ECONOMYAugust 18, 2026

The AI buildout rests on hidden debt

Elisabeth Krecké

Reading time: 10 min

Financing the AI buildout increasingly depends on debt that sits outside corporate balance sheets, spreading exposure beyond the technology sector.

Artificial intelligence may be a digital revolution, but its foundations are anything but virtual. Behind every AI application – from the simplest chatbot to the most advanced autonomous system – lies a large-scale physical infrastructure that must be built, financed, powered and maintained.

Financing that infrastructure increasingly means one thing: debt. And debt brings new risks, raising questions about how financially stable the AI boom really is.

Concrete, chips and power

Over the next few years, millions of servers will be needed to train and run more and more sophisticated models around the clock. This will require a massive expansion of data-center capacity worldwide. New facilities will rise. Existing campuses will expand. Ever more powerful clusters of graphics processing units will be deployed within them.

The build-out will consume vast quantities of semiconductors, storage systems, networking equipment, fiber-optic cables, cooling technologies and, above all, massive electrical power. In the emerging AI economy, compute – the combination of processing power, hardware, data storage and energy – is becoming both a strategic resource and a critical factor of production.

All of this demands substantial capital. According to McKinsey, global spending on AI-related data centers and supporting infrastructure could reach as much as $7 trillion by 2030. Some observers argue that this may become the largest peacetime investment project in history. It is certainly one of the largest wagers ever placed on a technology whose economics remain unproven.

Who pays for it? Who bears the risks? And who is left exposed if expectations prove wrong?

Boom or bubble?

If demand for AI were to slow even modestly, the implications could be significant. Data centers may be built ahead of need. Capacity may exceed demand. Capital may be stranded. Economists call this over-investment or capital misallocation. Investors call it losses.

One constraint rarely appears in financial models. Building and operating AI infrastructure requires enormous amounts of land, water and electricity. As the build-out accelerates, these resources will compete with other economic and social priorities. Local communities are beginning to notice. In some regions, resistance is already emerging.

If hundreds of billions of dollars end up tied to infrastructure that proves less necessary than expected – or becomes obsolete more quickly than anticipated – the consequences would not stop at the technology sector. They would reach lenders, investors and financial markets.

Do the economics of AI justify the scale of investment now underway? And if the sector were to experience instability – whether due to valuation shocks, business conflicts or technological shifts – could the fallout become systemic simply because so much capital is now involved?

Fragile business models

AI companies’ valuations alone give pause. In late 2025, Nvidia, the dominant designer of AI processors, became the world’s most valuable listed company, at times exceeding a market capitalization of $5 trillion. More than 1,000 AI start-ups are now valued above $100 million. Hundreds have achieved unicorn status with valuations above $1 billion. Perhaps markets are correctly pricing the future. Perhaps not.

In October 2025, SoftBank sold its entire Nvidia stake – 32.1 million shares – for $5.83 billion. One transaction proves little. Yet it suggests that at least some sophisticated investors may believe that valuations in parts of the sector have moved well ahead of fundamentals.

Valuation is only part of the story. Many of the industry’s most highly valued companies remain deeply unprofitable. For instance, OpenAI, whose release of ChatGPT helped ignite the current AI race, is valued at roughly $850 billion despite years of losses. In 2025 alone, the company is estimated to have lost around $21 billion at the operating level.

One company’s investment becomes another company’s revenue, and vice versa.

The gap between valuation and profitability is striking. So is the gap between revenue and spending ambitions. OpenAI generated roughly $13 billion in revenue in 2025. Yet CEO Sam Altman has spoken of infrastructure commitments reaching as much as $1.4 trillion over the next eight years. Such ambitions make sense only if revenues grow dramatically and continue to do so for years to come. There is little room for disappointment.

The problem is monetization: how to make money from AI. Millions of people use AI systems and businesses are experimenting with them. Adoption has been rapid, but popularity does not necessarily generate profits. Many consumers remain unwilling to pay for AI services. Many businesses are still searching for applications that generate measurable economic value. The technology has arrived. The business model is still being tested.

Circular capital flows

The AI boom rests on an optimistic assumption: Demand for compute will continue rising almost without limit. Yet part of that demand may be less independent than it appears. The AI sector is bound together by a dense web of commercial relationships and strategic partnerships. Companies invest in one another. They then use that capital to purchase goods and services from the very same firms that provided it.

