Is There an AI Bubble? Who Decides and Who Bears the Risks of the Technology Boom

 By Javier Surasky

In my classes, one question comes up frequently: “Is there an artificial intelligence bubble?” Behind it is an understandable question prompted by announcements of multibillion-dollar investments: how much of that money supports sustainable businesses, and how much depends on expectations that may not be met?

The evidence I review here points to signs of speculation in some companies over certain periods, but it is not enough to describe the entire sector as a single bubble. A technology can also prove useful while some of the investments made to develop it become unprofitable. To understand the consequences of that possible mismatch, we need to follow the money and the commitments it creates.

A deal announced by Meta offers a way into this web of relationships.

In October 2025, the company announced a joint venture with funds managed by Blue Owl to develop Hyperion, its data center campus in Louisiana, with an estimated cost of around $27 billion for its buildings and infrastructure. The funds would hold an 80% stake in the property-owning entity, while Meta would retain 20% and lease the facilities for its own use. Part of the financing would come from debt sold to investors in financial markets (Meta, 2025).

The ownership split seems straightforward, but the allocation of risk requires a closer look. Meta also provided a guarantee covering the future value of the facilities: if it decided to terminate or not renew a lease, and the other agreed conditions were met, it could have to compensate the property-owning entity if the property’s value fell below the threshold set in the contract. Ending the lease, then, would not necessarily mean an end to payments (Meta Platforms, Inc., 2026, Note 5).

A 20% ownership stake, therefore, does not establish that Meta bears 20% of the risk. To understand its exposure, we need to examine the contracts connecting the company to the property-owning entity and its financiers.

This is where I see a problem for international AI governance: the agreements that make its expansion possible also allocate decision-making power and financial responsibilities. The race to build infrastructure begins long before we can know what returns it will generate, but the obligations that will make its construction possible are agreed upon from the outset.

Ownership Tells Only Part of the Story

Hyperion shows how a single company can play several roles within a project: Meta holds an ownership stake, will lease the facilities, and guarantees part of their future value. Each role entails different commitments.

According to its financial report for the second quarter of 2026, the leases would begin in 2029, with an initial term of four years per property and renewal options. Meta also reported that it did not consider payments under the guarantees probable and had therefore not recorded a liability for them in its accounts. In other words, it had disclosed the commitment but had not yet recognized an amount as a payment obligation on its balance sheet (Meta Platforms, Inc., 2026, Note 5).

This gap between what a company owns, what it owes, and what it could become obligated to pay helps explain a recent warning from S&P. As Marc Jones (2026) reported for Reuters, Mark Habib, the rating agency’s global head of digital infrastructure, said that AI-related obligations could exceed what is currently visible. He referred to interconnected structures in which large technology companies back infrastructure projects and smaller companies. He also clarified that S&P had no estimate of the additional debt and did not consider changes to its ratings necessary at that time.

Knowing how much each company owes may not be enough to understand this web of financing, in which creditors and debtors, partners and competitors, are intertwined. To make sense of it, we need to examine who must pay, when, and what happens if one party defaults.

Signs of a Bubble and Their Limits

To speak of a bubble is to ask whether the prices of certain assets are sustained by expectations that future revenues are unlikely to justify. A rapid rise in share prices or a large investment cannot, on its own, settle that question.

Sarkar and Wells (2026) develop a method for detecting price behavior consistent with speculative components, accounting for persistent changes in volatility. Analyzing stock prices from 2022 to 2026, they identify signs of a bubble in companies such as Nvidia, Alphabet, TSMC, Broadcom, Palantir, and Micron, but not in others, such as Apple, Microsoft, and Amazon. These signs appear at particular points during that period and do not mean that all these companies showed them at the same time (Table 2).

My answer to the opening question, then, is that the evidence points to signs of speculation in parts of the AI-related market, but the studies examined do not support the claim that the entire sector constitutes a single, uniform bubble or allow us to predict a widespread crisis.

Uncertainty also extends to infrastructure investments. Contracts allow these investments to get underway before the expected revenues materialize, but the obligations may remain even if demand grows less than anticipated or equipment rapidly loses economic value.

Storm (2025) takes this concern as far as predicting that the strategy of scaling models by using more data and processors will fail. I think his work is best read as a critical thesis whose predictions depend on assumptions about revenues, costs, and technological change. Its contribution to this analysis is the warning that financial commitments can outlast the period during which an investment retains its expected value.

A technology can continue to develop and find valuable uses while some projects built to take advantage of it become unprofitable. The usefulness of AI and the sustainability of each investment require different assessments.

Stock performance alone also does not reveal how much a company owes for building its data centers or who would lose money if it could not pay. Understanding those risks requires returning to the accounts and contracts. The question of a bubble thus leads to another, equally relevant question: how losses would be shared if certain expectations are not met.

Financing Enables Influence

Following the money also helps identify who gains the ability to influence decisions.

In a study of private credit, Lalafaryan (2024) shows that funds lending to companies can influence their governance even when those companies are not in financial distress. The study examines corporate financing in general, but it offers an insight for our analysis: those financing digital infrastructure may also be “buying” influence over the projects.

That influence needs to be examined in each deal. Loan terms can restrict a borrower’s decisions, while access to new financing can determine which projects continue and which must be scaled back.

Financing is also a barrier to entry in the technology race. An actor able to obtain capital on favorable terms has more room to build infrastructure and wait for results than one that needs immediate returns. That difference can strengthen established competitors, although financing can also enable new actors to enter. Its effect depends on who secures the resources and on what terms.

The money that expands technological capabilities also helps determine who can develop them and how long they can wait for returns. Decisions about the future of AI are also made when the terms of its financing are negotiated.

