
The $60 Million Gap: Galaxy Digital's AI Pivot Is a Leverage Story, Not a Revenue Story
NeoBear
Galaxy Digital lost $85 million on crypto in Q2. Management projects $80 million in quarterly AI data center revenue. The symmetry is seductive. Headlines write themselves: crypto losses, offset by AI income. The numbers even line up — roughly $85M out, $80M in per quarter.
The symmetry is wrong.
One number is a net loss. The other is top-line revenue. They occupy different rows on the income statement, and they do not offset. But the pairing reveals something the market still is not fully pricing: Galaxy Digital is no longer a crypto financial services firm with an AI side project. It is a leveraged infrastructure operator with a crypto business attached. And the debt math is tighter than the "AI pivot" narrative admits.
SEC filings do not lie. They just require careful reading. Galaxy's Q2 filing confirms a transition that has been building for eighteen months. Phase I of its AI data center program is fully operational: 133 megawatts of critical IT load, leased to CoreWeave under a fifteen-year term. Revenue is accruing. Management guides to roughly $80 million per quarter from the lease at full run-rate.
Galaxy built its name as a full-spectrum crypto financial services platform: trading, asset management, investment banking, all under one roof. The AI data center operation, run through the Galaxy Helios subsidiary, has a completely different cost structure. This is not a software pivot. It is a physical buildout where land, power, and cooling systems are scarcer than code.
Phase II will more than double that footprint to 260 megawatts of additional capacity. The financing is in place: $3.507 billion in senior secured notes, priced at a 9.875% coupon, maturing in 2031. Galaxy Helios II LLC is the issuer and guarantor. The notes are secured by project assets and pledged equity.
That coupon is the most honest number in the entire report. Let me be precise about why.
I have spent enough years tracing smart contract state transitions to know that a flow chart is not a balance sheet. When the code bleeds, only the ledger survives. Apply the same discipline to Galaxy's AI project ledger.
Projected quarterly lease revenue: $80 million. Annualized: $320 million. That is gross contract income.
Management guides to a project-level adjusted EBITDA margin above 90%. The company is explicit that this is a guidance figure, not an audited result, and it excludes management fees. Even taking it at face value, project-level EBITDA comes to roughly $288 million per year.
Now the fixed charges. $3.507 billion in notes at a 9.875% coupon produces annual interest expense of approximately $346 million.
The gap is about $58 million per year. Before management costs. Before treasury segment losses. Before corporate overhead. The project's own EBITDA cannot cover its own debt service. Something else must fill that spread.
This is the calculation the market's "crypto loss versus AI revenue" framing ignores. The $80 million quarterly figure is gross contract value. It is not profit. It is not free cash flow. It is the starting point of a cost cascade that ends below zero.
The project-level EBITDA margin also assumes no downtime, no tenant credit events, and no renegotiation pressure over fifteen years. In my experience running DeFi yield positions since the Uniswap V2 migration days, every assumption that smoothes a model is a risk hiding in plain sight.
The structure adds another layer. CoreWeave is a single tenant. Galaxy's filing explicitly acknowledges that the data center segment is "initially highly dependent" on this one AI infrastructure client. A fifteen-year lease gives revenue visibility. It also concentrates counterparty risk into one name. CoreWeave just closed $20 billion in new financing, which keeps that engine running. It also confirms where institutional capital is flowing: out of Bitcoin exposure and into AI compute. Galaxy benefits from that rotation and feeds it at the same time.
Phase II compounds execution risk. The 260 MW expansion is due for handover in 2027. Construction complexity scales non-linearly with project size. The fine print states that the project company will begin incurring fixed cash costs from 2027 as milestones come due. If delivery slips, those costs continue regardless. Lease revenue does not start until capacity is live and accepted by the tenant. There is no partial credit for a building that is 80% complete.
Scale the capital math out further. High-density liquid-cooled data center capacity typically costs $25-30 million per megawatt to build. Phase II's 260 MW could therefore require $6.5-7.8 billion in total construction funding. The $3.507 billion in notes covers part of that bill at best. The remainder implies additional debt or equity somewhere down the road. That is not a hypothetical. That is a forward capital call written into the project's physics.
The bond market's pricing tells the rest of the story. A 9.875% coupon on senior secured paper is not the pricing of a utility-grade asset. It is the pricing of meaningful credit risk. Investment-grade corporate debt trades in the 4-6% range. The spread here is enormous. Note buyers are demanding compensation for duration, for construction execution risk, for single-tenant concentration, and for the possibility that AI infrastructure demand normalizes before Phase II goes live.
In 2022, I built a Python script to monitor on-chain liquidation thresholds across Aave and Compound after Celsius froze withdrawals. The lesson was not about the code. It was about leverage. Counterparty risk in lending markets is a function of the debt stack, not the quality of the promise. The same logic applies to physical infrastructure. The tenant's balance sheet matters less than the debt service schedule on the notes.
The interim numbers do not support the offset narrative either. Q2 adjusted gross profit from AI data centers was $20 million. The treasury segment posted a $42 million adjusted gross loss. Even on a segment basis, the AI business is a small positive contributor standing next to a much larger negative one. The $85 million net loss and the $80 million quarterly revenue guidance were never meant to be netted against each other.
The conventional reading says Galaxy is diversifying out of crypto volatility into stable, contracted infrastructure cash flows. I think the opposite. Crypto trading has known, measurable variability. The AI business replaces that with a different, arguably worse risk profile: enormous fixed costs, a multi-year construction period, and income that depends on one tenant's willingness to keep paying for fifteen years.
The "AI revenue offsets crypto losses" narrative is wrong on both sides of the equation. The $85 million is a net loss after all expenses. The $80 million is lease revenue before financing costs. They operate in different accounting dimensions. When you correct the mismatch, the story is not diversification. It is a leveraged bet that Phase II delivers on schedule and that CoreWeave remains solvent for the life of the lease.
The 9.875% coupon quietly confirms this. If the bond market had high confidence in the delivery schedule, the rate would be lower. The rate is the market's honest assessment of execution risk. That is not a whisper from a research desk. That is a yield spread on a $3.5 billion liability.
VanEck's warning about AI-linked miners receiving premium valuations before delivering most of their leased capacity applies here. Galaxy has actually delivered Phase I, which puts it ahead of the peer set. But equity markets may already be discounting Phase II success. When the debt costs nearly ten percent, the margin of safety is thin.
Q3 is the verification event. Management guided to roughly $80 million in quarterly run-rate revenue from Phase I. The actual reported numbers will show whether the leases are performing as contracted. That is the test. Watch the lease revenue line and the cash flow statement, not the press release.
I do not trust whispers; I trust verified hashes. Read the next quarterly report the way you would audit a reentrancy vulnerability: trace the path, check the math, and position for outcomes, not narratives. Yield is the shadow cast by risk taken. At 9.875%, that shadow is long.