The math didn't add up, but the fees did.
In the first half of 2025, Morgan Stanley booked $2.3 billion in underwriting fees from a single product category: AI infrastructure bonds. That's more than Goldman Sachs earned from all debt capital markets combined. The bank didn't invent a new algorithm or train a single model. It repackaged the risk of building data centers, wrapped it in the credit ratings of Big Tech, and sold it to pension funds and insurance companies.
Every rug has a seam you missed. Here, the seam is the assumption that AI demand will grow linearly forever.
Context: The Hype Cycle Meets the Debt Market
The AI industry has entered its capital-intensive adolescence. Training frontier models now requires clusters of 100,000+ H100 GPUs, each costing $30,000. Data centers consume megawatts equivalent to small cities. The total investment needed by 2028, according to Morgan Stanley's own estimates, is $2.9 trillion. That's not venture capital territory. That's bond market territory.
Wall Street noticed. In 2024, a handful of deals tested the waters. By mid-2025, AI-related debt issuance hit $236 billion — four times the prior year. The structure varies:
- Tech giant-backed bonds: Google, Meta, and NVIDIA lend their balance sheets by signing long-term compute leases, which are then securitized.
- Project finance bonds: Special purpose vehicles (SPVs) built around a single data center, with revenue tied to GPU rental contracts.
- Private credit: Direct loans from institutional investors, often off-balance-sheet for the borrower (Meta's $27 billion deal being the flagship).
The buyer base? Not crypto degens. Pension funds and life insurers — the ultimate stable capital. They saw a 7.75% yield on TeraWulf's bonds (a former Bitcoin miner pivoting to AI) and bought 4.7x the supply.
Hype burns out; structural integrity remains. But what is the structural integrity of a bond backed by a future that hasn't arrived?
Core: Systematic Teardown of the AI Bond Machine
Let's dissect the components. I will do what credit rating agencies should have done: stress-test the assumptions.
1. The Credit Enhancement Mirage
The most common structure: a data center operator (like TeraWulf) issues a bond. To make it investment-grade, a Big Tech company provides a "support letter" or a long-term lease. The bond is then rated based on the tech company's credit. This is the same trick used in mortgage-backed securities — wrapping subprime loans in AAA-rated tranches.
Example: TeraWulf's bonds carried a 7.75% coupon. The security was a lease agreement with Google. But Google's obligation is to pay for compute time, not to repay the bond. If TeraWulf defaults because its data center never gets built (permit delays, power shortages, equipment failure), the bondholders cannot demand payment from Google. They can only claim the physical assets — a half-built facility in a rural New York town. The "support letter" is not a guarantee. It's a letter.
2. The Scaling Law Bet
Every bond issued assumes that the demand for AI compute will grow at a compound annual rate of 50-70% for at least the next five years. This is extrapolated from the previous two years. But scaling laws are not physical laws — they are empirical observations that could break if: - A new architecture (e.g., Mamba, liquid neural nets) reduces compute requirements by 10x. - The market realizes that most consumer AI use cases don't need frontier models — a 70B parameter model is often overkill. - Regulatory restrictions halt data center expansion in key regions.
If demand growth slows to 20% per annum, the utilization rate of these new data centers drops below 60%, and the rental income cannot cover the debt service. The math didn't add up from the start; we just ignored the denominator.
3. The Off-Balance-Sheet Shell Game
Meta's $27 billion private credit deal is structured through an SPV that is legally distinct from Meta. The debt does not appear on Meta's balance sheet. This means: - Shareholders remain unaware of the true capital commitment. - Bondholders rely on an implicit promise that Meta will not let the SPV fail — but there is no contractual obligation. - If Meta's core advertising business weakens, the SPV could be cut loose without triggering a default at Meta.
This is identical to the Enron special-purpose entities. The difference is that here, the underlying asset is compute, not energy contracts. The opacity is the feature, not the bug.
4. The Concentration Risk in NVIDIA
Every AI bond is, in effect, a bet on NVIDIA's roadmap. If NVIDIA delays its next-generation GPU (Rubin, scheduled for 2026), the data centers optimized for that chip will be idle longer than projected. If a competitor (AMD, Intel, or custom ASICs) offers better price-performance, the long-term compute contracts become overpriced relative to spot market. The bond's cash flows are tied to a monopoly supplier — that's not diversification; that's a single point of failure.
5. The Energy Double Bet
Data center bonds also embed an energy price risk. The profit margin of a GPU rental is heavily exposed to electricity costs. A 30% spike in power prices (due to geopolitical events, carbon taxes, or grid strain) could wipe out the operator's margin. Few bond prospectuses include stress scenarios for energy prices. The disclosure is about construction costs and chip delivery — not kilowatt-hour futures.
Contrarian: What the Bulls Got Right
Emotion is the variable that breaks the model. But let's be fair: the bulls have a point.
- Real demand: Enterprise AI adoption is accelerating. Companies like JPMorgan, Walmart, and the U.S. government are deploying AI at scale. The compute requirement for inference alone could dwarf training demand.
- Asset-backed: Data centers are not vaporware. They are physical assets with residual value. Even if the AI bubble deflates, a modern data center can be repurposed for general cloud computing or HPC (high-performance computing). The land and power infrastructure have intrinsic value.
- Credit quality: The bonds backed by Google or Meta directly benefit from the parent company's ability to pay. Unless Google defaults on its own corporate bonds (unlikely), the lease payments should continue. The risk is not credit default; it's structural subordination.
- Liquidity premium: Pension funds need yield. 7.75% on a bond that is effectively backed by a AAA-rated tenant (if structured properly) is a genuine arbitrage in today's low-rate environment. The demand was 4.7x supply for a reason.
The bulls argue that this is not 2008 subprime. It's 1990s telecom fiber — real assets, real demand, but perhaps overbuilt. The difference is that telecom fiber had a clear service to sell; AI compute is a commodity whose price may collapse.
Takeaway: Accountability Call
Risk is not eliminated by ignoring it. The AI bond market will likely produce several high-profile defaults within 24-36 months. The terms are too loose, the projections too rosy, and the leverage too high. When the first major data center operator misses a coupon payment, the entire asset class will reprice. Pension funds will demand higher spreads. Issuance will dry up. And the AI industry's expansion will stall — not because of technology, but because of financial engineering.
The question is not whether the bubble will burst. It's who will be left holding the seam.
