Hook
On a Tuesday morning in late 2025, a little-known Bitcoin mining company called TeraWulf closed a bond offering that raised $350 million at 7.75% yield. The offering was oversubscribed 4.7 times. But the real story wasn’t the interest rate or the demand: it was a single page tucked into the offering memorandum—a letter from Google pledging to lease the company’s New York data center for at least seven years. That letter turned a former crypto miner into a quasi-utility for the AI revolution, and it turned a high-risk junk bond into a structured product that pension funds could buy with a straight face. I’ve been tracking mining companies since 2017, when I audited the Tezos code and watched the first wave of ICOs burn through capital. I’ve never seen a financial instrument that so neatly bridges two worlds: the hyper-volatile energy arbitrage of Bitcoin and the relentless capital hunger of artificial intelligence. This isn’t just a deal. It’s a narrative shift that will define the next chapter for both industries. Chasing the alpha through the digital fog, I’m looking at the balance sheets.
Context
To understand why TeraWulf’s bond matters, you have to rewind to 2022. The crypto winter slammed Bitcoin miners: prices dropped below $16,000, energy costs spiked, and many were forced to sell their mined coins at a loss or file for bankruptcy. Core Scientific, once the largest public miner, filed for Chapter 11. The narrative then was simple: miners are leveraged bets on Bitcoin’s price. But a smaller group of miners—including TeraWulf, Cipher Mining, and Hut 8—had something other than ASICs: they had land, power contracts, and cooling infrastructure. During the bear market, I spent six months interviewing operators in Texas and New York for a series I called “Crypto Under the Hood.” I saw empty warehouses with 200-megawatt capacity, sitting idle because Bitcoin mining margins had evaporated. One CEO told me, “We’re not a crypto company. We’re an energy company that happens to mine Bitcoin.” That phrase stuck with me. By 2024, the AI narrative exploded. Every tech company needed compute, and data center space was scarce. Suddenly, those idle power contracts became gold. TeraWulf pivoted, converting part of its upstate New York facility from ASICs to NVIDIA GPUs. The bond market noticed. Morgan Stanley, seeing a new asset class, started packaging these hybrid infrastructure plays into bonds—bundling the credit of Google, Meta, and Oracle with the physical assets of former miners. What emerged is a $236 billion market (as of mid-2025, four times larger than a year prior) that is reshaping the capital structure of both crypto and AI.
Core: The Architecture of the AI Bond
The core insight is that the AI bond is not a single product but a spectrum of risk transfer mechanisms, each designed to pull capital from the traditional fixed-income universe into the high-growth, high-uncertainty world of AI infrastructure. Let me break down three structures I’ve analyzed from the pitch books and offering documents.
Structure 1: The Credit Wrap. In TeraWulf’s case, the bond is issued by a special-purpose vehicle that owns the data center. The cash flows come from a long-term lease with Google. Google doesn’t guarantee the bond, but its commitment to pay rent for seven years provides the cash flow coverage. Moody’s and S&P treat this as investment-grade support. The 7.75% yield is a premium for the risk that Google might walk away (unlikely) or that the facility fails to meet operational specs. This structure is identical to how airlines securitize aircraft leases or how toll roads are financed. The innovation is the underlying asset: a former Bitcoin mine.
Structure 2: The Private Credit Behemoth. Meta, in 2025, raised $27 billion through private placements to fund its Louisiana Hyperion campus. This is off-balance-sheet financing. Meta uses a special-purpose entity that owns the land, buildings, and power equipment; Meta signs a take-or-pay contract for compute services. The debt is senior, secured, and pays around 5.5%—low because Meta’s own credit is effectively behind it. But here’s the twist: the bond buyers are pension funds and insurance companies that cannot hold crypto directly. They are now indirectly exposed to the success of Meta’s AI models. If Meta’s Llama 5 fails to gain traction, the campus still runs, but the margin pressure could strain the lessor’s ability to service the debt.
Structure 3: The High-Yield Frontier. Several smaller players, like Applied Digital and Bit Digital, have issued bonds with yields above 10%. These are pure bets on the AI compute market’s growth. They lack an anchor tenant like Google or Meta. Instead, they rely on spot-market demand for GPU cloud services. The buy-to-supply ratio for these bonds has fallen from 5x in February 2025 to under 2x by July. That’s a warning light. I’ve seen this pattern before: during the 2018 ICO craze, oversubscribed token sales evaporated as soon as Bitcoin retraced. The same psychology is at play here. The credit-default-swap market for Oracle debt—a proxy for tech infrastructure risk—surged to levels not seen since 2009. The market is pricing in a correction before it happens.
