Galaxy Digital's $85M Loss Isn't the Real Story — It's the 9.875% Coupon

Galaxy Digital lost $85 million on its crypto business last quarter. That is not the number worth obsessing over. The number worth obsessing over is 9.875%. That's the coupon on $3.507 billion of senior secured notes issued to finance Galaxy's AI data center expansion. Annual interest expense on that debt: roughly $346 million. Phase I annualized revenue guidance: roughly $320 million. A coverage gap opens before you count a single other cost. I've seen this pattern before. In 2022, while the industry argued about stablecoin design theories, I was mapping Celsius and BlockFi's off-chain exposure to Terra's collapse — not commentary, spreadsheets. We didn't need a leak to know which balance sheets were pumping water. Yields don't front-run in both directions. They just price the risk and wait. Galaxy Digital is now two companies stitched into one SEC-regulated shell. The first is the crypto financial services platform: trading, asset management, strategic investments — the franchise that earned it a seat at the institutional table. The second is a physical-asset operator: data center infrastructure, high-density liquid cooling, power procurement, megawatt math. Phase I of the AI build-out is live. 133 megawatts of critical IT load, fully operational, leased to CoreWeave under a 15-year contract. That's the revenue engine. Management guides to roughly $80 million in quarterly lease revenue from this arrangement — call it $320 million annualized. The Q2 print shows the ramp is still early: data center adjusted gross profit of just $20 million, adjusted EBITDA of $11 million. The run-rate promise, if it holds, is four times current production. Phase II is the bigger, heavier lift: 260 megawatts of expansion, funded partly by that $3.507 billion debt raise, targeted for 2027 handover. From that point, the project company absorbs fixed cash costs — power commitments, maintenance overhead, operational staffing — regardless of whether every megawatt is producing revenue. Construction risk gets codified directly into the income statement. The financing structure tells you what the debt market actually thinks. 9.875% on a 2031 maturity. That's the kind of coupon you see on a stressed asset with execution risk — not on a blue-chip utility. Investment-grade corporates borrow 400 to 600 basis points cheaper. The bond is priced for the possibility of failure. Galaxy is not alone in this pivot. Bitcoin miners from Riot to IREN to Cipher Mining are all selling the same AI-capacity story. VanEck flagged the dynamic: "AI-associated miners trade at premium valuations before most of their leased AI capacity is even delivered." Galaxy's edge is that Phase I is already delivering — but VanEck also named the curse: excess optimism priced into incomplete construction. Now let's do the arithmetic the press release skips over. Annual lease revenue run-rate, Phase I: $320 million. Annual interest expense: $346 million. That's a $26 million deficit at the top line before you pay for anything. Apply the project-level adjusted EBITDA margin — management's own guide of "greater than 90%" — and the EBITDA number lands at roughly $288 million. Still short of $346 million by about $58 million. That's a 22% EBITDA gap. Interest doesn't pause for ramp-ups. It doesn't care whether the narrative is "AI infrastructure supercycle" or "crypto winter." The coupon arrives on schedule. This is where I flag the media symmetry problem. The headline framing — $85 million crypto loss versus $80 million projected AI revenue — is a category error. One is a net loss; the other is top-line gross revenue. They don't offset. The $80 million number isn't profit. Margin assumptions, operating costs, and the rest of the corporate machine all get their cuts before anything services that 9.875% obligation. Now look sideways at the rest of the P&L. The Treasury segment, Galaxy's crypto-facing holdings, tallied a $42 million adjusted gross loss. The AI data center segment generated $20 million of adjusted gross profit. A 2:1 hole. The new engine is producing roughly half the cash damage the old engine is producing — not nearly enough to close the gap. And the dependency structure deserves more scrutiny. Galaxy's own quarterly filing concedes the data center segment is "initially highly dependent" on a single AI infrastructure customer. That customer is CoreWeave. The just-completed $20 billion financing round makes CoreWeave look bulletproof on paper. But CoreWeave is also carrying an aggressive debt load. Stack one leveraged balance sheet on top of another, and the combined structure becomes brittle in exactly the way that forced stress tests in TradFi. Here's the part of the timeline that keeps me alert. Phase II hands over in 2027. Until then, it consumes cash — construction costs, interest, procurement — with revenue back-loaded. From 2027, the project absorbs fixed costs even if delivery is phased. The 260 MW build is double Phase I, and large infrastructure projects don't scale linearly. They scale in complexity: power sourcing, chip availability, liquid cooling deployment, labor. Any slippage pushes revenue further right while the coverage gap compounds left. We didn't need insider information to see any of this. The coupon sheet told us. The question is why the equity market is reading a different document. Here's the decoupling that matters. Equity markets are paying a premium for the AI-crypto story. Wall Street is funding data centers that don't exist yet. Meanwhile, the debt market is demanding a 9.875% coupon on infrastructure that's already partially live. The credit side sees risk. The equity side sees upside. One market will be wrong. My money — professionally and, frankly, personally — sits with the credit interpretation. Yields don't negotiate with narratives. A coupon is a weekly reminder of liability structure. It says: leverage outstrips current cash flow, income is concentrated on one tenant, and construction of the next revenue phase is incomplete. The difference between a 5.5% coupon and a 9.875% is the fudge factor for human error. There is a second decoupling inside the "offset." The crypto loss is booked, realized, in the rearview. The AI revenue is guided, projected, in the windshield. One is accounting truth; the other is a promise. Until Q3 delivers a full quarter at run-rate — actual cash in, not margin guidance — the "offset" is hypothetical. The first empirical test lands in the next earnings report. I'm also watching the 2027 fixed-cost cliff. That's not a distant problem. That's a deadline encoded in a loan agreement. And deadlines in capital structure always win. Q3 is the first real audit. Watch the coverage ratio, not the headlines. Watch whether the $80 million quarterly run-rate becomes repeatable revenue, how fast Phase II burns cash, and whether the 9.875% coupon starts to look like a bargain or a warning. Galaxy is not a bad AI trade. It's early in a good trade with a heavy capital structure and single-tenant risk. The market narrative says the pivot is working. The bond says the wait is expensive. Both can be true. The question is how long the balance sheet can hold them together.