Over the past seven days, a single financial instrument has attracted more scrutiny than most DeFi protocols’ total value locked: Galaxy Digital’s $3.5 billion senior secured notes, issued to finance two AI data centers in Texas. The coupon is 9.875% — roughly 2.5 times the yield of a 10-year U.S. Treasury. Annual interest alone amounts to $346 million, a figure that dwarfs the operating margins of many mid-cap crypto exchanges. The maturity date is August 2031, but the first principal payment is not due until 2027, contingent on the second facility’s completion.
This is not a token sale. It is not a liquidity mining program. It is a traditional, high-leverage debt structure wrapped in the narrative of AI-x-crypto convergence. Data does not negotiate; it only reveals. And the data here reveals a fragile architecture: one that depends on flawless construction timelines, continuous AI demand growth, and the absence of cross-contamination from Galaxy’s volatile crypto trading book.
Context: The AI Infrastructure Gold Rush
The crypto industry has spent 2024 and 2025 pivoting toward physical infrastructure. DePIN projects like Akash Network and io.net promise decentralized compute, but their total addressable GPU capacity remains fractional compared to hyperscale data centers. Galaxy Digital, led by Mike Novogratz, chose a different path: partner with CoreWeave, a cloud provider specializing in GPU-accelerated workloads, and build two facilities in Texas with a combined critical IT load of 260 megawatts and a total utility capacity of 400 megawatts.
The funding vehicle is Galaxy Helios Data Centers II LLC, a special-purpose entity that isolates the project’s assets and liabilities from Galaxy’s core operations. The debt is secured by first-priority liens on the project’s assets — land, buildings, power infrastructure, and GPU clusters. According to the offering documents, approximately 20% of the proceeds are reserved for a debt service reserve account, ensuring interest payments during the construction phase. The remaining 80% will fund the first phase, scheduled for completion in the first half of 2026, with the second phase commencing delivery in the first half of 2027.
At first glance, this structure appears conservative: secured by hard assets, with a reserve buffer. But a forensic examination reveals multiple layers of uncollateralized risk, masked by the optimistic narrative of AI ubiquity.
Core: Systematic Teardown of the Debt Structure
Let us begin with the leverage ratio. $3.5 billion in debt against a project that, at completion, is expected to generate annual revenue from compute leases. Assuming a conservative 70% utilization rate and an average rental price of $2.50 per GPU-hour (halved from current spot rates), the facility would need to host approximately 100,000 H100-equivalent GPUs. At current market prices, that GPU inventory alone exceeds $3 billion — meaning the debt fully covers the hardware, leaving zero margin for construction cost overruns, which are endemic to large-scale infrastructure projects in Texas. The state’s grid, ERCOT, has a documented history of instability; the additional 400 megawatts of load will likely require dedicated substations and backup generators, adding millions in unbudgeted expenses.
Second, the interest coverage ratio. Galaxy Digital must pay $346 million annually from the reserve account or from its own cash flows starting immediately. The reserve account covers roughly the first two years of interest, assuming no other draws. But by the time the second phase begins in 2027, the reserve will be depleted. At that point, the project must generate positive cash flow from the first facility to cover the debt service on the entire $3.5 billion. If the first phase is delayed by even six months — a common occurrence in data center construction — the interest payments will drain the reserve, forcing Galaxy to either inject fresh capital or default on the notes. The offering documents acknowledge this risk in legal boilerplate, but the market has priced it as a remote possibility. Historical data suggests otherwise: according to industry reports, 45% of large-scale data center projects exceed their original budget by at least 20%, and 30% suffer delays exceeding one year.
Third, the repayment mechanism. The notes are structured with an initial annual amortization rate of 4%, meaning the first principal repayment of $140 million is due in 2028, one year after the second phase begins. If the second phase is not operational, the amortization schedule can be adjusted, but only with the consent of a majority of noteholders — a process that could trigger a liquidity crisis and forced restructuring. This is not a hypothetical scenario. In 2022, a similar debt issuance for a crypto mining facility in Texas defaulted within 18 months due to rising energy costs and falling Bitcoin prices. The difference here is that the collateral is AI compute, which is perceived as more resilient. But AI compute demand is a function of venture capital funding into AI startups — a cycle that peaked in 2024 and is already showing signs of contraction. Data does not negotiate; it only reveals the thin line between bullish conviction and financial recklessness.
Fourth, the counterparty concentration. CoreWeave is the sole operator; Galaxy is the sole owner. If CoreWeave suffers an operational failure or loses its key customers (reportedly including Microsoft and OpenAI), the revenue stream collapses. The notes do not specify any minimum revenue guarantees or take-or-pay contracts with end users. This is a single-asset, single-tenant risk, wrapped in a special-purpose vehicle. Institutional investors have bought these notes based on the reputation of the sponsors, not on audited financial projections of the underlying project. In my eighteen years of industry observation, I have seen such structures unravel quickly when the narrative shifts from growth to survival.
Contrarian: What the Bulls Got Right
To be balanced, one must acknowledge the argument for this deal. The bulls contend that AI compute demand is structurally undersupplied for at least the next five years. They point to hyperscalers like Microsoft and Amazon doubling their capital expenditures on data centers in 2025, each exceeding $50 billion annually. Galaxy Digital, by securing a 260-megawatt facility with pre-committed GPU orders, positions itself to capture a slice of that demand, with a potential annual EBITDA of $800 million to $1.2 billion by 2028. At 5x EBITDA, the project’s equity value could exceed $5 billion, offering a significant upside to Galaxy’s shareholders.
Furthermore, the interest rate of 9.875%, while high, is lower than what most crypto-native investors would accept for a comparable risk. Institutional bond buyers, accustomed to negative real yields for years, saw this as an attractive risk-adjusted return in a rising rate environment. The fact that the notes are senior secured, with a first-priority lien, provides a legal cushion that most crypto lending protocols lack. In a worst-case liquidation, the GPUs alone could fetch 60-70% of their original cost, covering a significant portion of the principal.
Finally, Galaxy Digital’s management has a track record of executing large-scale projects. Mike Novogratz raised over $1 billion for his crypto merchant bank. CoreWeave’s founders have deep roots in the traditional cloud industry. The combination of crypto capital and traditional infrastructure expertise is a bridge that few have successfully built. If this project delivers on time, Galaxy will have proven that crypto can fund real-world assets without resorting to tokenization or speculative ICOs.
Takeaway: The Accountability Call
The deal closes. The checks are cashed. The GPUs are ordered. But the true test begins in 2027, when the first principal payment is due. If the AI demand cycle holds, this will be remembered as a masterstroke. If it falters, the $3.5 billion will join the growing list of overleveraged infrastructure projects that regulators use to justify tighter oversight of crypto-related lending.
The question that remains unanswered is not whether the data centers will be built — they likely will — but whether the capital structure can absorb the inevitable shocks of construction delays, energy price spikes, or a slowdown in AI venture funding. The bondholders are betting on the narrative, not the numbers. Data does not negotiate; it only reveals. And it will reveal the truth in the quarterly reports two years from now.
