Key Points
- The investment required to develop AI could reach $1,000 per person on the planet by 2030, i.e., up to $7.9 trillion according to McKinsey’s highest estimates.
- In the United States, this effort could amount to 3.2% of GDP per year through 2032, three times higher than during the 1990s telecommunications boom and six times that of the early 20th-century electrification wave.
- According to the Federal Reserve, US data centers raised $680 billion in debt in the first ten months of 2026, more than $2 billion per day.
- Without these investments in data centers, US growth would likely be lower by about one percentage point and roughly on par with euro area growth.
- More than half of AI company investments between 2021 and 2025 came from other AI companies, according to the Bank for International Settlements.
Testez le Grand Continent. Les idées et les analyses qu’on ne trouve nulle part. Mais que vous verrez bientôt partout. Dans votre boîte mail à partir de 8 euros par mois
The scale of the investment effort required to raise computing capabilities to a level compatible with the expected development of AI is dizzying. The estimates gathered by McKinsey range from $3.7 trillion to $7.9 trillion over the period 2025-2030. To give a more concrete sense of scale, the high end of this range represents an investment of $1,000 per person on the planet. According to recent work by Stijn Van Nieuwerburgh, in the United States these investments could reach 3.2% of GDP annually between 2025 and 2032, i.e., three times the effort required by the telecom boom of the late 1990s and six times that of the early 20th-century electrification.
In the first instance, frontier AI companies could finance virtually all of their investments from their profits. Since 2025, however, even hyperscalers, which typically boast abundant cash reserves, have had to resort to borrowing. According to the Federal Reserve, debt financing of all kinds aimed at developing data centers—bank loans, private debt and corporate bonds—reached $680 billion in the United States in the first ten months of 2026. Activities related to advanced computing technologies accounted for nearly a third of all issuance of corporate bonds rated Investment Grade since the start of the year.
The rise of AI-related debt has not yet reversed the downward trend in total debt held by American firms: debt not tied to AI is shrinking rapidly relative to GDP, but the overall decline is slowing. The debt ratio of US non-financial corporations to GDP fell by more than one percentage point a year in 2023 and 2024. This fell by half a point in 2025, then again since the start of 2026. If the investment pace envisaged by McKinsey and Stijn Van Nieuwerburgh comes to fruition, it is very likely that overall corporate indebtedness will have to rise again. Last summer, we argued that this new wave of private debt, colliding with already massive public financing needs, would push interest rates higher worldwide, risking to “stifle” more traditional sectors.
We focus here on the risks inherent to this AI debt frenzy. Even if a slowdown in deployment at the technology frontier can be justified—to allow time to implement the necessary regulatory guardrails—the global economy has become “dependent” on the AI investment boom. Without investments in data centers, American growth would likely be lower by about one percentage point. Beyond borders, many firms thrive by supplying equipment to this ecosystem. A sharp drop in the valuation of tech stocks would be painful for investors well beyond the United States, including in Europe where exposure to the U.S. stock market has tripled in five years.
Moreover, circularity, or endogeneity, of the AI ecosystem, in which financial ties among a relatively small set of players intertwine with commercial links, as well as the multi-layered financing structures of many AI-related projects, makes the sector dependent on the fate of a handful of players. The complexity of the financial arrangements ultimately masks the viability foundation of this growth. The final demand for AI services arising from outside the tech sector and its ability to cover the cost of debt must also be considered. The demand generated by technological innovation can follow a geometric rather than linear progression, but small early missteps in judging the pace of this progression can translate into substantial gaps after a few years.
CoreWeave: A Case Study
Just as Jensen Huang’s departure from Nvidia triggered a fresh wave of selling in the tech sector, market sentiment about AI had initially improved thanks to quarterly results and CoreWeave’s forecasts beating expectations. We take this company as a starting point to examine, through a concrete case, some of the challenges currently facing the entire AI ecosystem.
