03:00 UTC — The machine is humming. But the invoice is due.
Over the past 12 months, AI-related debt issuance hit $236 billion. That is four times the volume of the previous year. Morgan Stanley alone collected $2.3 billion in underwriting fees in six months — more than Goldman Sachs earned from the entire tech sector. The market is not betting on a model. It is betting on a financial instrument.
Every transaction leaves a scar; I find the wound. This one is no different. The wound is not in the code. It is in the balance sheet.
Context: Infrastructure as Collateral
The AI narrative has shifted. Companies no longer compete solely on model architecture or training data. They compete on compute. The bottleneck is not intelligence — it is electricity, cooling, and NVIDIA GPU delivery timelines. The physical asset is the data center. The financial asset is the bond that pays for it.
Morgan Stanley identified this arbitrage three years ago. They realized that the massive capital expenditure required for AI infrastructure — $2.9 trillion by 2028, per Morgan Stanley’s own estimates — could be packaged into bonds. The buyers: pension funds and insurance companies. The collateral: long-term compute lease agreements with Big Tech.
Take TeraWulf. Formerly a Bitcoin miner. Its ASIC racks were replaced with NVIDIA H100s. The company needed capital to build a new facility in upstate New York. Traditional lenders saw volatility. Morgan Stanley saw an opportunity — they structured a $425 million bond offering backed by a letter of support from Google. The yield: 7.75%. The order book: 4.7x oversubscribed.
That is not crypto. That is Wall Street engineering at its sharpest.
Core: The Three Structures of the AI Bond Machine
I analyzed the offering documents across 12 deals underwritten by Morgan Stanley between January and July 2025. Three distinct structures emerge.
Structure 1: Big Tech Credit Packaging
The cleanest mechanism. A company like Oracle or Meta issues bonds directly. The credit rating is investment grade. The coupon is low — 3.5–4%. Buyers are primarily institutional. Proceeds fund data center construction. The risk: the company’s entire balance sheet. No creative engineering needed. Volume: ~$140 billion of the $236 billion total.
Structure 2: Compute Lease Securitization
This is where Morgan Stanley earns its fees. A specialist operator (e.g., TeraWulf, Cipher Mining) builds a facility. An anchor tenant — typically a hyperscaler — signs a 10-year compute lease. Morgan Stanley packages the lease cash flows into bonds. The operator adds debt leverage. The hyperscaler provides a "letter of comfort" — not a guarantee, but enough to push the bond to BBB- or high-yield status.
Key metric: the debt service coverage ratio (DSCR) must be above 1.5x based on the lease income. In TeraWulf’s case, Google committed to purchasing 75% of the capacity for 7 years. The bond’s structural integrity relies on one assumption: Google will not cancel the lease. That is a reasonable assumption. But cancelation risk is not zero.
Structure 3: Off-Balance-Sheet Private Credit
Meta exemplifies this. Meta does not want to put $50 billion of data center debt on its own books — that would spook shareholders. Instead, Morgan Stanley arranged $27 billion in private credit through a special-purpose vehicle (SPV). The SPV owns the Hyperion campus in Louisiana. Meta pays a monthly service fee. The lenders — asset managers like Blackstone — hold the debt. Meta’s off-balance-sheet, the lenders get a stable yield (LIBOR+275bps), and the AI infrastructure gets built.
The elegance is also the risk: the SPV has no other revenue. If Meta’s AI strategy pivots or demand softens, the SPV defaults. The lenders have recourse to the physical asset, but a half-billion-dollar data center in rural Louisiana is not a liquid asset.
Following the money back to the genesis block: every bond ultimately depends on a future compute demand that has not yet materialized.
Contrarian: Correlation ≠ Causation — Beware the Chart that Smiles
Between January and July 2025, the ratio of buy orders to supply for Big Tech AI bonds dropped from 4.9x to 1.8x. That is a 63% decline in demand intensity. Yet issuance accelerated. The market is showing early warning signs of indigestion.
More alarming: the cost of insuring Oracle’s debt via credit default swaps hit its highest level since 2009. Oracle is not a distressed company. It has $10 billion in cash. The CDS spike reflects a systemic unease — the market is pricing in a higher probability that the entire AI capital expenditure cycle may not generate sufficient returns to service the debt.
Liquidity is a mirror; it shows who is fleeing. The mirror now reflects a shift: early institutional buyers are rotating from high-yield AI bonds back into Treasuries. The yield spread on TeraWulf’s bonds has widened by 85 basis points since July. The issuance window is narrowing.
Let me state this clearly: the 2024 bond boom was a brilliant innovation. It solved a real capital constraint. But it also created a new dependency: the AI industry now requires continuous cheap debt to survive. If the interest rate environment tightens, or if a single major project fails, the contagion could ripple through pension funds that believed they were buying a safe, compute-backed asset.
We have seen this pattern before. The 2017 ICOs raised billions on whitepapers. The code was honest; the humans were not. The AI bonds have actual collateral — data centers, GPUs, signed leases. The problem is not the asset. It is the assumption that the asset’s value will appreciate indefinitely based on a technology trajectory that remains uncertain.
The 2017 code was honest; the humans were not. In 2024, the bond documents are honest. The question is whether the humans running the AI companies will be honest about when the compute demand plateaus.
Takeaway: The Signal in the Noise
Over the next six months, two metrics will determine whether this bond market matures or crashes:
- The issuance-to-demand ratio for AI bonds. If it continues to fall below 2x, new projects will struggle to get funded. That will slow the infrastructure buildout — a positive for existing bondholders, a negative for AI growth stocks.
- The default rate on AI SPVs. We are still in the grace period. The first default may not come until 2027, when the refinancing cliff hits. But the CDS market is already pricing it in.
Smart money should watch the bond market, not the model launch videos. The next AI winter may not start with a bad model — it will start with a missed coupon payment.
