Data centers are more expensive than states? Will the massive investments in artificial intelligence ever pay off?

Disappointing AI results: Why billions in business investments have so far yielded hardly any profits

The dangerous money cycle of AI: How tech giants are manipulating their success

While hyperscalers like Amazon, Microsoft, Alphabet, and Meta are pumping unprecedented sums into the expansion of artificial intelligence, with investments totaling over a trillion US dollars, the actual economic benefits for end customers remain alarmingly low. To finance this gigantic infrastructure, the industry is increasingly resorting to complex leasing arrangements and risky shadow debt, which is already leaving cracks in the balance sheets of even financial giants. This discrepancy between astronomical capital expenditures and stagnant profitability is fueling fears of an unprecedented misallocation of capital. The following analysis sheds light on the opaque web of special purpose vehicles, circular revenues, and geopolitical boundaries—and shows why the real weak point of the AI ​​boom lies not in the technology itself, but in its risky financing chain.

Related to this:

• Renting instead of building: Why the secret AI winner Anthropic doesn't own its own data centers

When data centers become more expensive than states and nobody knows who will ultimately pay

The question of whether the massive investments in artificial intelligence will ever pay off became the central axis of the global economic debate in the summer of 2026. Within just a few quarters, the capital expenditures of major American technology companies had reached a scale that would have been unthinkable three years earlier, while the actual returns on these investments appear vague at best, and at worst highly circular and obscured by accounting practices. This analysis organizes the available facts, identifies the limits of current knowledge, and provides a reasoned assessment of where this development might lead, without prematurely passing final judgment on success or failure.

An investment wave that is changing the balance sheet structure of the global economy

The four largest hyperscalers – Amazon, Microsoft, Alphabet, and Meta – have repeatedly revised their investment plans for the 2026 calendar year upwards over the past few months. Following reports for the second quarter of 2026, the combined forecast for the full year ranges from approximately $650 billion to $770 billion, with some bank analysts, including Oracle and other cloud providers, arriving at figures close to $850 billion. Microsoft is planning for around $190 billion, Amazon for roughly $200 billion to $220 billion, Alphabet for $180 billion to $190 billion, and Meta for $125 billion to $145 billion, all for the current calendar year. Compared to approximately $410 billion in 2025, this represents an increase of roughly 77 percent in just twelve months. Cumulatively, since the start of the current AI investment cycle in 2023, these four companies alone have spent approximately $1.1 trillion in capital by mid-2026, according to calculations by the Financial Times, marking a structural break with the technology industry's previous, low-capital software business model.

These sums are remarkable because they fundamentally alter the nature of the companies involved. Corporations known for decades for their extremely high free cash flows and low capital intensity are now moving toward a heavy industrial business model more reminiscent of steel mills, energy suppliers, or telecommunications network operators than traditional software companies. Alphabet reported negative free cash flow in the second quarter of 2026 for the first time in years—a signal widely discussed in the financial sector because it demonstrates that even the world's most financially sound companies are reaching the limits of self-financing. Analyst firms such as Goldman Sachs and Morgan Stanley predict that the necessary infrastructure investments will increase to several trillion US dollars between 2029 and 2031, making the question of financing sources beyond core business operations unavoidable.

How balance sheet manipulation could become a systemic risk

A key feature of the current investment cycle is the increasing shift of debt from traditional corporate balance sheets to special purpose vehicles, joint ventures, and complex leasing structures. The most prominent example is Meta's Hyperion data center project in Louisiana, financed through a special purpose vehicle with participation from the debt fund Blue Owl and other partners such as PIMCO. The loan volume is in the range of $27 to $30 billion, is structured through long-term leases, and therefore does not appear in its entirety on Meta's consolidated balance sheet. This model explicitly serves to protect the parent company's financial ratios, while the economic liability effectively remains.

Rating agencies and central banks have become significantly more critical of this practice in recent weeks. The Bank for International Settlements warned in its annual report that this form of hidden debt – often referred to as shadow debt – systematically understates the true debt levels of the technology sector and, in extreme cases, could trigger a sudden investment crisis. Analysts from Moody's, Morgan Stanley, and Japanese business publications such as the Nikkei arrive at widely varying estimates of the total volume of off-balance-sheet liabilities, leases, and special purpose vehicles, ranging from $1.2 trillion to well over $3 trillion, depending on which contractual categories are included. This enormous range demonstrates one thing above all: there is currently no uniform, verifiable methodology for quantifying the actual economic risk of these structures, and traditional debt ratios such as the net debt-to-operating profit ratio only partially capture this risk.

