Rafomac News, Tech & Trends That MatterAI chips like Nvidia’s H100 have plunged in rental value by up to 70% in just two years, yet Wall Street keeps lending against them for decades. This growing AI debt structure carries echoes of the 2008 financial crisis, raising questions about the true risk behind the tech boom.
Why Is Wall Street So Eager to Lend on AI Infrastructure?
Morgan Stanley projects global data centre spending between 2025 and 2028 will hit $2.9 trillion. Of that, $1.6 trillion is earmarked for computing hardware, like GPUs ticking away inside AI servers. But these hyperscalers — the likes of Microsoft, Amazon, Alphabet, and Meta — only generate around $1.4 trillion in operating cash flow to cover their costs.
This leaves a giant $1.5 trillion financing gap big enough to topple most stock exchanges. The answer? Wall Street’s private equity firms, asset financiers, and bond markets are flooding in, lending hundreds of billions to bridge the difference. Private credit lenders, including Apollo and Blackstone, have scaled from near zero to over $200 billion outstanding loans for AI firms. The five largest tech companies hammered out $121 billion in corporate bonds in 2025 alone — nearly four times their annual average for the prior four years. Global AI debt issuance is on track to double again in 2026, approaching $570 billion.
Money is pouring into AI infrastructure, but credit rarely comes without strings attached. Unlike traditional debt sitting plainly on corporate balance sheets, AI borrowing increasingly hides behind complex legal structures designed to keep liabilities off the books.
The Debt Structures Masking Real Risks
Take Meta’s Hyperion Campus in Louisiana — a sprawling $30 billion data facility. Instead of Meta borrowing directly, the debt is held by a special entity called Big Net Investors LLC, designed to be ‘bankruptcy remote’. If this vehicle fails, Meta’s broader finances stay untouched. Blue Owl asset managers control 80%, with Meta holding 20%. Around $27 billion in debt has been issued, mostly anchored by PIMCO’s $18 billion investment. S&P rates this paper an A+, slightly below Meta’s own rating, with a fixed coupon of 6.58% and maturity stretching to 2049 — over two decades away.
Oracle follows suit, managing $248 billion of lease commitments via third-party entities alongside more than $130 billion in direct debt. Credit watchers like S&P and Moody’s flag Oracle’s outlook as negative, with five-year credit default swaps soaring to 16-year highs. Anthropic’s chip-backed debt layers another example: senior debt guaranteed by Broadcom, arranged through a special-purpose vehicle to mask direct obligations.
Behind the scenes, Nikkei’s 2026 investigation exposed the five biggest tech giants holding $1.35 trillion in reported debt — plus another $1.65 trillion locked in off-balance-sheet AI obligations, growing eightfold over four years. Bond investors are swallowing A+ notes backed by leases from companies themselves rated below investment grade. This slicing and dicing of weak credit pools into tranches, with senior slices stamped as safe, eerily mirrors the financial engineering that triggered the 2008 crisis.
Collateral Trouble: The AI Hardware Depreciation Dilemma
All these loans are backed by physical assets — GPUs and servers — but their value is collapsing fast. Operators are depreciating GPUs over 4 to 6 years, but real-world data shows rental income for chips like Nvidia’s H100 falling 50-70% within two years. Amazon even cut the useful life assumptions of some server fleets. That means a multi-year debt maturity often outlives the commercial viability of the asset securing it.
A key nuance emerges from CoreWeave, the largest AI cloud provider. It raised an $8.5 billion term loan in March 2026 rated investment grade thanks to contracts with Meta, seen as a solid counterparty. Yet, just seven weeks later, a $3.1 billion facility backed by less creditworthy clients was rated junk. Moody’s and Fitch ratings reflect tenant strength, not the hardware itself. And these clients often lease the equipment through long-term, multi-billion dollar commitments, sometimes guaranteed by chip vendors themselves.
In August 2026, Nvidia unveiled a platform to mobilise over $500 billion in third-party capital alongside giants like Apollo and Goldman Sachs, pushing AI infrastructure securitisation to new heights. Larry Fink from BlackRock compared this financial engineering step to the birth of mortgage-backed securities in the 1970s. Yet skeptics point out nobody has seen GPU residual values over a full tech cycle — a blind spot in risk modelling.
Bitcoin Miners Morph Into AI Landlords
Publicly traded Bitcoin miners have pivoted to AI leasing to survive spiralling crypto mining losses. Mining costs have hovered between $75,000 and $88,000 per Bitcoin in 2025 and 2026, often surpassing market prices. In response, miners have signed multi-billion dollar, decade- or two-decade AI colocation leases with companies like CoreWeave, Anthropic, and Hut 8.
Core Scientific inked a $10.2 billion AI colocation deal and raised $3.3 billion in notes paying 7.75%. But losses mount: it posted a $347 million net loss in early 2026 and sold significant Bitcoin reserves to fund the AI transition. TeraWulf and Hut 8 have similarly inked enormous lease deals, betting as much as 70% of future revenue on AI contracts.
Converting mining sites to AI-grade liquid cooling costs $8 million to $15 million per megawatt — a huge leap from $700,000 to $1 million per megawatt for mining. Van Eck estimates a $50 billion funding gap on promised builds already. While CoreWeave agreed last year to purchase Core Scientific for $9 billion in stock, helping tidy up obligations, warning signs abound: CoreWeave’s credit default swaps imply roughly a 50% chance of default over five years despite investment-grade ratings on some debt tranches.
What This Means for Investors and Markets
Billions of dollars of AI spending are pushing credit quality lower across tech giants, with $460 billion in direct debt and around $1.2 trillion in lease commitments currently in play. Four US senators recently asked regulators to scrutinise Big Tech’s dependence on opaque debt structures.
Data centers’ share of new commercial mortgage-backed securities is surging, with JPMorgan projecting $30-40 billion annually in AI-related securitisations for the next two years. Morgan Stanley expects $130 billion net issuance over 2026–2028. The pipeline widens but the credit beneath looks weaker.
If you hold Bitcoin, the effects may be indirect. But investors in publicly traded miners or funds containing them face direct exposure to long-term lease receivables backed by rapidly aging hardware. This fragile layering echoes the financial dominoes that triggered past market crashes.
Defaults won’t unravel this web immediately. Instead, losses shift through bankruptcy-remote vehicles to structured securities sold to pension funds and insurers. Each step appears safer, yet the overall risk accumulates where asset values erode faster than debts mature. The AI frenzy’s foundation may soon wobble beneath its own weight.
Is this the start of a new era in tech finance or a dangerous bubble waiting to burst? The echoes of 2008 resonate loudly in the AI infrastructure boom.
Watching how this unfolds promises to be one of the defining market stories of our time.