CoreWeave is a “neocloud.” Put simply, the company operates and sells compute capacity. It does not directly own data centers: once the owner has provided the building, the electrical hookups and the cooling system, CoreWeave supplies GPUs—acquired, for example, from Nvidia—operates the facility, manages cloud access, and sells compute services to its clients.
This business model places CoreWeave at the heart of the AI value chain. The company prospers if it can secure long-term payment commitments from its clients that cover both operating costs and, crucially, the cost of the debt needed to build its compute capacity. Contractual cash flows and the value of the equipment serve as collateral for its borrowings. If the GPUs remain usable at the end of a client’s commitment and the debt has been repaid, CoreWeave will then have a stock of capital free of any obligation, which it can re-commercialize for another client.
The company is still far from this ideal. At present, its net debt, excluding operating lease contracts, stands at seven times its EBITDA over the last twelve months. An entity like CoreWeave faces substantial risks. Beyond the usual credit risk—the possibility that a client fails to honor its payment commitments—fluctuations in interest rates can also weaken it given its leverage. But it also bears technological risk. If the GPUs it buys become obsolete more quickly than anticipated, its future capital stock—and thus the bedrock of future profitability—shrinks, and clients may move to operators closer to the technological frontier when commitments expire.
Revenue and adjusted EBITDA for CoreWeave surged in the third quarter—even as net loss widened substantially and the cost of servicing its debt rose sharply. More importantly, the company announced it expected roughly $9 billion in upfront prepayments and deposits from its clients over the next three months. These early receipts help reduce borrowing needs. That clients are willing to pay in advance to secure future compute capacity also reassured a market that had begun to worry about the robustness of demand for AI-related services. Yet a key question remains: do these upfront payments necessarily reflect the strength of final demand, i.e., demand from outside the AI ecosystem? CoreWeave did not specify which clients agreed to these upfront payments, but its existing client base is largely dominated by AI-related firms themselves. This is understandable: a middleman like CoreWeave focuses on selling compute to AI model developers or hyperscalers such as OpenAI, Anthropic, or Meta, which then offer services to clients outside the tech ecosystem. CoreWeave highlights expanding its client base beyond the “heart of AI.” But, in itself, high revenue that mainly reflects internal demand within the tech sector—which can itself be funded by debt—does not constitute proof of the financial viability of the AI sector. We will return to this.
Circularity has, thus far, been one of the hallmarks of the AI revolution. CoreWeave illustrates the endogenous nature of this sector. Nvidia is its supplier—selling GPUs to it—but also its shareholder, with a 10% stake. Nvidia also commits to purchase compute capacity that CoreWeave cannot sell to its customers under an agreement initially valued at $6.3 billion. In short, Nvidia is simultaneously a supplier, a potential ultimate customer, and an investor in CoreWeave.
Endogamy: a Teething Problem of Technological Innovation
It is by no means an isolated case. According to a Bank for International Settlements study, between 2021 and 2025, 55.2% of the investments received by AI companies, in value, came from other AI companies. Moreover, 46.4% of the value of cross-company investments were accompanied by business relationships within the supply chain. The study covered all kinds of arrangements linking financial relationships to business relationships, regardless of the direction of those relationships: the investor could be a supplier or a customer of the beneficiary company, or both. The dominant form of circularity is the one that unfolds downstream: an investor provides products or services to the company in which it invests.
These arrangements are not exactly new. Historical parallels can be drawn during major waves of technology when innovators must stimulate demand for their own products by supporting downstream users. The most relevant parallel seems to be the 1920s electrification era. General Electric—the engine of the stock frenzy of that decade—sold electrical equipment to utilities operating power plants, but also diversified into households by venturing into radio receiver manufacturing. To stimulate sales of receivers, GE needed content for users. That was RCA, founded by GE, which marketed receivers but also created record companies and radio stations. Blurring the lines between suppliers, customers, competitors, and partners, RCA eventually partnered with Westinghouse—GE’s direct competitor—to create NBC, the National Broadcasting Corporation. In short, mass electrification of the interwar period was characterized by strong endogenous relationships among its main players. One can see a parallel with the deal that gave birth to NBC—two competing suppliers joining forces to support their customers’ demand—in the December 2023 investment by Nvidia and AMD in the same AI developer, Essential AI.