The problem becomes more concrete and measurable in the case of Oracle. In July 2026, the rating agency S&P Global downgraded Oracle to BBB- – the lowest level within the investment-grade range, just above the junk bond threshold. This step was justified by the massive debt financing of the AI ​​infrastructure expansion and, particularly significantly, by the extreme customer concentration: Approximately half of Oracle's contractually bound but not yet invoiced revenues – the so-called Remaining Performance Obligations – which have grown to $638 billion, are attributable to the customer OpenAI alone. This figure increased by 363 percent within a year, from approximately $138 billion to $638 billion, demonstrating how heavily Oracle's future earnings are concentrated on the solvency of a single, not yet profitable, customer. As a result, credit default swaps on Oracle bonds reached historic highs, a clear warning signal from the capital markets.

The second-tier stress test: When cloud providers themselves become a risk

Alongside the established hyperscalers, a second group of companies has emerged: the so-called neoclouds. These are specialized providers of graphics processing power (GPU), the most prominent of which is CoreWeave. These companies borrow enormous sums to purchase GPUs and then lease the computing power to model developers and businesses. As of June 30, 2026, CoreWeave had total debt of approximately $35.6 billion, spread across numerous secured and unsecured credit lines and convertible bonds. In the second quarter of 2026 alone, net interest expenses amounted to about $640 million—an amount sufficient to turn a near-profitable operating result into a net loss of approximately $626 million.

At first glance, this risk is mitigated by an impressive order backlog of approximately $104 billion, representing contractually guaranteed future revenues. However, closer examination reveals that the projected annual revenue range of $12.4 billion to $13.2 billion for 2026 means that CoreWeave can actually convert significantly less than 15 percent of this order backlog into revenue in any single year. This puts the size of the backlog into perspective and makes the question of the debt maturity structure all the more pressing. CoreWeave recently increased its investment budget for 2026 to $35 billion to $39 billion, forcing the company to continuously raise new debt capital in a rising interest rate environment, while a significant portion of its customer base is concentrated among a few large clients. A refinancing window is emerging for the years 2026 to 2028, which is considered one of the most vulnerable points in the entire AI financing system.

Here you will learn how your company can implement customized AI solutions quickly, securely and without high entry barriers.

A managed AI platform is your all-inclusive, worry-free solution for artificial intelligence. Instead of dealing with complex technology, expensive infrastructure, and lengthy development processes, you receive a ready-made solution tailored to your needs from a specialized partner – often within just a few days.

The key advantages at a glance:

⚡ Rapid implementation: From idea to ready-to-use application in days, not months. We deliver practical solutions that create immediate added value.

🔒 Maximum data security: Your sensitive data stays with you. We guarantee secure and compliant processing without sharing data with third parties.

💸 No financial risk: You only pay for results. High upfront investments in hardware, software, or personnel are completely eliminated.

🎯 Focus on your core business: Concentrate on what you do best. We take care of the entire technical implementation, operation, and maintenance of your AI solution.

📈 Future-proof & scalable: Your AI grows with you. We ensure continuous optimization and scalability, and flexibly adapt the models to new requirements.

More information here:

• The Managed AI Solution - Industrial AI Services: The Key to Competitiveness in the Services, Industry and Mechanical Engineering Sectors

Why the sales figures of the model suppliers remain the real mystery

The demand side of the AI ​​economy—that is, whether enough end customers will ultimately be willing to pay for AI services to justify the infrastructure investments—remains the most difficult part of the entire debate to verify. Anthropic and OpenAI both report very high annualized revenue rates, but the published figures vary considerably depending on the source. It remains unclear whether these are net or gross figures, actual billed revenues, or merely contractually agreed-upon volumes. Several reports indicate that Anthropic has now overtaken OpenAI in annualized revenue, while OpenAI is simultaneously grappling with substantial operating losses. These have been estimated at around $21 billion for 2025 and approximately $12.3 billion for the second quarter of 2026, based on leaked internal figures rather than audited public financial statements, as neither company is fully regulated by the stock market.

A particularly discussed special case is xAI, which merged with SpaceX in February 2026, followed by SpaceX's IPO in June 2026. This structure allows for the vertical integration of model development, the X platform, and the company's proprietary compute infrastructure called Colossus. The actual product revenues of the chatbot Grok are significantly smaller than the income from renting computing capacity to third parties, including even competitor Anthropic. This example illustrates that in the current AI cycle, economic success is increasingly achieved through structural integration and diversification of revenue streams, and less and less through independently proving that a language model can be operated profitably on its own.

