In 2025, the cost to insure Oracle’s debt hit levels last seen during the global financial crisis. That same week, Morgan Stanley booked $2.3 billion in fees from AI infrastructure bonds. These two facts are connected by a single, fragile thread: the belief that AI’s growth will outpace its debt service.
Investment banks have discovered a new asset class. They bundle long-term compute contracts from companies like Meta and Google into bonds, then sell them to pension funds and insurers. The pitch is simple: AI is the new electricity, and these data centers are the new power plants. The numbers are staggering. Morgan Stanley alone has raised $236 billion in AI-related debt this year — quadruple last year’s volume. They project $2.9 trillion will be needed by 2028. The fees are juicy: $2.3 billion in six months, catapulting them past Goldman Sachs.
The structure mirrors the crypto lending boom of 2021. Back then, companies like BlockFi and Celsius packaged user deposits into yield products, backed by promises of arbitrage profits. The underlying assets were volatile, but the narrative was smooth. Here, the collateral is compute power — but the cash flows depend on a single variable: the continued appetite for AI training and inference. The debt is secured by physical assets: land, power capacity, GPU clusters. But the real collateral is the future revenue from those GPUs, not the chips themselves. When demand dips, the chips’ resale value plummets. We saw this with ASIC miners during the crypto winter.
The market is already flashing warning signals. In February, investors bought nearly five times the supply of large-cap tech bonds. By July, that ratio had dropped to less than two times. This is not a trivial dip — it’s a structural shift in risk appetite. Meanwhile, the cost to insure Oracle’s debt against default hit its highest point since 2009. Oracle is not a distressed company. It generates $50 billion in annual revenue. The CDS spike reflects a systemic concern: the market is re-evaluating the debt load of even the most stable AI players.
Consider the TeraWulf case. This former bitcoin miner issued $7.75% bonds to build an AI data center in upstate New York. The bond was oversubscribed 4.7 times, partly because Google provided a backstop letter. A 7.75% coupon on a 4.7x oversubscription screams artificial scarcity. Pension funds bought this paper believing Google would never let it fail. But what happens when Google’s own compute needs shift? Based on my audits of institutional custody, I’ve seen how “too big to fail” narratives create a false sense of safety. The truth is, no tech giant has unlimited appetite to absorb losses from partner defaults.
Meta’s $270 billion private credit arrangement for its Louisiana Hyperscale campus follows the same pattern: debt off the balance sheet, hidden from shareholders, backstopped by future earnings. It’s a clever piece of financial engineering, but it introduces off-balance-sheet leverage that mirrors the Enron-era synthetic structures. The difference? Enron’s assets were pipelines; Meta’s assets are GPUs that depreciate 40% annually.
The bulls argue that physical infrastructure has tangible value. A data center can be repurposed for less intensive workloads. Land and power contracts hold value even if AI hype subsides. That’s partially true. But the debt is priced on the assumption of high utilization at premium compute rates. If utilization falls to 60%, the cash flow doesn’t cover interest. The bonds trade at distressed levels, and the pension funds suffer. The contrarian angle: the bond market may be a leading indicator, not a trailing one. When demand for compute contracts weakens, the first sign will be a widening of credit spreads — not a drop in Nvidia’s stock price.
The final irony: the same institutional investors that shunned crypto for being unregulated are now holding AI bonds that are structurally similar to crypto lending products. Both rely on a narrative of exponential growth. Both use complex structures to hide risk. Both create moral hazard by outsourcing due diligence to credit ratings agencies that don’t understand the underlying technology.
Complexity hides the body. The question is not whether this market will correct — it is whether the correction will be slow and manageable, or fast and systemic. Read the code, not the pitch deck. Here, the code is the debt covenants, the compute contract terms, and the rights of bondholders in a default scenario. Few investors have read them. They bought the narrative.
When the compute demand dips, who will be left holding the bag — the tech giants, or the pension funds that trusted the pitch deck?