The BIS study acknowledges that much academic literature justifies certain forms of circularity. In particular when it comes to innovative products and young firms with limited track records, customers and suppliers can better understand a company’s prospects than traditional investors. Circularity allows established firms to “lend” their financial credibility to new entrants and to make the investments required financially viable. It can also help align incentives among the different players. AI projects often meet very specific technical specifications, and suppliers who tailor their products to their clients’ particularities can incur significant irrecoverable costs if those clients ultimately withdraw from the project: they would then have difficulty offering their products to other clients. All of this leaves the question of the overall risk volume in the ecosystem intact, while making it harder to identify who bears the ultimate risk.
Anatomy of the Multi-Layered Structure
One of the defining features of the current surge in AI-related investment is the number of “tiers” within its financial structure. Rather than pursuing a “vertical integration” in which hyperscalers or neoclouds directly own data centers, these centers are frequently held by project companies, or special purpose vehicles (SPVs), whose initial equity is often provided by private equity funds. The SPV then borrows by combining traditional financing—bank loans, market-traded bonds—and alternative sources, such as private debt. The credit quality of this debt rests on the multi-year commitments of hyperscalers to lease the data center. In his illuminating study, Stijn Van Nieuwerburgh uses Hyperion, a major data-center project, as an example. The SPV that owns it, Beignet, despite lacking an independent credit history by design, issued a privately placed $27 billion bond. Standard & Poor’s assigned it an A+ rating, just one notch below Meta, which has committed to a sequence of leases covering the bond’s life.
Van Nieuwerburgh’s study also emphasizes the role of accounting rules that allow hyperscalers to foster the build-out of the capital stock they need without unduly burdening their balance sheet. Lease contracts do not have to be recognized on the balance sheet immediately, but only when they take effect, i.e., when the data center is fully constructed and operational. During the project’s construction, lease commitments that underpin its financing remain “off-balance sheet.” While this can lead to underestimating the system’s total debt, the logic is that the owner’s failure to complete the project would relieve the hyperscaler of its obligations. The lease structure allows the parties to share technological risk. In Hyperion’s case, successive lease contracts with Meta include an escape clause. At intervals of a few years, Meta can disengage—e.g., if the technology becomes obsolete—while compensating the owner for lost revenue from resale of the assets. The technological risk can thus be distributed across the different layers of the arrangement, but it cannot be entirely eliminated.
The circularity creates two distinct risks. The first, of course, is that the fate of the entire sector could depend on a handful of participants. One concern that emerged after Jensen Huang’s departure from Nvidia is that his successor might reduce the company’s support for other ecosystem players. This concern was reinforced by reports that Nvidia’s chief financial officer, who has since become interim chief executive, had warned Huang against certain circular arrangements. The second risk is that the complexity of the financial structures could end up masking the fundamental viability of the AI boom: the final demand for AI services from outside the technology sector, and its ability to cover the cost of debt.
To illustrate, return to our detour into the 1920s electrification. If GE benefited from the demand for equipment and content spurred by RCA—and the downstream effects on electricity consumption—the revenues ultimately came from the spending of “ordinary” businesses and households. Employees spent a portion of their pay on radios, while companies allocated a growing share of their advertising budgets to this new medium. NBC reported in 1935 that radio network advertising revenues had been multiplied by ten between 1927 and 1932. At the outset of the boom, however, the forecasts probably anticipated even faster growth. On average, between 1927 and 1930, radio advertising revenues nearly doubled each year. If that pace had continued, by 1934 they would have been eight times higher than their actual level. Although radio networks weathered the Great Depression better than traditional media, they could not escape NBC’s financial difficulties in 1932. The demand sparked by technological innovation often follows an exponential trajectory, but it is particularly hard to forecast its pace in advance.