A Network of Connections Can Also Transmit Losses

Distributing commitments among several actors helps raise capital for large investments, but it does not guarantee that risks remain contained.

On October 13, S&P warned of a growing correlation in the financial performance of major cloud providers, semiconductor manufacturers, and other companies issuing AI-related debt. That connection reduced the benefits of diversifying investments within the sector (S&P Global, 2026). Put more simply, investing in several companies may offer less protection than expected if they all depend on the same growth expectations being met.

The literature on financial networks helps explain the mechanism. Drawing on evidence from German banks and a theoretical model, Elliott et al. (2021) show how certain connections among institutions with similar risks can increase risk across the system. The ties that allow risks to be shared can also transmit financial difficulties.

The study does not analyze AI financing, but it allows us to formulate a hypothesis for examining it. If expected revenues fail to materialize, some projects could be scaled back or postponed. Their suppliers would lose sales, and some borrowers would find it harder to obtain new financing. The extent of the damage would depend on the money their creditors had committed, the guarantees available, and the capacity to absorb losses.

A company can protect itself through a contract and transfer part of the risk to another. Assessing the safety of each deal also requires looking at what happens on the other side of that protection.

The State’s Role in the Web

The Hyperion guarantees examined so far are private commitments made by Meta. To understand when a state becomes exposed, we need to look at the terms on which it receives investment and offers support.

Competition to attract data centers can lead governments to commit resources before a project’s results are known. When they offer tax incentives, they forgo certain revenues. When they provide guarantees, they may become obligated to pay if the conditions that trigger them are met. It also matters which public commitments would remain if the company reduced its investment or abandoned the project.

The announced value of an investment does not answer those questions. Nor does it establish, on its own, that the investment is a “good deal” for the public.

International inequality makes that assessment more demanding. At the October meetings of the IMF and the World Bank, the relationship between AI, market concentration, and financial stability was on the economic agenda (International Monetary Fund [IMF], 2026a). Kristalina Georgieva noted that economies integrated into AI supply chains were performing better, while low-income countries and fragile states faced particular difficulties when they were left out of the technology boom and depended on energy imports (IMF, 2026b).

Being left out has costs, but the desire to participate does not remove the need to assess the terms of entry. The studies included here do not estimate the financial impact of these projects on developing countries. That is why I argue that understanding public and private commitments must be part of the assessment made before receiving an investment.

Assessing Risks Before Making Commitments

National financial supervisory authorities need information to identify who finances each company, what commitments bind the parties, and who would bear the losses if one of them could not pay. International cooperation can help trace relationships that cross jurisdictions and are only partially visible to a national authority.

For governments supporting these projects, the task includes disclosing payment commitments that could be triggered in the future. Commercial confidentiality needs to be discussed alongside these supervisory needs, especially when public resources are involved.

My position is that this financial dimension must be integrated into international AI governance. We do not need to wait until we have settled whether there is a bubble across the entire sector to examine the obligations already being assumed. Hyperion shows why an ownership percentage is insufficient to understand a company’s responsibilities. The same attention to contracts must inform the assessment of the obligations a state takes on.

A government may celebrate a new technology investment and later discover that some public commitments survive the investor’s withdrawal. Knowing what those commitments are and how much they could cost should come before the announcement, while there is still room to negotiate who decides and who bears responsibility for each decision and the uncertainties it entails.

References

Elliott, M., Georg, C.-P., & Hazell, J. (2021). Systemic risk shifting in financial networks. Journal of Economic Theory, 191, Article 105157. https://doi.org/10.1016/j.jet.2020.105157

International Monetary Fund. (2026a, October 16). AI and the global economy: From risk to opportunity. https://www.imf.org/en/meetings/2026/annual/schedule/2026/10/16/ai-global-economy-212373

International Monetary Fund. (2026b, October 14). IMF managing director’s briefing on the global policy agenda press briefing. https://mediacenter.imf.org/news/imf-managing-director-s-briefing-on-the-global-policy-agenda--press-briefing/s/3cf4bc1c-ae60-45b3-aa3b-c11d2d6d92db

Jones, M. (2026, October 19). Top hyperscalers likely to have more AI-linked debt than now known, S&P Global warns. Reuters. https://www.reuters.com/business/top-hyperscalers-likely-have-more-ai-linked-debt-than-now-known-sp-global-warns-2026-10-19/

Lalafaryan, N. (2024). Private credit: A renaissance in corporate finance. Journal of Corporate Law Studies, 24(1), 41–95. https://doi.org/10.1080/14735970.2024.2351230

Meta. (2025, October 21). Meta announces joint venture with funds managed by Blue Owl Capital to develop Hyperion Data Center. https://about.fb.com/news/2025/10/meta-blue-owl-capital-develop-hyperion-data-center/

Meta Platforms, Inc. (2026). Quarterly report for the quarterly period ended June 30, 2026 (Form 10-Q). U.S. Securities and Exchange Commission. https://www.sec.gov/Archives/edgar/data/1326801/000162828026050705/meta-20260630.htm

Sarkar, A., & Wells, M. T. (2026). Is there an AI bubble? Robust date-stamping for periods of exuberance. Frontiers of Mathematical Finance, 10, 1–42. https://doi.org/10.3934/fmf.2026005

S&P Global. (2026, October 13). Elevated sovereign yields; AI divide in credit markets; and political advertising in the 2026 US midterms. https://www.spglobal.com/en/research-insights/market-insights/daily-update-oct-13-2026

Storm, S. (2025). The U.S. is betting the economy on ‘scaling’ AI: Where is the intelligence when one needs it? International Journal of Political Economy, 54(4), 425–452. https://doi.org/10.1080/08911916.2026.2616133