GPU Infrastructure Financing

GPU financing differs from traditional asset lending because lenders assess both asset value and operator cash generation. Pure asset-backed lending fails because GPU markets are illiquid (bid-ask spreads 20-30%, forced sale recovers 70-80% of list price), resale value collapses quickly (Blackwell $30K today vs. $10-15K in 2027), and hardware alone ignores the real value driver: operational cash flow. GPU debt is hybrid: 60-70% LTV asset-backed (GPU collateral, recovery value $30-40M at year 6), 30-40% cash-flow backed (contracts, CFADS, revenue).

Example $75M cluster financing: $50M senior debt at 70% LTV, 400bps, 5-year tenor (collateral: GPU fleet + operator contracts, secondary recovery $18-20M salvage at 50% rate); $15M mezzanine at 55% LTV, 700bps, 5-year; $10M equity. Lenders expect recovery from operator revenue and DSCR >1.25x. If operator defaults year 4, recovery is 50-100% depending on operational health at default (asset liquidation plus contract assignment/litigation).

Mid-market operators price at SOFR + 300-600bps depending on track record, customer concentration, contract quality, and leverage. Tier-1 (proven revenue, Tier 1-2 contracts, <3x debt/EBITDA) = SOFR + 300-350bps senior, +500-600bps mezzanine. Tier-2 (emerging revenue, mixed contracts, 3-4x debt/EBITDA) = SOFR + 400-500bps senior, +650-800bps mezzanine.

Tier-3 (unproven, spot-heavy, >4x debt/EBITDA) = SOFR + 500-650bps senior (or unavailable), +800-1000bps mezzanine. Tenor is 5-7 years (matching 4-6 year hardware depreciation). 5-year is standard; 7-year available for strong operators; <5 year adds refinancing risk (hardware is half-depreciated in years 4-5, operator needs capital for next cycle). SOFR floating-rate risk is material: 5% SOFR + 400bps = 9% ($4.5M annual service on $50M), rising to 6% costs +$0.5M (11% increase).

Operators hedge via fixed-rate swap (5.50-6.50% fixed, adding 150-200bps upfront). Example sculpted 5-year amortisation (50% principal years 1-3, 50% years 4-5): Years 1-2 = $2.5M principal + $2.0M interest = $4.5M service; Year 3 = $2.5M principal + $1.5M interest = $4.0M; Years 4-5 = $6.25M principal + $0.6M interest = $6.85M annually. Average ~$5M/year.

Mid-market operators (CoreWeave, Lambda, Crusoe, FluidStack) lack balance sheet to fund $50-500M clusters alone. CoreWeave raised $250M+ to deploy $2-3B clusters (requires $250M equity + $500-750M debt + strategic partnerships).

Lenders require sponsor equity injection comfort and strategic anchor guarantees (e.g., Anthropic offtake commits 50% capacity, supporting DSCR). Real structures: FluidStack-Google (Google offtake + capex co-investment; lender trusts Google revenue floor).

Cipher Mining-TeraWulf (crypto/BTC collateral backs loan; cloud revenue is secondary). Crusoe Energy-AMD (AMD 0% financing on accelerators, deferred payment until revenue, reducing upfront capex need). Strategic partners provide credibility, revenue visibility, utilisation-risk reduction, capex co-investment. Standalone operators must raise 30-40% equity vs. 15-20% with strategic backing, significantly diluting investor returns.

Debt covenants are restrictive: minimum DSCR 1.25x+, maximum leverage debt/EBITDA ≤3.5-4.0x, DSRA 6-12 months debt service, MRA 3-6 months capex, capex ≤ depreciation + % EBITDA without lender consent. Covenants lock cash for debt service, reserves, capex; distributions to equity are prohibited whilst debt outstanding. Example $75M cluster: $5M annual CFADS, $5M debt service, $4.5M capex = $14.5M required cash retention annually.

