GPU Cluster Financial Modelling

GPU cluster financials follow standard project finance discipline borrowed from infrastructure and renewable energy. The four key metrics are Cash Flow Available for Debt Service (CFADS), Debt Service Coverage Ratio (DSCR), Life of Loan Cash Flow (LLCR), and Loan Loss Coverage Ratio (LLCR benchmark).

CFADS represents net operating cash flow after operating expenses, taxes, and working capital adjustments; for GPU clusters, this is utilisation-adjusted revenue minus power costs, staffing, depreciation, and facility charges. DSCR = CFADS divided by annual debt service (principal + interest); target DSCR ranges 1.20-1.40x for mid-market GPU operators, meaning cash flow should cover debt service by 20-40%.

Lower DSCR (1.10-1.20x) is available for operators backed by strategic anchors (hyperscalers, large research institutions, sovereign wealth) or high-utilisation baseline; higher DSCR (1.40-1.60x) required for unproven operators without anchor revenue. LLCR measures ability to repay debt over full loan tenure (typically 5-7 years for GPU infrastructure): sum of discounted CFADS across loan term, divided by outstanding debt. Lenders typically target 1.10-1.35x LLCR, meaning cumulative cash flow should exceed debt by 10-35%. GPU cluster LLCR analysis must account for hardware refresh cycles (every 3-6 years, current-gen Blackwell at $30-40K per accelerator), utilisation decay patterns (year 1-2 baseline, years 3-5 gradual 2-5% annual decline as newer inference workloads migrate), and pricing decay (current-gen commands premium pricing, previous-gen H100 pricing has typically declined 15-25% annually against margin compression).

GPU cluster debt structures employ a cash waterfall (cash cascade) with priority hierarchy: (1) operating expenses and vendor obligations, (2) DSRA (Debt Service Reserve Account), (3) loan principal and interest, (4) MRA (Maintenance Reserve Account), (5) equity distributions. This waterfall protects lenders before equity receives distributions.

Sculpted repayment structures debt maturity to match revenue and utilisation profiles. Years 1-2 show lower principal amortisation (50-60% debt remains) because utilisation and pricing are premium; years 3-5 show higher amortisation (80-95% retired) as utilisation and pricing normalise. Level principal repayment (equal annual amortisation) would create cash shortfall risk in years 3-5.

DSRA typically holds 6-12 months of debt service, funded from equity at close or as percentage of annual CFADS, ensuring one poor utilisation year does not trigger covenant default. MRA holds 3-6 months of capex (hardware refresh, liquid cooling, network cards), preventing operators from raiding cash flow for unexpected equipment replacement.

GPU cluster revenue per MW deployed is a key benchmarking metric: $20-30M/MW/year for mid-case utilisation and blended pricing. At 80% utilisation and $5.00/hour blended pricing, a 10MW cluster generates ~$35M annual revenue.

Margin structure: gross margin (revenue less power) is 70-75%; EBITDA margin (gross less staffing, facility, networking, depreciation) is 35-50% for scaled operators, falling to 15-25% for sub-5MW operators without anchor revenue. For lenders, CFADS margin should target 30-45% at baseline utilisation, enabling 1.25-1.40x DSCR at 300-600bps over SOFR.

Revenue sensitivity analysis must model utilisation downside (70%, 60%), pricing compression (10%, 20%), and competitor entry in years 3-5. Baseline DSCR of 1.30x can slip to 1.00-1.15x under moderate stress.

The Culley Five Golden Rules (developed for infrastructure project finance) translate to GPU clusters as: (1) Get your money back from project cash flow alone. GPU CFADS, without sponsor equity top-ups or asset sale, must cover debt service. (2) Have recourse to sponsor balance sheet if project cash flow fails. Mid-market GPU operators often lack balance sheets; lenders require parent guarantees or sponsor equity cushion.

(3) Understand technology and operational risk. GPU cluster operational risk (utilisation, power, staffing) must be quantified; lenders typically require 3-year operational history or related infrastructure track record. (4) Ensure contract/revenue quality is high. Spot and short-term revenue is lower quality than take-or-pay or multi-year contracts; DSCR requirements increase 15-25% for spot-heavy portfolios.

(5) Enable exit. GPU accelerators depreciate from current-gen to previous-gen in 18-24 months, requiring clear refinancing or resale pathways. Lenders must model salvage value and secondary market for older hardware.

5-year GPU cluster financial models require detailed monthly or quarterly cash flow projections. Key line items: (1) utilisation (80% year 1-2, declining to 70-75% by year 5), (2) blended pricing decay (starting $5.00-$5.50/hour, declining 8-12% annually), (3) power costs ($1.20-$1.50/hour, indexed to regional electricity), (4) staffing ($2-3M annually for 10-20MW cluster), (5) facility and maintenance ($0.50-$1.00/hour opex equivalent), (6) hardware refresh capex (year 3-4, amortised $3-8M).

Sensitivity tables must vary utilisation (-10%, -20%), pricing (-10%, -20%), power cost (+20%, +40%), and competitor entry timing. Stress case models simultaneous utilisation decline (to 60%), pricing compression (20%), and power cost inflation (30%). Lenders typically accept stress-case DSCR to 1.00-1.10x assuming sponsor absorbs shortfall.

Monte Carlo simulation of utilisation and pricing volatility informs probability distribution of DSCR and LLCR outcomes, replacing single-point forecasts.

Target DSCR for mid-market GPU clusters: 1.20-1.40x; higher DSCR required for unproven operators or spot-heavy revenue

Sculpted repayment frontloads principal in years 1-2 (premium pricing, high utilisation) and accelerates amortisation in years 3-5 (margin compression, utilisation decline)

Revenue baseline: $20-30M/MW/year at 80% utilisation; CFADS margin should sustain 30-45% after opex and taxes

Hardware refresh cycles (3-6 years) and pricing decay (current-gen to previous-gen) create refinancing risk; exit strategies critical to lender comfort