GPU Lease vs Rent vs Buy: Financial Modeling Framework for AI Infrastructure Decisions at Every Scale

GPU acquisition falls into three categories. Buying (Capex) means purchasing GPUs outright at full hardware cost, capitalizing the asset on the balance sheet and depreciating it over 3-5 years. The buyer owns the hardware for its useful life and bears all utilization risk. At mid-2026 pricing, an H200 SXM GPU costs approximately $28,000-32,000 at wholesale, and a B200 NVL costs $45,000-55,000. A full 8-GPU H200 node with NVLink, CPU, and memory totals approximately $280,000-350,000. Buying makes sense when utilization is predictably above 70% and the organization has the capital budget and balance sheet capacity.

Leasing (operating or finance lease) means contracting with a lessor to use GPUs for a fixed term (12-36 months) at a fixed monthly payment. Operating leases keep the hardware off the lessee's balance sheet. Finance leases transfer substantially all ownership risks to the lessee. A 24-month operating lease for an 8-GPU B200 node at mid-2026 rates is approximately $12,000-16,000 per month with a $0 buyout. The total 24-month cost is $288,000-384,000 versus a $400,000-480,000 purchase price. The lease premium covers the lessor's cost of capital and residual value risk.

Renting (on-demand or spot) means paying for GPU time by the hour, day, or week with no long-term commitment. Renting is pure Opex with zero balance sheet impact and maximum flexibility. At ClusterBid mid-2026 spot rates, H200 GPUs rent for $3.07/GPU/hr. A 24-month continuous rental of 8 H200 GPUs costs approximately $1.29 million at spot rates or approximately $646,000 at 12-month reserved rates ($1.54/GPU/hr). The rental premium over buying is the cost of the flexibility to scale down or change GPU types on short notice.

An operating lease for GPU hardware has three financial parameters: the lease term (12-36 months), the implicit interest rate (typically SOFR + 300-500 basis points for GPU lessors in 2026), and the residual value assumption (what the lessor expects the GPU to be worth at lease end, typically 35-50% of purchase price for H100/H200 and 45-60% for B200 given stronger demand). The monthly lease payment equals the GPU purchase price minus the present value of the residual, amortized over the lease term at the implicit rate.

For a 24-month operating lease on a B200 (purchase price $50,000, residual 50% or $25,000, implicit rate SOFR + 400bps), the monthly payment is approximately $1,100-1,300 per GPU. Over 24 months, the total payment is $26,400-31,200, compared to the $50,000 purchase price. The $50,000 buyer spends $50,000 upfront and owns a GPU worth $25,000 after 24 months, for a net cost of $25,000. The lessee spends $26,400-31,200 and returns the GPU. The lease premium of $1,400-6,200 is the cost of balance sheet treatment (no asset, no debt) and residual value protection (if the B200's market value drops below 50% due to Rubin architecture launch, the lessor absorbs the loss, not the lessee).

Rental pricing for GPUs has three tiers. Spot pricing fluctuates with supply and demand. In mid-2026, H200 spot is $3.07/GPU/hr, B200 spot is $5.45/GPU/hr. On-demand (list) pricing is approximately 1.2-1.5x spot: $3.68-4.60 for H200 and $6.54-8.18 for B200. Reserved pricing (1-12 month commitment) is 15-30% below spot: $2.15-2.61 for H200 and $3.82-4.63 for B200. The spread between spot and reserved reflects the provider's capacity risk: spot can be preempted with 2-30 minute notice on most platforms.

The cost-to-rent over 24 months at spot rates is approximately $43,000 per H200 GPU and $76,000 per B200 GPU. For an 8-GPU H200 node, 24 months of continuous spot rental is $344,000 versus a $300,000 purchase price. The $44,000 rental premium buys the ability to stop renting at any time, switch to B200s if the workload changes, or scale down without selling hardware on the secondary market. For workloads running fewer than 12-18 hours per day, rental is almost always more economical than purchase even on a per-hour basis because the GPU is not consuming depreciation while idle.

Buying GPUs capitalizes the asset and depreciates it over its useful life. For financial reporting, GPU servers are typically depreciated over 5 years (IRS MACRS GDS class 00.11 for computers). For internal modeling, a 3-year useful life better reflects the technology cycle: H100 GPUs purchased in early 2025 lose approximately 30-40% of their residual value when B200 launched in late 2025. The annual depreciation for an 8-GPU H200 node at $300,000 is $60,000 per year on a 5-year straight-line basis or $100,000 per year on a 3-year basis.

