A 100 H100 cluster costs $2.5M to $3M in GPU silicon alone at $25,000 to $30,000 per H100 SXM5, and runs to $3.3M to $4.2M fully built when you add 8-GPU HGX servers at $250,000 to $320,000 each. Most companies do not write that check from cash. The financing structure you pick changes the all-in $/GPU-hour, shifts the buy-vs-rent breakeven, and decides whether the asset hits your balance sheet.
This guide compares the five financing paths used by institutional AI buyers in 2026, with the formula you can plug your own cost of capital and lender quote into.
TL;DR
Five financing paths:
Cash purchase. Lowest absolute cost on paper, highest opportunity cost on capital. Best when your cost of capital is genuinely below current equipment finance rates.
Equipment loan. Debt-financed CapEx, asset and matching liability on balance sheet, full ownership at term-end. The default structure for 50+ GPU buyers with stable cash flow.
Capital lease. Economics close to an equipment loan, with ownership flexibility at term-end. Useful for specific tax positioning.
Operating lease. Vendor-funded, lowest upfront cash, highest total cost because the lessor takes residual value risk and prices it in. Partial off-balance-sheet treatment under ASC 842.
Cloud rental. No capital outlay, fully opex, no ops burden. The right default below 20 GPUs or with variable workloads.
The honest median answer for 2026: for institutional deployments above 50 GPUs with sustained utilization above 75%, a debt-financed owned cluster usually beats reserved cloud on cost. Below 20 GPUs or under 65% sustained utilization, reserved cloud rental beats all four owned paths. The financing path does not change whether you should own, it raises the utilization bar at which owning starts to win.
For the buy-vs-rent framework: Buy vs Rent GPUs.
What does it mean to finance AI compute?
Financing AI compute means choosing the capital structure that funds your GPU hardware. The five common paths differ on four dimensions: who owns the asset at term-end, what hits the balance sheet, how the cash leaves the business, and what the cost of money adds to the all-in $/GPU-hour.
| Path | Asset owner at term-end | Balance sheet treatment | Cash pattern |
|---|---|---|---|
| Cash purchase | Buyer | Asset, depreciated over useful life | Full upfront |
| Equipment loan | Buyer | Asset and matching loan liability | Down payment, then debt service |
| Capital lease | Buyer (typically with bargain purchase option) | Asset and lease liability | Periodic lease payments |
| Operating lease | Vendor | Right-of-use asset and lease liability (ASC 842) | Periodic lease payments |
| Cloud rental | Cloud provider | Opex only | Pay-as-you-go or reservation prepayment |
The line between capital lease and operating lease matters less than it did before 2019. Under ASC 842 (US) and IFRS 16 (international), most leases longer than 12 months now create a right-of-use asset and matching lease liability regardless of classification. The "off-balance-sheet" label that operating leases used to carry is now partial: the income statement treatment still differs (lease expense for operating leases vs depreciation plus interest for loans and capital leases), but the balance sheet shows both.
Who offers GPU equipment financing in 2026?
Three categories of providers compete for institutional AI deployments.
OEM finance arms. Supermicro Capital, Dell Financial Services, HPE Financial Services. Offer equipment loans and leases bundled with hardware procurement. Easiest to execute because the financing closes alongside the purchase order. The financing premium tends to fund the OEM's sales advantage, so this path is usually convenient rather than cheap.
Banks and traditional equipment finance lenders. Standard secured equipment loans with typical equipment finance terms: multi-year amortization, a down payment in the 20 to 30% range, and rates that depend on buyer credit and current Treasury benchmarks. Established enterprises see the better end of the spectrum; growth-stage AI companies see tighter collateral terms and higher rates.
Specialty AI infrastructure lenders. A newer category that emerged through 2024 and 2025. Lenders structured specifically around GPU collateral, with an explicit view on H100 and H200 residual value. Useful for buyers that traditional equipment finance doesn't underwrite, or for structures where the lender's willingness to model GPU-specific depreciation matters more than rate.
Operating leases through OEMs price above equivalent equipment loans because the OEM is also taking residual value risk. The lessor absorbs the gap between the contractual residual assumed at term-start and the actual secondary-market price at term-end. Given the depreciation curve described in H100 Depreciation (H100 SXM5 retains roughly 50 to 60% of original value at 36 months, with conservative cases down to 30%), that residual risk is real and lessors price it in.
How does financing change the all-in cost of a GPU cluster?
The standard owned-GPU effective cost formula from Buy vs Rent GPUs and Colocation Economics treats hardware as if you paid cash. The capital-structure-adjusted version adds the cost of money to the hardware amortization line.
