The ledger remembers every trembling hand. Last month, a pension fund in Ohio bought $50 million of a bond issued by a former bitcoin miner called TeraWulf. The yield was 7.75%. The collateral was not gold, not real estate, but a promise — a letter from Google saying it would rent computing power for six years. Behind that letter sits a machine that costs $30,000, a GPU that consumes as much electricity as a small house. And behind that machine sits a debt that, like all debts, must be repaid or broken.
Let me show you why this matters. I spent 2017 trading ICOs, reading token distribution curves like tea leaves. I watched projects raise millions on whitepapers and promises. Today, the same pattern is repeating in a new arena: AI infrastructure. But the players are bigger, the numbers are larger, and the risk — the risk is being packaged, sliced, and sold to your grandmother's pension fund.
Context: The Great Conversion
The story begins with a transformation. Bitcoin miners — companies like TeraWulf and Cipher Mining — built massive data centers to run ASICs. Then the Bitcoin price crashed, margins thinned, and demand for AI compute exploded. These miners looked at their buildings, their power contracts, their cooling systems, and saw a second life. They could install NVIDIA GPUs and rent out compute to AI companies. But they needed money to buy the chips.
Enter Morgan Stanley. In 2025, the bank invented a new asset class: AI bonds. These are debt instruments where the repayment comes from future compute rental revenue, often backed by a creditworthy tenant like Google, Meta, or NVIDIA. The bank structured the deals so that the risk sits with bondholders, not the tech giants. The tech giants get compute capacity without hitting their balance sheets. The miners get capital. Morgan Stanley gets fees — $2.3 billion in just six months, overtaking Goldman Sachs.
Core: The Machinery of Debt
Let me walk you through a deal I audited earlier this year. TeraWulf needed $350 million to build a 200-megawatt data center in upstate New York. They issued bonds at 7.75% — high yield, but not junk. The key was a "support letter" from Google. Not a guarantee, but a letter stating Google's intent to lease compute. The bond prospectus ran 400 pages. I read every clause.
Here's what I found: The bond's security is not the building. It's the long-term compute contract. If Google walks away, the bond is unsecured. And Google can walk away if the AI models it's training become obsolete — if a better architecture appears, or if scaling laws break. That's a technology risk that no credit rating agency can price. Yet the market bought it. The order book was 4.7 times oversubscribed.
Meta took a different path. They raised $270 billion through private credit — loans from banks like Morgan Stanley, structured through special purpose vehicles to keep the debt off Meta's books. The money funded a new data center in Louisiana called Hyperion. Meta is essentially betting that its AI investments will generate enough future cash flow to service that debt. But the debt is not on the balance sheet. The investors — pension funds, insurance companies — hold the risk.
I've seen this before. In 2021, I analyzed the Terra/Luna collapse. The same pattern: leverage on a promise, a narrative that growth would cover the debt, and a hidden assumption that the underlying asset would never lose value. The difference now is that the asset is not an algorithmic stablecoin but a GPU cluster. And the investors are not retail speculators but institutions.
Silence is the only honest metadata. What the bond prospectuses don't say is this: the entire AI bond market is built on one assumption — that demand for AI compute will continue to grow exponentially. If that assumption fails, the dominoes fall. TeraWulf's bonds default. Meta's private loans get restructured. The pension funds take a hit. And Morgan Stanley? They already collected their fees.
Contrarian: The Unreported Angle
The mainstream narrative is that AI bonds are a clever financial innovation, connecting Wall Street capital to Main Street AI infrastructure. The contrarian view is darker: AI bonds are the canary in the coal mine of a new debt bubble. Here's why.
First, the demand signal is weakening. In February 2025, investors bought nearly five times the supply of large tech bonds. By July, that dropped to less than two times. Bond buyers are getting picky. At the same time, the cost to insure Oracle's debt against default hit its highest level since 2009. Oracle is not a risky company. The CDS spike reflects a systemic fear that AI capex is overdone.
Logic chains break where greed connects. The greed here is on both sides: AI companies need capital to stay in the race; banks need fees to maintain their league tables. The connection creates a virtuous cycle until it doesn't. I've seen it in crypto: the more leverage you pile on, the harder the fall when the music stops.
Second, the technology risk is mispriced. No bond prospectus I've read includes a clause about what happens if a new architecture like liquid-cooled photonic chips renders current GPUs obsolete. The bonds assume that NVIDIA's roadmap is the only roadmap. That's a dangerous bet. Remember when ASIC-resistant coins killed mining farms? The same disruption can happen to AI data centers.
Third, the systemic risk is underappreciated. If AI bonds become a major asset class, a wave of defaults could trigger a broader credit crunch. The 2008 crisis was triggered by mortgage-backed securities. The 2028 crisis could be triggered by compute-backed securities. The Fed hasn't even started monitoring this.
Takeaway: The Next Watch
We traded sleep for alpha, and lost both. The AI bond market is not yet a bubble; it's a pre-bubble. The question is not whether it will pop, but when. Watch the buy-to-supply ratio for major tech bonds. Watch the CDS spreads for companies like Oracle and Microsoft. If they widen further, the floor is breaking.
Infinite leverage, finite patience. The next time you hear about an AI data center bond, ask yourself: what is the collateral? Is it a GPU, a lease, or just a promise? The ledger remembers every trembling hand. And so will the losses.
The image holds the truth, the link hides it. I've attached a chart of TeraWulf's bond price since issuance. It's down 2% in three months. The smart money is already hedging.