« Estimating the Curve’s Rise, the Heat of the Flame »
A skeptical reader would probably object that we chose an extreme example there. The 1930s crisis was deep and prolonged. Yet we can examine another episode, less dramatic: the rise and fall of the telecommunications sector between 1997 and 2003. This sector was a central element of the “new economy” narrative. While the rise in investment spending was far more modest than what we currently observe around AI, the race to expand bandwidth to meet the advent of the Internet had a visible impact on GDP in the late 1990s and early 2000s. While most commentators at the time focused on Internet startups and their unsustainable valuations, a 2003 study by economists at the Federal Reserve Bank of Richmond found that “the telecommunications sector represents a much larger share of gains and losses in market capitalization than Internet companies.” According to this study, the telecom boom did not meet the authors’ definition of a “bubble.” In a bubble, it is the price expectations—which are nothing more than the belief that prices will continue to rise, without a plausible economic explanation—that drive investment. The authors suggest that investments were based on a fundamentally correct expectation: demand for telecommunications services would rise massively. But they turned out to be excessive because the market was overly optimistic about deregulation effects, did not fully account for technological progress, and misread the precise magnitude of the real increase in demand.
The advances in optical transmission allowed a single fiber to carry roughly forty times more data in 2000 than in 1996. Operators therefore built excess capacity, whereas a more intensive use of existing capacity could have met demand. This was a classic instance of materializing a technological risk. At the same time, operators were encouraged by the 1996 Telecommunications Act, which was meant to enable them to combine their new long-haul fiber networks with access to customers’ homes and offices, paving the way for bundled offers. But lawsuits and uncertainties surrounding access rules delayed this process and forced operators to revise their expected revenues downward. They were also mistaken about the pace of the geometric rise in Internet demand. A common expectation was that Internet usage would double every ninety days. Reliable data on this topic were hard to come by in the 1990s, but the Richmond Fed study cites an average annual doubling. The gap obviously has serious consequences. As the authors note: “Growth rates of 100% every three months would have implied that Internet usage was multiplied by 17 million between 1994 and 2000. Forecasts based on these growth rates and the 1994 Internet usage data imply that every Internet user in 2000 would be continuously streaming video.”
Back to the present. Midterm elections, scrutiny of data centers and highly publicized cases of “model misalignment” increase the probability of tighter regulation, or at least greater oversight, of AI deployment modalities. The technological risk is inherent to AI: less capital-intensive solutions delivering the same computing capacity would challenge current investment plans. Projections of future demand for computing from outside the tech sector are, by their nature, fragile. One study estimates total AI expenditures in the United States at $280 billion in 2026, up 40% from the previous year, based on a survey of average spend per employee. But the essential question remains the growth pace going forward. AI adoption is now widespread—the authors note that 78% of firms across seven countries use at least one AI technology—without necessarily assuming a linear relationship between outside-the-tech-sector adoption and demand for computing capacity. Firms could intensify their use of AI by turning to ever more sophisticated models as their employees become familiar with the new tools. Conversely, if non-tech-sector firms decide to allocate their AI development budgets to “older” models, the share of computing capacity devoted to training cutting-edge models could decline, leaving more room for inference, i.e., handling end-user requests. The capacity would then have less need to grow.
Let us be clear: the telecom episode suggests that the market has reason to examine regulatory developments carefully, and that technological progress can ultimately reduce capital needs rather than raise them. We do not claim, however, that the AI ecosystem is necessarily on the cusp of a painful correction. Current expectations for computing capacity requirements may be accurate. Our point is more modest. Feeling reassured by observing a “circular demand”—for example, when chipmakers’ valuations are supported by new commitments from hyperscalers or neoclouds to add data centers—can be misleading. The right indicator is demand from the rest of the economy, for which precise data are scarce at this stage, and the outlook difficult to quantify.