// text
Effective $/GPU-hour =
(Hardware_CapEx + Financing_cost - Residual_value)
÷ (Useful_hours × Utilization)
+ Power_$/hr + Colo_$/hr + Ops_$/hr
Where:
Financing_cost = sum of interest expense over financing term
for a loan or capital lease
= (Total lease payments) - Hardware_CapEx
for an operating lease
= 0 for cash purchaseThe intuition: every $1 of capital tied up in GPUs has an opportunity cost equal to your cost of capital. Cash purchase looks "free" of financing cost in the formula, but the opportunity cost is real. Treating cash as if it has a 0% cost of capital is the most common modeling error finance teams make on infrastructure decisions.
Worked example: cash-purchase baseline for a 100 H100 cluster
The cleanest way to use the formula is to compute the cash-purchase baseline first, then layer financing on top with your specific lender quote.
Inputs:
| Variable | Value | Source |
|---|---|---|
| Hardware | 100 H100 SXM5, $2.5M to $3M in GPUs; $3.3M to $4.2M fully built | H100 Depreciation, HGX server pricing |
| Residual at 3 years | 50 to 60% of original (30% conservative, 70% optimistic) | H100 Depreciation |
| Utilization | 70 to 75% | Sustained production workload |
| Power, colo, ops | $0.41/GPU-hour | Colocation Economics base case |
| Reserved cloud comparison | $1.30 to $1.80/GPU-hour | Mercatus GPU Index, 3-year reserved at long-tail providers |
| Time horizon | 3 years | Standard amortization window |
Cash-purchase effective cost:
At 70% utilization on a 3-year horizon, the cash-purchase effective cost for a 100-H100 cluster lands at roughly $2.30 to $2.50 per GPU-hour optimized, including hardware amortization, power, colocation, and operations. The full breakdown is in 100 H100 Cluster TCO.
Layering financing on top:
Any debt-financed path (equipment loan, capital lease, specialty lender) adds interest expense to the numerator of the hardware amortization line. The size of that addition depends on three inputs you control: the principal financed, the effective rate, and the amortization term. Plug your specific lender quote into the formula above.
Operating leases work differently. Instead of adding interest to a known hardware base, the lessor quotes a total lease payment stream that includes hardware cost, the lessor's cost of capital, the lessor's residual risk pricing, and the lessor's margin. Compare the total lease payments to the equivalent equipment loan total payments over the same term to see the all-in premium.
When does each financing option make sense?
Cash purchase: cost of capital below current equipment finance rates
Cash purchase only beats a financed alternative when your cost of capital is genuinely lower than the financing rate you'd pay. For most AI companies in 2026, the cost of capital sits above current equipment finance rates, which means cash purchase is more expensive in opportunity-cost terms even though it looks cheaper on the income statement.
Cash is the right answer when you have idle capital with no better deployment, or when you specifically want to avoid restrictive lender covenants on a strategic deployment.
Equipment loan: the balanced default for 50+ GPU buyers
The standard structure for institutional GPU purchases. Asset and matching liability on balance sheet, predictable monthly debt service, full ownership at term-end. Tax-deductible interest. Compatible with most buyer balance sheets and most lender appetites.
Best fit when you have stable cash flow, the deployment runs at sustained utilization high enough to justify owning, and you want full ownership of the residual asset.
Capital lease: tax and ownership flexibility
Economics are close to an equipment loan. The differences are accounting treatment and term-end options. Some capital leases include bargain purchase options that effectively transfer ownership at a nominal price; others let the lessee return or buy at fair market value. For buyers with specific tax positioning needs, or uncertainty about whether they'll want to keep the asset at term-end, capital leases offer flexibility.
Operating lease: off-balance-sheet, fastest deployment
The right answer when speed and balance-sheet treatment matter more than total cost. The lessor handles procurement, ownership, and residual value risk. The lessee gets working hardware fast and a clean monthly opex line.
Total cost runs above an equivalent equipment loan because the lessor is pricing residual value risk and earning their margin on the deal. Operating lease is the right choice when capital is tightly constrained, when balance-sheet leverage ratios matter (covenant compliance, debt capacity preservation), or when the deployment is genuinely short enough that ownership doesn't make sense but reserved cloud doesn't fit.
Cloud rental: variable demand or under 20 GPUs
Cloud rental wins for fleets under 20 GPUs, variable workloads, sub-3-year horizons, and any deployment where operational simplicity beats marginal cost optimization. Reserved 3-year capacity at long-tail and specialty providers ($1.30 to $1.80 per GPU-hour for H100 in 2026) captures most of the ownership economics with none of the operational burden.
How financing affects the buy-vs-rent decision
The buy-vs-rent breakeven assumes cash purchase. For a 100-GPU H100 cluster, that breakeven sits at roughly 75 to 80% sustained utilization on a 3-year horizon, comparing cluster economics to reserved 3-year cloud capacity from long-tail providers.
Financing shifts the breakeven up. Every dollar of interest expense added to the hardware line raises the effective $/GPU-hour, which means a higher utilization rate is required before owning beats renting. The size of the shift depends on the financing path: equipment loans add a smaller shift than capital leases, which add a smaller shift than operating leases. Operating leases tend to push the breakeven into a range where reserved cloud is the better answer for most institutional deployments.
