specific work environment. It deserves more attention than general assurances that “we will all become ten times more productive.”
At the same time, the usefulness of a product does not automatically determine the fair value of the company selling it. Customers may be receiving value, while investors have paid too much for future earnings. Competition may make the service more affordable while simultaneously making it harder to repay the costly infrastructure.
For a Bulgarian company, editorial office or agency, the practical conclusion is to measure the entire process: the time needed to create, check and revise content, the quality of the final result and the actual cost. The number of pages produced is a convenient metric. The client has usually commissioned completed work.
If the bubble bursts, the bill will not stay in the chat
In an adverse scenario, losses would depend on the specific investments and contracts. Shareholders could suffer declines in valuations. Creditors could face greater difficulty servicing their debts. Delayed investments could affect suppliers and employment on related projects.
On October 15 Kristalina Georgieva warned of the possible international spread of losses in the event of a severe correction in the highly interconnected AI landscape. Cross-border investments mean that geographic distance from Silicon Valley is not sufficient to ensure financial isolation.
Another possible outcome remains: the promised productivity gains materialize, but more slowly, more unevenly and at different companies emerging as winners than expected. Even such a development could punish expectations that were bought at a high price.
To assess whether the risk is turning into an actual crisis, specific changes must be monitored: are customers paying sustainably, how are profits developing after costs, are contracted purchases being carried out, and on what terms are loans being refinanced? A new record in a technology demonstration does not answer these questions.
As of November 12, 2026 the IMF is monitoring market movements but considers it premature to conclude that a broader reassessment of the long-term profitability and productivity of AI is beginning, Reuters reports. The risk has documented grounds. When the bubble will burst, however, we still do not know for certain…