Operator cannot distribute (cash available $5M - $14.5M requirement = negative). If utilisation drops 80% to 70%, CFADS falls to $3.5M, triggering debt service shortfall, covenant breach, and lender acceleration. DSRA (Debt Service Reserve Account) held in escrow ($2.5-5M example) earns minimal rate, typically 1-2% below debt cost.

Lender effectively charges operator for reserves privilege. Mid-market operators find this onerous; large reserves + leverage obligations reduce free cash for growth/distributions, creating equity drag. Smaller operators refinance poorly as hardware depreciates and new clusters deploy. DSRA stays large (based on original debt service), reducing free cash available, throttling growth.

Accelerators depreciate 15-25% annually (Blackwell $30K → $15-20K year 2 → $5-10K year 4). At year-3 of 5-year loan, $75M capex is now $30-40M (50-55% value). Refinancing gap opens: lender will advance 70% LTV on remaining $30-40M = $21-28M, vs. $50M outstanding senior debt.

Gap of $22-29M closes via equity injection, mezzanine (700-900bps), asset/customer sale, or restructuring/haircut. Equipment leasing avoids refinancing risk. Lessor (Blackstone Credit, ArcLight) owns hardware, operator pays fixed lease ($X/year, 5-7 year term matching capex repayment).

Cost: 5-10% premium vs. asset-backed debt, but eliminates refinancing hassle and balance sheet leverage. CoreWeave reportedly explores leasing; Together AI uses debt/equity mix to manage refinancing burden; hyperscalers self-finance (equity/retained earnings) and avoid risk entirely. Fast depreciation also creates equity gap: if debt finances only 60% capex (lender risk aversion), equity funds 40%.

This is >15% standard for infrastructure (power plants finance 70-80%). High equity requirement reduces investor returns and makes mid-market operators dependent on venture capital or strategic partnerships. Lease structures trade capex transparency for refinancing stability; asset-backed financing trades refinancing risk for lower cost of capital. Capital structure choice reflects operator maturity and access to patient capital.

Debt terms directly correlate with revenue quality and contract mix. A cluster with 80% Tier 1-2 (availability/take-or-pay) contracts can finance at 70% LTV, SOFR + 350bps, 5-7 year tenor.

Same cluster with 80% on-demand/spot revenue can only finance at 50-55% LTV, SOFR + 550bps, 4-5 year tenor. The difference in financing capacity is material: $75M capex with 80% take-or-pay financing = $52.5M debt vs. same capex with 80% on-demand = $37.5-41M debt, requiring $13.5-15M more equity (5-8 percentage point return dilution on equity).

This is why contract quality is strategic competitive advantage: operators with anchor customers (Anthropic, Databricks, Google) can fund growth at lower cost of capital, scaling faster than operators without. Real deal structures: (1) CoreWeave reportedly announced 'anchor customers' (OpenAI, Anthropic, others) committing multi-year, multi-billion GPU capacity purchases; these commitments reduce DSCR requirements for lenders, supporting higher leverage and lower cost of debt. (2) Lambda announced $200M Series B funding including credit facility (debt + equity blend); credit facility likely has embedded customer concentration risk (if top 5 customers leave, credit agreement is triggered/breached). (3) Crusoe + AMD deal reportedly includes AMD's credit quality and customer pipeline supporting debt terms. Lesson: revenue quality (as captured in customer contracts, DSCR headroom, revenue concentration) is the single largest determinant of borrowing costs for GPU infrastructure.

GPU debt = 70% asset-backed (hardware collateral, 50-70% LTV) + 30% cash flow collateralised (contracts, 1.25x DSCR), not pure equipment financing

Pricing: SOFR + 300-600bps senior (tier-1 operators 300-350bps, tier-3 500-650bps), SOFR + 500-1000bps mezzanine, 5-7 year tenor

Mid-market operators require strategic backers (Google, hyperscaler offtakes, co-investment) to access debt markets; without backer, equity requirement >30-40% vs. 15-20% with strategic anchor

Fast depreciation (15-25% annually) creates year-3-4 refinancing risk; asset base shrinks 45-50%, refinancing gap must be closed by equity injection or restructuring