A recurring point of criticism in expert discussions concerns so-called circular financing. Nvidia invests directly or indirectly in customers like OpenAI, who in turn purchase computing power from cloud providers. These providers themselves procure Nvidia chips and are sometimes involved with the same model providers. This interconnectedness means that a significant portion of the reported revenue is generated within a closed loop, where capital flows in a circular fashion without necessarily generating an additional external payment from a genuine end customer. Critics see this as a distortion of actual market demand, while the companies involved emphasize that these investments represent economically sound capital allocations within a growing value chain.

Related to this:

• Why traditional data centers are no longer sufficient for the AI ​​boom

What corporate clients actually get for their money

While trillions are being invested on the supply side, the picture on the demand side among actual user companies is considerably more sobering. A global survey published by McKinsey in August 2026, based on approximately 1,719 executives, concludes that 80 percent of organizations perceive individual productivity gains from AI, but only 37 percent can detect any measurable effect on operating profit – a figure that has remained virtually unchanged from the previous year. Only 6 percent of companies qualify as "high performers" according to McKinsey's definition, meaning organizations that attribute at least 5 percent of their operating profit to the use of AI; this figure is also unchanged from the previous year. This stagnation in hard financial metrics stands in stark contrast to the continued increase in companies' investment budgets and provides the skeptics of the debate with a strong empirical argument.

In parallel, the Domino Enterprise AI Report for 2026 reports that 57 percent of the surveyed companies are experiencing a return on investment that lags behind their actual expenditures, a plateau that has been observed since 2025. External forecasts from Gartner further predict that by 2027, more than 40 percent of so-called agentic AI projects—that is, projects with autonomously acting AI systems—are likely to be discontinued. These figures support the thesis that, so far, artificial intelligence has not been a decisive competitive advantage for most companies, but rather an experimental cost center with uncertain returns. At the same time, the same data does not imply that the technology is worthless overall, as a small group of companies are indeed achieving significant results, indicating a widening gap between successful and unsuccessful implementations.

Consolidation is taking place, but not where it is expected

A common expectation in public debate is that financially struggling model developers will sooner or later be acquired by wealthier competitors, potentially triggering a chain reaction of market consolidation. However, actual observations over the past twelve months reveal a different pattern. Instead of outright takeovers of frontier model developers by hyperscalers, minority stakes, long-term computing capacity contracts, and so-called partial acquihire deals—where essentially talent and licenses are acquired without taking over the entire company—dominate. The real wave of consolidation is instead taking place at the application level, for example, through Salesforce's acquisition of customer service AI provider Fin, ServiceNow's purchase of Moveworks, and IBM's acquisition of Confluent, supplemented by acquisitions of large Indian IT service providers specializing in AI-driven business processes.

This observation is economically insightful because it shows that the perceived economic value currently lies less in the development of new basic models, which are increasingly seen as interchangeable, but rather in the ability to embed these models in concrete business processes usable by end customers. An additional limit to consolidation arises from geopolitical factors: In April 2026, it was revealed that a planned acquisition of the Chinese AI company Manus by Meta had been halted or significantly hampered by a government order from China. This demonstrates that cross-border acquisitions in the AI ​​sector are increasingly coming under geopolitical scrutiny. Further rumors, such as Stripe's potential interest in acquiring the OpenRouter platform or Nvidia's interest in Hugging Face, remain unconfirmed speculation and should not be confused with documented transactions.

Regulation as an additional cost factor, but not a short-term obstacle

At the regulatory level, the European Union is currently furthest advanced with its AI Regulation. Since August 2, 2026, transparency obligations for general-purpose basic models have been in effect, along with corresponding enforcement by the European AI Office. The originally stricter obligations for AI applications classified as high-risk were postponed to December 2027 and August 2028, respectively, through the so-called Digital Omnibus procedure. This extension means that while companies will have to bear documentation and transparency costs in the short term, the actually cost-intensive compliance requirements for high-risk applications will only take effect after a considerable delay. For the core question of return on investment examined here, this means that regulatory pressure will not be a decisive factor for potential consolidation in the foreseeable future, even though it is likely to influence the long-term cost structure of the industry.

Two camps, two worldviews: Bubble or foundation of a new industry?

The public debate can be roughly divided into two camps, which differ fundamentally in their interpretation of the same data. The skeptical camp, which includes some credit analysts, well-known investors like Michael Burry, and numerous social media commentators, draws parallels to the dot-com bubble of the early 2000s or the overinvestment in telecommunications networks during that period. Their core argument is that current investment sums far exceed realistically achievable end-customer revenues and that a single data center with a capacity of one gigawatt incurs construction costs of an estimated 80 to 100 billion US dollars, while the achievable annual revenues are more in the range of 10 to 12 billion US dollars. This results in a payback period that is significantly longer than the actual technical and economic lifespan of the graphics processing units (GPUs) used.