The utilization risk is the largest hidden cost of buying. If the purchased GPUs run at 50% utilization (common for research teams with variable workloads), the effective cost per GPU-hour is double the 100%-utilization rate. At $300,000 purchase for 8 H200s depreciated over 3 years, the cost per GPU-hour at 100% utilization is $3.56/hr. At 50% utilization, it is $7.12/hr, more expensive than spot rental at $3.07/hr. Most teams under 200 GPUs cannot maintain the 70%+ utilization needed for purchase to beat reserved rental on a cost-per-GPU-hour basis. Teams above 500 GPUs with predictable training pipelines (24/7 model training with scheduled jobs) can achieve the 80%+ utilization that makes purchase the clear financial winner.

The break-even between renting and buying depends on three variables: utilization rate, holding period, and GPU type resale value. For an H200 GPU at $30,000 purchase price with 3-year useful life and 50% residual after 3 years, the effective cost per GPU-hour is $2.85 at 80% utilization (purchase winner over spot) and $5.70 at 40% utilization (rental winner). The crossover point where purchase beats 12-month reserved rental for H200 is approximately 65% utilization. For B200 at $50,000 with 4-year useful life and 55% residual, the crossover is approximately 60% utilization due to the stronger residual value assumption.

Leasing is the best option for teams that need 12-36 months of committed capacity but do not want balance sheet exposure or residual value risk. This describes most AI startups and mid-size enterprises in 2026: they need guaranteed GPU capacity for their core training workload but cannot absorb a $2-5M balance sheet hit or the risk that Rubin architecture will collapse H200/B200 residual values. The lease premium (typically 5-15% above the equivalent buy-then-resell cost) is the price of balance sheet neutrality and residual value insurance.

Buying GPUs adds a fixed asset to the balance sheet and increases leverage if financed. For enterprise teams with $5M+ annual GPU spend, the balance sheet impact can affect debt covenants and return-on-asset metrics. The tax treatment is favorable: Section 179 allows expensing up to $1.22 million of GPU equipment in the purchase year (2026 limit), and bonus depreciation allows 60% additional first-year depreciation for 5-year property. A $5M GPU purchase generates approximately $3.7M in first-year tax deductions, reducing effective after-tax cost by roughly 25% for a 21% corporate tax rate company.

Leasing keeps the asset off the balance sheet for operating leases (ASC 842 has tightened the criteria, but leases with terms under 12 months or with less than 90% of the asset's fair value in lease payments still qualify for off-balance-sheet treatment). Rent is fully deductible as an operating expense with no balance sheet impact. For venture-funded AI startups, renting is strongly preferred because investors evaluate burn rate (Opex) rather than Capex, and a $5M GPU purchase would represent a large fraction of total funding. For publicly traded companies, the balance sheet treatment matters less than the earnings impact: buying is cheaper over 3 years at high utilization, and depreciation expense is predictable and consistent.

For teams running 1-16 GPUs: rent on-demand. The flexibility premium is minimal at this scale, and the overhead of hardware procurement, installation, and maintenance is not worth the 10-20% savings from committing. ClusterBid hourly rental at spot rates provides the lowest total cost for this segment because utilization is inherently variable for small-scale deployments.

For teams running 16-128 GPUs: mix reserved rental for the base load (60-70% of capacity on 1-3 month commitments) with spot rental for burst (30-40%). The reserved rate saves 15-25% versus spot on the base load. Leasing is worth evaluating if the base load is stable for 12+ months and the team wants to avoid spot price volatility. Buying only makes sense at this scale if utilization exceeds 75% and the team has 12+ months of predictable workload history.

For teams running 128-1,000+ GPUs: buy the base load (the equivalent of 60-80% of peak capacity) and rent the rest on reserved or spot terms. A well-structured fleet at this scale has 50-60% owned hardware (depreciated over 3 years), 20-30% leased hardware (12-24 month terms, absorb burst and growth), and 10-20% spot rental (experimentation, variable batch workloads). The ClusterBid marketplace offers all three acquisition paths with unified billing across spot, reserved, and lease contracts.