The initial phase of the AI boom was a cash game defined by eye-watering, headline-grabbing injections. Shortly after OpenAI launched ChatGPT, Microsoft threw an initial billion dollars at the startup. Venture capital firms quickly followed suit, investing multiples of that figure. As ChatGPT’s viral adoption exploded, so did its operational costs. Microsoft doubled down, committing another $10 billion (all figures in USD).
It’s textbook Silicon Valley: tech giants, elite funds and micro-venture capitalists throwing billions of dollars at model startups. For investors, these were massive, highly speculative bets — but they were entirely equity risks. If a startup failed, the equity was wiped out, but nobody owed a commercial bank a dime. The collateral damage was paper wealth.
That era is over. The scale of what it costs to build out the physical infrastructure required to sustain the AI revolution has outstripped the financial capacity of even the world’s wealthiest technology companies.
The five largest technology titans building AI infrastructure — Amazon, Alphabet, Meta, Microsoft and Oracle, known as the hyperscalers — are on track to spend more than $1 trillion in 2025 and 2026 alone. These companies are spending $30–$40 billion per quarter. Each.
This spend now vastly exceeds the cash these behemoths are generating from their core operations. An increasing share of the AI infrastructure buildout is being financed with borrowed money — a structural shift that has triggered warnings from the Bank for International Settlements about financial stability risks. Meanwhile, the Bank of Canada has cautioned that leverage among non-bank financial institutions continues to grow.
To bridge the gap, even these historically cash-rich companies have fundamentally altered their behaviour. Between 2021 and 2024, the core group of hyperscalers had a modest footprint in credit markets, averaging just $28 billion in annual debt issuance. However, in 2025, this figure rose to a combined $121 billion in corporate bonds — over four times their previous four-year average.
In fact, AI-related infrastructure debt accounted for nearly 30% of all net issuance in the entire U.S. investment-grade corporate bond market in 2025. By mid-2026, that pace will have accelerated to a staggering $170 billion year-to-date. They are no longer borrowing to fund stock buybacks or strategic acquisitions; they are borrowing just to maintain their aggressive capital expenditure.
Heavy infrastructure companies
This transition changes the fundamental nature of technology investing. Historically, software companies were Wall Street darlings because they operated on an asset-light business model. They boasted massive gross margins, required little physical capital to scale and generated free cash flow rapidly because their product was replicable code.
AI has completely inverted this paradigm, turning tech giants into heavy infrastructure companies overnight. Today, their balance sheets look less like nimble software firms and more like electric utilities.
They are tied to capital-intensive operations, heavy hardware requirements and astronomical, continuous capital expenditures that compress margins through soaring power and chip costs. Yet, unlike traditional utilities, they lack the guaranteed, rate-regulated profits that shield energy companies from market volatility.
While tech behemoths tap the public bond markets, mid-tier cloud providers and specialized data centre operators face a different hurdle. Lacking the balance sheets of Alphabet or Microsoft, these companies are turning to the rapidly growing private credit market for billions of dollars in structured loans.
Here’s the challenge: In some cases, borrowers are pledging graphics processing units as collateral for the debt.
This creates a massive economic blind spot. Unlike standard industrial collateral like a piece of real estate, a commercial aircraft or a shipping vessel, a microchip possesses a steep depreciation curve. The moment Nvidia or Advanced Micro Devices releases a next-generation architecture, older chip models can lose value rapidly.
If a mid-tier cloud provider experiences even a minor dip in customer demand, its leasing revenue will collapse. Private lenders could be left holding warehouses full of rapidly aging, power-hungry hardware with liquidation values well below the amounts they financed.
The real estate side of the AI boom is equally leveraged. Modern data centres are no longer just passive warehouses with servers. They are highly specialized industrial plants requiring immense power grids, dedicated electrical substations and advanced liquid-cooling systems. Real estate investment trusts and private equity firms are financing these multi-billion-dollar builds, assuming that Big Tech will rent these facilities under long-term leases for decades to come.
Many lenders view long-term hyperscaler leases as exceptionally high-quality collateral because of the tenants’ strong credit ratings. That assumption could be tested if AI infrastructure demand slows materially.
If the software market decides it does not require this much raw computing capacity in three years, those lease commitments could be renegotiated or abandoned. Highly leveraged real estate developers and their credit syndicates will be left holding specialized, vacant structures with massive stranded-utility costs.
Perhaps the greatest systemic risk is that the financial system cannot accurately measure how deep this debt goes. An immense portion of this financing is occurring via private credit syndicates, shadow banks and off-balance-sheet vehicles rather than public, regulated, commercial banks. The true leverage is opaque. It does not show up on standard public market dashboards or regulatory disclosures.
This sets the stage for a dangerous, interconnected feedback loop. When corporations realize the immediate return on investment for AI software is lagging, they may rationalize their IT budgets and slow down spending.
This stagnation in software spending could cause a deceleration in cloud computing consumption for the hyperscalers and secondary cloud providers. As revenues dry up, mid-tier providers risk defaulting on their chip-backed private loans, while data centre developers may miss their debt service payments.
In a severe downturn, this could trigger a sudden asset devaluation and a subsequent liquidity freeze for private credit funds, rippling backward into the broader financial system.
The efficiencies promised by generative AI are not a myth. However, how an economy pays for a technological breakthrough dictates the macroeconomic outcome.
When a boom is funded strictly by equity, a market correction is painful with diminished stock portfolios and lower paper net worth.
When a boom is funded by multi-layered debt, a market correction can trigger an entirely different beast: corporate bankruptcies, structural defaults, asset foreclosures and systemic liquidity freezes.
The core question for investors right now is not whether AI is revolutionary. The real question is whether the actual revenue the software generates can outrun the historic mountain of debt being used to build it.