The implication: if you're going to finance the purchase, you need higher confidence in sustained utilization before owning wins. Operators that drop into operating leases without running the breakeven math often find themselves locked into hardware payments while reserved cloud would have been cheaper at their actual utilization.
For deployments that can credibly project sustained utilization above 80%, equipment loans usually still produce the lowest effective cost. The financing premium is real but smaller than the structural advantage of owning at high utilization. For deployments that can't credibly project that level of utilization, reserved cloud is the right answer regardless of financing path.
How this connects to broader AI infrastructure economics
Financing sits between the asset-level economics and the market structure questions.
- Asset side: H100 Depreciation covers what the underlying asset is worth over time. Residual value drives every financing structure.
- Cost side: Colocation Economics covers the power, space, and operations cost that runs in parallel with the financing payment.
- Decision side: Buy vs Rent GPUs covers the binary choice that financing modifies but does not invert.
The financing question only matters if you're owning. If you're renting, the financing decision is the cloud provider's problem. Reserved cloud reservations are sometimes prepaid (a working capital decision) but rarely require dedicated financing arrangements.
Frequently Asked Questions
How do companies finance AI compute in 2026?
AI compute is financed five ways: outright cash purchase, equipment loans, capital leases, operating leases, and cloud rental. Equipment loans and capital leases are the most common structures for institutional GPU purchases above 50 GPUs. Operating leases sit partially off-balance-sheet but cost more than an equivalent loan because the lessor takes residual value risk. Cloud rental is the default for fleets under 20 GPUs or variable demand.
Can you lease H100 GPUs instead of buying them?
Yes. OEMs (Supermicro, Dell, HPE) offer capital and operating leases for H100 servers, typically over 3 to 5 year terms. Specialty AI infrastructure lenders also fund cluster purchases with structures tied to GPU collateral. Operating leases price above equivalent equipment loans because the lessor absorbs residual value risk on hardware that depreciates 40 to 50% over a 3-year hold.
Who offers GPU equipment financing in 2026?
Three categories: OEM finance arms (Supermicro Capital, Dell Financial Services, HPE Financial Services), banks and traditional equipment finance lenders, and specialty AI infrastructure lenders that underwrite specifically against GPU collateral. OEM finance is the easiest to execute but typically the most expensive. Banks offer the best rates to established enterprises. Specialty lenders fill the gap for buyers traditional equipment finance doesn't underwrite.
Is GPU financing on or off the balance sheet?
Equipment loans and capital leases put the asset and matching debt on the balance sheet. Operating leases under ASC 842 (US) and IFRS 16 (international) also create a right-of-use asset and matching lease liability for leases longer than 12 months, so the off-balance-sheet treatment is partial. The income statement treatment still differs: lease expense for operating leases vs depreciation plus interest for loans and capital leases.
Should an AI startup buy or lease GPUs?
For most AI startups under 50 GPUs with variable workloads, reserved cloud rental beats both buying and leasing. For larger deployments with sustained utilization above 75%, the math swings toward owning, and within owning, equipment loans usually beat operating leases on total cost. The decision hinges on utilization confidence, cost of capital, and whether you want the asset on the balance sheet.
How does financing affect the buy-vs-rent breakeven?
Cash purchase keeps the breakeven at roughly 75 to 80% sustained utilization on a 3-year horizon for a 100-GPU H100 cluster. Any debt-financed path adds interest expense to the hardware line, which pushes the breakeven utilization higher. Operating leases push it higher still because the lessor also prices in residual value risk. The financing path doesn't change whether to own, it raises the utilization bar at which owning starts to win.
Can you finance GPU clusters with venture debt?
Yes. Specialty lenders structure asset-backed GPU debt against H100 and H200 collateral. Venture debt typically carries warrants, which raises the effective cost above the headline rate. For established companies, traditional equipment financing through OEMs or banks is usually cheaper. Venture debt is most useful when the buyer doesn't qualify for traditional secured equipment financing.
Use current GPU pricing in your model.
Mercatus GPU Index publishes real-time H100, H200, and B200 cloud pricing, refreshed daily. Plug current rates into the formula above before you sign a financing term sheet.
Methodology
Hardware cost assumptions ($25,000 to $30,000 per H100 SXM5 in 2026, residual value 50 to 60% at 36 months in the mid case). Server pricing references published HGX 8-GPU server quotes from major OEMs ($250,000 to $320,000 fully built). Reserved cloud pricing references Mercatus GPU Index May 2026 snapshot ($1.30 to $1.80/GPU-hour at long-tail providers, 3-year reserved). Power, colocation, and operations cost references base case ($0.41/GPU-hour). Lease accounting treatment references ASC 842 (US, effective 2019) and IFRS 16 (international, effective 2019). Last verified: 2026-05-19.