The more optimistic camp, which includes analysts from Barron's, Vontobel, parts of Goldman Sachs, and Reuters, points out that the current valuation levels of the companies involved are significantly lower compared to the multiples of the dot-com era and that the financing corporations have considerably stronger operating cash flows than the telecommunications companies of that time. A report by the research firm Exponential View, distributed via Bloomberg, argued in early summer 2026 that AI-related revenues outside China had already exceeded the estimated write-downs of hyperscalers and neoclouds in the first quarter of 2026, albeit with very thin margins. Counter-analyses, such as those by the research platform WireSift or analyst Joseph Klement, conclude, however, that there is a structural gap of more than one trillion US dollars per year between the necessary and the actual end-customer revenue.

Significantly, both sides lack a unified, robust definition of what should even be considered a bubble, as well as a reliable forecast of end customers' actual willingness to pay for the years 2027 to 2030 and a transparent metric for the actual utilization of the newly created computing capacities. This knowledge gap is not merely an academic footnote, but the real reason why the debate is so polarized: Both sides argue with assumptions that are partly plausible, but not entirely verifiable empirically.

The real weak point lies not in the models themselves, but in the financing chain

A comprehensive review of the available data yields a nuanced yet well-founded assessment. The assertion that all previous investments have already paid for themselves cannot be supported by the available figures – neither at the level of the model developers, who are predominantly still operating at a significant loss, nor at the level of broad corporate demand, which, according to several independent surveys, has stagnated for over a year. At the same time, the opposing assertion of an imminent systemic collapse is equally unsubstantiated, as there has been no documented payment default, no failed follow-up financing, and no forced takeover of a major market participant. What can actually be observed are the first, clearly measurable stress signals at the system's periphery: Oracle's downgrade, the extremely concentrated customer relationships of individual cloud providers, the growing role of off-balance-sheet financing structures, and the high interest burdens of capital-intensive neoclouds like CoreWeave.

The most likely development is therefore less a single dramatic break than a gradual, but increasingly visible, differentiation within the industry. Companies with diversified revenue streams, a broad customer base, and strong balance sheets—especially the established hyperscalers with their profitable core business in traditional cloud computing—have sufficient buffers to weather multi-year periods of low return on equity. In contrast, more leveraged, customer-centric players such as individual neoclouds or infrastructure providers dependent on a few large customers, like Oracle, face a significantly higher risk of encountering serious difficulties in the upcoming refinancing windows between 2026 and 2028 if the growth of underlying end-customer demand does not accelerate considerably. Any consolidation is therefore likely to occur gradually through refinancing pressure, contract renegotiations, and selective distress sales of infrastructure capacity, rather than through a sudden, all-encompassing chain reaction as conjured up in some crisis scenarios on social media.

Three indicators are particularly crucial for monitoring this development: the actual maturity structure of debt held by capital-intensive infrastructure providers, the development of customer concentration in the largest cloud contracts, and whether the share of companies with a measurable profit contribution from AI applications will finally rise above the stagnant level of around 37 percent in upcoming surveys. Only when a clear trend reversal emerges can a reliable assessment be made as to whether the current wave of investment is actually creating a new industrial base infrastructure or whether it will ultimately go down in economic history as one of the most expensive misallocations of capital.

In industrial B2B, sustainable business relationships rarely emerge overnight. They develop step by step – through visibility, professional relevance, recurring touchpoints, and growing trust. Xpert.Digital's 4-stage model addresses precisely this: It offers a structured path that begins with a manageable entry point and can evolve into deeper collaboration in business development if needed.

Instead of relying on loud marketing promises, this model puts the relationship at the forefront. Companies start with clearly defined, easily calculable measures and then decide, based on their own experience, how far they want to expand the collaboration. A key factor for this undisturbed trust-building process: The platform completely avoids annoying advertising ads, so the editorial focus remains solely on the companies' expertise.

More information here:

• From visibility to trust: Your scalable path with Xpert.Digital

☑️ Our business language is English or German

☑️ NEW: Correspondence in your native language!

 

I and my team are happy to be available to you as your personal advisor.

You can contact me by filling out the contact form here wolfenstein@xpert.digital:or simply call me at +49 7348 4088 965. My email address is

I'm looking forward to our joint project.

 

 

☑️ SME support in strategy, consulting, planning and implementation

☑️ Creation or realignment of the digital strategy and digitization

☑️ Expansion and optimization of international sales processes

☑️ Global & Digital B2B trading platforms

☑️ Pioneer Business Development / Marketing / PR / Trade Fairs