Technical Scepticism from a Code-First Perspective: As someone who has audited Solidity and built smart contracts, I can tell you that the risk in these bonds is not the credit rating—it’s the underlying technology assumptions. Every bond assumes that the demand for AI compute grows at a compound annual rate of 40-70% for the next five years. That is based on the continued scaling of large language models (Scaling Laws). If the AI field hits a plateau—if a new architecture reduces compute requirements by 10x—then the demand for these data centers could collapse. The bonds are structured to survive a mild slowdown, not a paradigm shift. I’ve mapped the invisible architecture of value: these bonds are essentially puts on the future of NVIDIA’s GPU roadmap.
Data That Tells the Story: The surge in AI bond issuance correlates with a 40% drop in Bitcoin mining’s share of total U.S. data center power. Miners are selling their ASICs and using the freed capacity for GPUs. But the conversion is not trivial. A Bitcoin mining rig is a single-purpose machine that runs 24/7; a GPU cluster requires complex networking, liquid cooling, and variable loads. The bond market is betting that former mining engineers—many of whom I’ve interviewed—can learn to manage these systems. Some can’t. I’ve seen one facility where the cooling system failed within a week of converting. The bond investors will never know, because the backup generator kept the contract alive. The next failure might not be so lucky.
Contrarian: The Bond That Bites Back
Every narrative has a counter-narrative, and this one is hiding in plain sight. The contrarian view is that the AI bond market is a bubble within a bubble, and it will crash before the underlying tech matures. Let me lay out three blind spots.
Blind Spot 1: The Supply of Capital vs. the Supply of Real Demand. Pension funds have trillions to allocate, but the demand for AI compute could be lumpy. The hyperscalers (Google, Meta, Microsoft) are building their own data centers at a breakneck pace, potentially oversupplying the market by 2027. TeraWulf’s bond depends on Google’s lease, but if Google’s internal capacity hits a surplus, they could exit the lease early (subject to penalties) or renegotiate. The term sheets have break clauses. Investors see the Google name and assume safety, but the structure is more fragile than it appears.
Blind Spot 2: The Energy Trap. Bitcoin miners are experts at curtailing power—they can shut down instantly when grid prices spike. AI data centers cannot. They need steady, reliable 24/7 power. Converted mining sites often lack the redundant substations and uninterruptible power supplies required for AI. Retrofitting costs can equal the original build. The bond proceeds often assume a simple swap, but the capital expenditure for retrofitting can double. I’ve seen projects where the bond’s debt service consumes 80% of the cash flow in the first two years, leaving zero margin for maintenance. One missed GPU delivery could trigger default.
Blind Spot 3: The Cultural Mismatch. The bond market is built on predictability and documentation. Crypto mining, by contrast, is built on resilience, volatility, and flexibility. The two cultures clash. When I interviewed a former miner now running an AI hosting site, he told me, “The auditors asked me for a five-year forecast of GPU utilization. I laughed. In mining, if you have 80% utilization, you’re doing great. In AI, they want 95% or they start asking about margin calls.” The bond market imposes a rigidity that may suffocate the very agility that made these assets valuable in the first place.
My Contrarian Bet: I am not bearish on AI infrastructure itself. I am bearish on the idea that the bond market can accurately price the technology risk. The current spread compression (where risky bonds yield only a few hundred basis points more than Treasuries) reflects a narrative optimism, not a technical reality. When a single model release (like OpenAI’s GPT-5 underperforming) shakes confidence, the re-pricing could be violent. Stories that move money faster than code—and that story is starting to fray.
Takeaway
Anthropology of the tokenized soul: what we are witnessing is the financialization of a belief system. The AI bond market is not just funding data centers; it is funding a future where intelligence-infused compute is the new commodity. The Bitcoin miner, once the denigrated climate villain, is now the unexpected hero of this story—a provider of energy and land in a world hungry for more. But the question that keeps me up at night is not whether these bonds will pay—it’s whether the underlying narrative can withstand the next bear market. When the music stops, who will own the stories? And more importantly, who will own the debt?
