The numbers are staggering, and they should make you nervous. Global data center dealmaking surged past $61 billion in 2025, breaking last year’s record of $60.8 billion amid what S&P Global calls a “global construction frenzy.” More than 100 transactions closed in just the first eleven months of the year. And the debt financing underpinning this buildout has reached levels that would have seemed reckless five years ago.
Debt issuance for data center infrastructure nearly doubled to $182 billion in 2025, up from $92 billion last year. Meta alone has raised $62 billion in debt since 2022, with nearly half of that total issued this year. Google and Amazon added another $29 billion and $15 billion respectively. Oracle, the hyperscaler with the weakest balance sheet of the bunch, has been borrowing so aggressively that Barclays analysts predict the company may run out of cash by November 2026 if current trajectories continue.
This is either the greatest infrastructure buildout since the transcontinental railroad or the beginning of a spectacular bubble. The honest answer is that nobody knows which yet.
The AI Demand Story Is Real, But So Are the Warning Signs
The fundamental driver of this spending frenzy is artificial intelligence. Training large language models requires massive computing power, and the hyperscalers are racing to build enough capacity to meet projected demand. McKinsey estimates that $7 trillion in data center investment will be required by 2030 just to keep pace. Google, Meta, Microsoft, and Amazon have together spent $112 billion on capital expenditures in the past three months alone.
The scale of this buildout has started to move macroeconomic indicators. Some analysis suggests that AI data center expenditure has now surpassed the total impact from U.S. consumer spending on GDP growth in 2025. That’s an extraordinary shift in the composition of economic activity, driven almost entirely by a handful of technology companies betting on a future that hasn’t fully materialized.
And therein lies the concern. The hyperscalers are spending trillions on infrastructure to serve AI workloads, but concrete business models and tangible returns from generative AI remain elusive. Investors are increasingly demanding evidence that this spending will translate into revenue, and the patience is wearing thin.
The Debt Structures Are Getting Creative and Risky
What’s particularly striking about the 2025 data center boom is how the financing has evolved. The hyperscalers aren’t just issuing corporate bonds. They’re using increasingly complex structures to keep debt off their balance sheets and obscure the true extent of their leverage.
Microsoft, through its AI Infrastructure Partnership with BlackRock, Global Infrastructure Partners, and MGX, deployed the partnership’s first major investment in October: a $40 billion acquisition of Aligned Data Centers, the largest data center deal ever. The clever part? The 70% debt leverage sits at the fund level, not on Microsoft’s corporate balance sheet. The debt doesn’t appear in Microsoft’s 10-K filings.
Meta went even further with Blue Owl Capital, securing $27 billion in debt financing for its Hyperion data center project while keeping the entire obligation off its balance sheet through a sale-leaseback arrangement. Meta gets to lease back the completed 4-million-square-foot facility while someone else technically owns the debt.
These structures aren’t illegal, but they raise legitimate questions about transparency. Howard Marks at Oaktree Capital has questioned whether investors are being adequately compensated for 30 years of technological uncertainty when they buy hyperscaler debt yielding only about 100 basis points above Treasuries.
Power Constraints Are Becoming the Real Bottleneck
Money isn’t the only constraint on data center expansion. Electricity is. AI workloads are extraordinarily power-hungry, and the grid simply cannot deliver enough electricity to support the projected buildout in many locations.
An estimated 10 gigawatts of new data center capacity is projected to break ground globally in 2025, but securing reliable power supplies has become the primary challenge. Meta’s Hyperion project in Louisiana involves a $4 billion gas plant just to power the campus. Google signed a $3 billion hydropower deal to secure 3 gigawatts of renewable capacity.
This power scarcity creates an interesting dynamic. Struta, an S&P analyst, expects that already-built data centers with secured power supplies will become more valuable precisely because new construction faces such severe energy constraints. “I wouldn’t be surprised if already high valuations get even higher,” he told CNBC.
The implication is that the data center market may be heading toward a bifurcation: existing facilities with power become premium assets, while projects still seeking energy connections face delays and cost overruns.
The Bubble Question Won’t Go Away
In Bank of America’s November fund manager survey, 46% of respondents cited “AI bubble” as the top tail risk facing markets. That was down slightly from a record 54% in the prior month’s survey, but the concern has been growing steadily throughout 2025.
JP Morgan Chase estimates that up to $7 trillion of AI spending will ultimately be financed with borrowed money. The bank’s strategists have noted that the question isn’t which capital market will finance the AI boom, but how financing will be structured to access every capital market simultaneously.
There’s a historical pattern here that should concern investors. Special purpose vehicles and off-balance-sheet financing typically proliferate late in credit cycles, not early. The creative deal structures appearing now look more like late-cycle financial engineering than prudent early-stage infrastructure investment.
What Could Go Wrong
The biggest risk isn’t that AI fails to become important. It’s that the infrastructure buildout overshoots actual demand, leaving hyperscalers with expensive facilities they can’t fill and debt they struggle to service.
Oracle’s situation illustrates the vulnerability. The company’s debt-to-equity ratio has reached 500%, and it would need 7.4 years of operating income to pay off its debt at current levels. Investment-grade thresholds typically sit at 2.5 to 3.5 times. Bond markets are already pricing Oracle debt like junk, not because of the absolute debt level, but because of the debt relative to Oracle’s capacity to service it.
The ripple effects could extend beyond the hyperscalers themselves. Smaller companies in the AI ecosystem, like data center REITs and GPU-focused startups, are also borrowing aggressively. Guy LeBas at Janney Capital Management notes that “pro-cyclical lending experiences tend to be the ones that end in tears.”
For now, the money keeps flowing and the data centers keep rising. S&P expects even more robust M&A activity in 2026. But at some point, the AI revolution will need to start generating actual returns rather than just infrastructure spending. When that moment of reckoning arrives, the difference between visionary investment and speculative excess will become painfully clear.
