In February 2025, TeraWulf sold $500 million in bonds at 7.75% yield. The order book was 4.7 times oversubscribed. The buyers? Pension funds and insurance companies. The collateral? A promise from Google to rent computing power that doesn't exist yet.
This is the new alchemy of the AI era. And it smells like 2008.
I’ve spent nine years dissecting crypto hype cycles. I watched ICOs promise the moon with nothing but a whitepaper. I scraped NFT wash trading data in 2021 and proved 40% of volume was fake. But this—this is different. The AI bond machine is not a scam. It’s a financial innovation so elegant that it’s dangerous.
Context: From Bitcoin Miners to AI Landlords
The story begins with an identity crisis. By 2024, Bitcoin mining had become a race to the bottom. Halving cycles squeezed margins. ASIC rigs lost value overnight. Then came the AI boom—a hungry beast that needed data centers, power, and cooling. Companies like TeraWulf, Cipher Mining, and Hut 8 realized their infrastructure could pivot. The same land, substations, and cooling towers that once served SHA-256 could now serve NVIDIA H100s.
Morgan Stanley saw the opportunity first. In 2024, the bank began packaging AI-related debt into bonds. The recipe: take a tech giant’s credit (Google, Meta), combine it with a long-term compute lease from a data center operator, and securitize the whole thing. By mid-2025, Morgan Stanley had underwritten $236 billion in AI bonds—four times the volume of the previous year. Their fee: $2.3 billion, enough to leapfrog Goldman Sachs in investment banking rankings.
Core: The Forensic Teardown
Let me walk through the architecture. I’ve analyzed three structures in detail.
Structure 1: The Tech Giant Backstop
Take TeraWulf’s bond. The issuer is a former Bitcoin miner. The buyer gets a 7.75% coupon. But the real insurance is a “letter of comfort” from Google, committing to lease compute capacity at TeraWulf’s Lake Mariner facility. Google doesn’t guarantee the bond—it just promises to pay rent. The credit rating agencies (Moody’s, S&P) treat this as near-investment grade. The pension fund sees a stable yield with a tech giant’s footprint.
I ran the numbers. TeraWulf’s Lake Mariner has 500 MW of power capacity. Google committed to 200 MW for the first phase. The remaining 300 MW is speculative—unbuilt, unleased. At 7.75% interest on $500 million, TeraWulf must pay $38.75 million annually. If Google’s lease covers, say, $15 million of that, the company needs $23.75 million from unleased capacity. One missed AI cycle, one scaling law plateau, and that gap becomes a cash incinerator.
Structure 2: The Meta Private Credit
Meta’s approach is different. They arranged $27 billion in private credit through Morgan Stanley for their Hyperion data center campus in Louisiana. The key innovation: off-balance-sheet financing. Meta created a special-purpose vehicle (SPV) that owns the buildings and power contracts. Meta then signs a long-term lease with the SPV. The debt never appears on Meta’s balance sheet. Shareholders don’t see the leverage. But the lenders—pension funds—are exposed to Meta’s willingness to pay rent for 20 years.
I check the motive here. Audits check syntax; journalists check motive. Meta’s bet is that the AI compute demand for their LLMs will grow exponentially. If that bet fails, the SPV defaults, and the pension fund holds a half-built data center in Louisiana. The land itself is worth 10 cents on the dollar.
Structure 3: The Speculative Developer
The third structure is the riskiest. Pure-play data center developers borrowing against future tenant commitments. These are the “airdrops” of the bond world—no tech giant signature, just a pro-forma with optimistic utilization rates. The market is already blinking. In February, investors bought 5x the supply of large tech bonds. By July, that ratio fell to 2x. The credit default swaps on Oracle popped to their highest since 2009. Data leaves footprints; hype leaves only dust.
The Numbers That Keep Me Up at Night
- $2.9 trillion: Morgan Stanley’s target for AI infrastructure investment by 2028.
- 4.7x: TeraWulf’s bond oversubscription ratio. Strong demand, but for how long?
- 23x: The multiple of Morgan Stanley’s AI debt sales year-over-year.
- 1x: The number of independent technical audits I could find on these bond structures. Zero. No one is checking if the compute contracts are enforceable if the AI model fails.
I spent 2022 auditing a Layer-2 bridge that ignored a critical overflow bug because VCs pushed for launch. That bridge now has $0 in TVL. The same pattern is repeating here. The AI bond machine operates on faith in scaling laws. But code is law only until someone finds the loophole. In this case, the loophole is a global recession, a new chip architecture, or a model that doesn’t need 1 million GPUs.
Contrarian: What the Bulls Got Right
Let me give credit where it’s due. The bulls argue that AI compute demand is structurally insatiable. They point to Microsoft’s $100 billion in AI capex, Google’s 1 million GPU clusters, and the fact that every Fortune 500 company is scrambling to deploy models. The bonds are backed by real assets: land, power substations, fiber optics. Even if a tenant defaults, the data center has residual value as a modern industrial asset.
They also argue that 7.75% yield in a 4% risk-free world is a compelling trade. Pensions need returns. AI bonds offer a spread that beats 10-year Treasuries without the volatility of equities. And the tech giant involvement (Google, Meta) provides a de facto backstop. If Google’s AI division falters, Google’s search revenue still pays the lease.
The logic is sound—until it isn’t. Every bond market crash starts with a “one-off” event. A miner misses a payment. A developer fails to secure a tenant. The rating agencies downgrade a tranche. Then the leverage cascades. The pension funds that loaded up on these bonds face a liquidity crunch. The SEC investigates. Congress holds hearings. And the term “AI bond” becomes synonymous with “subprime.”
Takeaway: The Accountability Call
Morgan Stanley is not evil. It’s solving a real problem: funding the grid that will power the next industrial revolution. But beneath every whitepaper lies a buried intent. The intent here is to shift risk from tech companies to savers—pensioners, retirees, insurance policyholders—while collecting billions in fees.
As Satoshi might have asked: if the code of the bond contract fails, who bears the loss? The pensioner, not the banker. The AI bond machine is a lever for progress, but it’s also a trap for those who confuse financial engineering with technological innovation.
Truth is not distributed; it is discovered. And I’ve discovered that the next crash will start not in a smart contract, but in a spreadsheet.
Check the chain, ignore the chat. But here the chain is a balance sheet, and the chat is a credit rating. Proceed with extreme caution.