Nvidia offers a striking example. The company has taken equity positions in the AI developers and cloud providers whose expansion, in turn, drives demand for Nvidia’s chips. Among them is a $30 billion investment in OpenAI’s most recent funding round, alongside smaller stakes in CoreWeave, Nebius and Anthropic.

One company’s investment becomes another company’s revenue, and vice versa.

Some observers argue that this resembles a system in which suppliers effectively subsidize their largest customers, helping sustain – or, as critics would say, artificially inflate – demand for their own products. Industry insider Stefano Mainetti describes parts of the sector as operating within “closed circuits” that facilitate a form of “sophisticated demand engineering.”

The mechanism is powerful. Growth attracts investment, investment generates demand, and demand justifies further investment; the cycle can continue for a long time. The danger is that it may create the appearance of inexhaustible demand even as vulnerabilities accumulate beneath the surface.

Credit supercycle

Five American hyperscalers have emerged as the dominant builders of AI infrastructure: Amazon, Alphabet (Google’s parent company), Meta, Microsoft and Oracle. According to Morgan Stanley, taken together, they are expected to spend roughly $800 billion on capital expenditure in 2026 and close to $1.1 trillion in 2027.

For now, only these hyperscalers appear capable of funding investments of such magnitude from operating cash flows alone. Yet even they are borrowing more. In 2025, the five issued a record $121 billion in debt – more than four times their average annual borrowing between 2020 and 2024. Their borrowing has become so large that they now rival major American banks as the dominant issuers in the investment-grade bond market.

Outside this small group of top-rated global firms, the picture looks less reassuring. Many AI companies are borrowing heavily despite generating little or no profit. A prominent example is xAI, Elon Musk’s artificial-intelligence venture. Its willingness to raise debt while losses continue to mount has led some critics to describe the strategy as reckless. The AI boom has become a credit story – or, as financial analyst Stephen Innes has put it, a credit supercycle.

More on artificial intelligence

• Data centers become the new frontlines of global power
• OpenAI, public interest and the problem of tech governance
• Economic, fiscal and monetary policies need a radical rethink

Hidden debt

The largest AI borrowers face a difficult task: how to persuade investors that rising debt is the price of building the future rather than a sign of growing fragility.

Part of the answer lies in increasingly complex financing structures designed to keep debt off corporate balance sheets. At the center are so-called special purpose vehicles (SPVs). These are typically backed by private credit, institutional investors and major banks.

The arrangement works roughly as follows. The SPV owns the underlying assets: land, data centers, servers, computing equipment and so on. The technology company then leases the infrastructure under long-term contracts.

The debt sits largely inside the SPV rather than on the company’s balance sheet. The company’s lease payments generate the cash flow used to service that debt, while the financing is ultimately provided by private-credit funds, insurers, pension funds and other institutional investors.

The appeal is clear. AI companies gain access to enormous sums while preserving their credit ratings and borrowing capacity. Critics see something else, with some describing these structures as the AI boom’s hidden debt. Others call them accounting tricks. The Bank for International Settlements uses a more measured term: shadow borrowing.

There is nothing inherently improper about these arrangements. SPVs have financed infrastructure, real estate and energy projects for decades. Yet their growing role in AI raises an obvious question: If billions of dollars of obligations sit outside conventional debt metrics, how much leverage is accumulating across the sector?

Dispersing the risk

Banks, meanwhile, are becoming more cautious. According to the Financial Times, even some of the world’s largest lenders – including JPMorgan Chase, Morgan Stanley, SMBC and MUFG – are approaching internal limits on exposure, with some reportedly beginning to “choke” on a growing “glut of data-center debt.”

Their response is familiar. First comes syndication: Multiple banks share a single large loan. Risk is spread across several balance sheets. Then comes distribution, as banks increasingly sell portions of these loans to non-bank investors, sometimes at a discount.

They are also making greater use of significant risk transfer instruments, or SRTs. These structures allow banks to transfer part of the credit risk on loan portfolios to outside investors without removing the underlying assets from their balance sheets. Buyers typically include private-credit funds, hedge funds, pension funds and insurance companies.

The process does not stop there. According to the Bank of England, part of bank and non-bank lending to the AI sector is subsequently repackaged into collateralized loan obligation structures, diffusing exposures across the financial system. The debt is sliced, bundled and sold on. With each step, the location of the risk becomes harder to identify. If conditions deteriorate, losses may surface far from where the original loans were made.

The pattern has been seen in previous credit cycles.