Hook
The buy-to-supply ratio for AI-linked bonds dropped from 4.7x in February to under 2x by July. That’s a 57% collapse in demand within five months. Not a crash. Not a default wave. Just a quiet repricing of enthusiasm. But in the world of structured credit, such signals are the first tremors before the fault line shifts. We watched the leverage unwind in crypto—first UST, then the cascading liquidations on Aave. Now the same pattern is emerging in a market that pension funds and insurers consider "safe." The AI debt boom is not a technology story. It is a liquidity story. And liquidity, as we learned in May 2022, can vanish faster than a model can retrain.
Context
Morgan Stanley has executed nearly $23 billion in AI-related debt underwriting fees in the past six months—surpassing Goldman Sachs and trailing only JPMorgan. The total issuance of AI debt has hit $236 billion, four times the volume from the same period last year. By 2028, Morgan Stanley estimates that $2.9 trillion in capital needs to flow into AI infrastructure to meet compute demand. This is not venture capital. This is Wall Street turning AI compute into a fixed-income asset class. The structures vary: large tech credit wraps (NVIDIA, Google), compute lease securitizations (TeraWulf with Google backing), and private credit off-balance-sheet vehicles (Meta’s $27 billion deal for a Louisiana data center). The buyers are pensions and insurers—the same institutions that bought subprime CDOs in 2006. The product is different. The mechanics are eerily familiar.
Core
Composability is a double-edged sword. In DeFi, composability meant protocols stacking on each other until a single liquidator triggered a chain of insolvencies. In AI debt, the composability is between tech giant credit ratings, long-term compute contracts, and speculative demand for GPU clusters. The TeraWulf example is textbook: a former Bitcoin miner converts its ASIC barns into GPU data centers, signs a support letter with Google (not even a full lease guarantee), and then issues 7.75% unsecured bonds. The bonds are bought 4.7x oversubscribed. Why? Because investors treat Google’s implicit backing as collateral. They are betting on two things: Google’s solvency and the continued demand for AI inference. But Google’s support letter is not a guarantee. It’s a letter. In the 2022 Terra collapse, "support" from the Luna Foundation Guard was also just a promise. Algorithms don’t fail; models do. The model here is that AI compute demand grows at 40% CAGR indefinitely. That’s a strong assumption, especially when the underlying transformer architecture is facing diminishing returns on scale.
The Meta layer extends the analogy further. Meta’s $27 billion private credit deal for its Hyperion campus is executed through a special purpose vehicle (SPV) that keeps the debt off Meta’s balance sheet. This is exactly how Enron hid leverage: off-balance-sheet entities tied to operating leases. The difference is that Meta’s revenue is real and growing. But the structure matters because it separates the credit risk from the tech company’s core debt. If the SPV defaults, Meta is not technically on the hook—unless it chooses to honor the compute contract. This creates a moral hazard. The investors (pension funds) are betting on Meta’s reputation rather than its contractual obligation. Trust is the new currency. But trust has a half-life in financial markets, usually equal to the next earnings miss.
The demand signal decay is the most important data in the article. The drop from 4.7x to 2x oversubscription in five months is a classic risk-off shift. It doesn’t mean the market is crashing; it means marginal buyers are demanding a higher premium for the same risk. Combined with the Oracle CDS cost hitting its highest level since 2009, we see a compression of liquidity in the very instruments that underwrite the AI buildout. Cross-border payments are evolving—but the evolution of finance is not always towards more efficiency; sometimes it’s towards more fragility. The same institutional investors buying AI bonds are the ones that backstopped the repo market in 2020. When they start pulling back, the entire funding chain tightens. In crypto, we saw this in 2022 when market makers withdrew liquidity from DeFi protocols. The result was a repricing of risk that took months to play out.
The physics behind the finance. TeraWulf’s transition from ASIC to GPU is not just a business pivot; it’s a recognition that the physical assets (power, land, cooling) have more value than the digital assets (Bitcoin) they were built for. AI data centers demand power density that rivals small cities. The 2.9 trillion figure implies building the equivalent of 500 large nuclear reactors-worth of data center capacity. That energy must come from somewhere. The AI bond market is effectively securitizing future electricity consumption. If energy prices spike—due to geopolitics or grid constraints—the operating costs of these data centers could exceed the compute revenue. Macro trends ignore micro-hype. The Federal Reserve’s interest rate policy determines the discount rate for these long-duration bonds. A 50-basis-point hike could feasibly wipe out the net present value of a 20-year compute lease. Pension funds buying AI bonds are making a bet on sustained low rates and ever-cheaper renewable energy. Both are questionable.
Contrarian Angle
The consensus view is that AI debt is a prudent way to fund growth, and the demand drop is just a normal cyclical adjustment. The contrarian take: This debt supercycle is actually a sign of strength. By institutionalizing AI infrastructure finance, Wall Street is forcing discipline—projects must pass credit scrutiny, not just hype tests. The buy-to-supply drop is healthy because it means underwriters can’t just dump any deal; they need to structure better. And the Oracle CDS spike reflects company-specific risk (Oracle’s aggressive cloud buildout) rather than systemic AI risk. The real danger is not default but opportunity cost: if the 2.9 trillion is deployed into suboptimal architectures (e.g., data centers built for today’s chips instead of tomorrow’s), then the entire capital base becomes stranded. But that’s a long-term risk. In the short term, the AI debt market is providing a much-needed bridge between tech innovation and patient capital. Composability is a double-edged sword—but it can also be a shield. When structured correctly, these bonds could become a new benchmark for digital infrastructure, much like railroad bonds did for the 19th century.
Takeaway
The AI debt supercycle is the macro story of the decade for digital assets. It is creating a parallel financial system where compute is the underlying asset, and tech giants are the counterparties. The bubble burst, the lessons remain. But which bubble? The AI hype bubble, or the bond market itself? When the music stops—and it will, because leverage cycles always end—the question will be: did the models deliver enough returns to service this debt, or will the debt become the anchor that sinks the scaling ship? Look closer at the liquidity pools. They are flowing through Wall Street, not just on-chain. And that changes everything.
