Whether to rent or buy GPUs comes down to two numbers: what one GPU costs to own per hour, and the rental rate for the term you would sign. Against an on-demand rate, owning costs less only above a break-even utilization, the share of the year's hours the hardware does useful work. Against a committed term, the renter also pays for every reserved hour, so utilization drops out and the test becomes cost per reserved hour [14].
Built from dated public inputs, owning one HGX B200 GPU costs about $2.43 an hour at full use, or $2.84 with NVIDIA's software subscription, before staff, cluster networking and insurance [15, 16, 18, 24, 25, 27, 29]. Break-even utilization against on-demand and index B200 rates on 24 October 2026 runs from 17% to 44% [3, 8, 9, 36]. Committed B200 rates span about $2.69 to $14.21 per GPU-hour [1, 13].
Every worked number below is our arithmetic and illustrates the method.
Rent or buy GPUs: hourly cost to own
The method borrows the shape of the simple levelized cost of energy from the National Laboratory of the Rockies [34]. A capital recovery factor (CRF) turns a purchase price into an equal yearly charge that repays the price with interest over a set life. For an interest rate i over n years, CRF = i(1 + i)^n ÷ ((1 + i)^n - 1) [34]. Utilization takes the place of the energy formula's capacity factor:
Owning cost per used hour = (price × CRF + fixed yearly costs) ÷ (8,760 × utilization) + variable cost per used hour
This note charges power at full draw every hour, as a fixed cost, so break-even utilization is the owning cost at full use divided by the rental rate. CBRE reports that minimum power floors of 60% to 85% of allocated power are becoming increasingly common in U.S. colocation contracts [28].
Inputs for one HGX B200 GPU, with dates
| Input | Value used | Source and date |
|---|---|---|
| Server price | $443,278 for 8 GPUs plus eight 400G ConnectX-7 cards at $1,815: $57,225 per GPU | Exxact configurator listing, 24 Oct 2026 [18] |
| Life | 5 years | Book lives: WhiteFiber 5, CoreWeave 6 [13, 14] |
| Financing rate | 9.0% | IREN, fixed rate, non-investment-grade GPU financing, Aug 2026 [15] |
| Power draw | 14.3 kW maximum for 8 GPUs, 1.79 kW per GPU | NVIDIA DGX B200 specification [16] |
| Power usage effectiveness (PUE) | 1.54 | Uptime Institute 2025 survey, weighted average [25] |
| Electricity | 9.64 cents per kWh | EIA, U.S. industrial average, Aug 2026, preliminary [24] |
| Colocation | $196.25 per kW-month | CBRE, 250 to 500 kW in primary North American markets, H2 2025 [27] |
| Software, optional | $18,000 per GPU for 5 years | NVIDIA AI Enterprise list price [29] |
| Cost line | Per GPU-hour at full use |
|---|---|
| Capital: $57,225 × CRF 0.2571 ÷ 8,760 | $1.68 |
| Power: 1.79 kW × 1.54 × $0.0964 | $0.27 |
| Colocation: 1.79 kW × $196.25 × 12 ÷ 8,760 | $0.48 |
| Owning cost | $2.43 |
| With the software subscription ($0.41) | $2.84 |
Sources: [15, 16, 18, 24, 25, 27, 29, 34]. The arithmetic is ours.
Treat this as a floor for GPU total cost of ownership. It leaves out network switches and cables, racks and installation, which the configurator does not offer [18], as well as staff, storage and insurance. Where a colocation rate already covers cooling, the PUE uplift counts it twice. What an HGX server price includes covers the hardware scope, and rack power and liquid cooling covers the space bill.
Break-even utilization against on-demand rates
| B200 rate, 24 Oct 2026 | Per GPU-hour | Break-even utilization | With software |
|---|---|---|---|
| AWS p6-b200, on-demand, N. Virginia | $14.24 | 17% | 20% |
| CoreWeave HGX B200, on-demand, North America | $8.60 | 28% | 33% |
| Lambda 8-GPU B200 instance, on-demand | $6.69 | 36% | 42% |
| Silicon Data B200 neocloud index | $6.45 | 38% | 44% |
| CoreWeave HGX B200, spot | $4.26 | 57% | 67% |
Sources: [3, 8, 9, 36]. Break-even is owning cost divided by the rate.
A 2024 study of two research clusters with 24,000 A100 GPUs reported average utilization of 83% and 85% [30]. At 85%, owning costs about $2.85 per used hour. At 30%, about $8.08.
Committed terms: compare cost per reserved hour
CoreWeave's annual report describes its committed contracts as typically take-or-pay, "requiring payment regardless of the level of utilization" [14]. An owner pays regardless too, so compare $2.43 to $2.84 with each committed rate.
| Committed B200 option | Per GPU-hour | What the commitment buys |
|---|---|---|
| WhiteFiber service order, 384 GPUs, 24 months, Jan 2026 | about $2.69, derived | Reported contract, may include services [13] |
| CoreWeave committed usage | from $3.44 | Up to 60% off on-demand, by inquiry [8] |
| AWS EC2 Instance Savings Plan, 3 years | $5.35 to $6.15 | Committed spend, no capacity reservation [4, 5] |
| Lambda 1-Click Clusters, 2 weeks to 1 year | $8.87 to $9.86 | Cluster with InfiniBand fabric and head nodes [9, 10] |
| AWS EC2 Capacity Block | $14.21 | Reserved capacity, up to six months, prepaid [1, 2] |
The WhiteFiber figure is our arithmetic, $18.1 million over 384 GPUs and 17,520 hours [13], and sits inside the owning range once software is counted. The CoreWeave figure applies the full stated discount [8].
A longer commitment is not always cheaper per hour. AWS's B200 Capacity Block rate sits 0.2% below its on-demand rate [1, 3], and Lambda's B200 clusters cost more per GPU-hour than its 8-GPU instance [9]. A Capacity Block reserves instances for up to six months [2], a Lambda cluster adds a 400 Gb/s InfiniBand fabric and three management nodes [10], and a Savings Plan reserves no capacity [5].
A committed term also ends first: a 24-month order covers 40% of a five-year life. The rent-to-value note covers residual value in its full report.
Rental options outweigh ownership costs
Moving any one ownership input across its sourced range shifts B200 owning cost by no more than $0.47 per GPU-hour.
| Input moved | Range from sources | Owning cost at full use |
|---|---|---|
| Server price per GPU | $54,167 to $65,324 | $2.34 to $2.66 |
| Life | 6 to 5 years | $2.20 to $2.43 |
| Financing rate | 6.0% to 9.0% | $2.30 to $2.43 |
| Electricity | 7.21 to 24.12 cents per kWh | $2.36 to $2.82 |
| Colocation | $190 to $235 per kW-month | $2.41 to $2.52 |
Sources: [13, 14, 15, 18, 19, 21, 24, 25, 26, 27]. The base case is $2.43.
Rental rates move far more. B200 rates in these sources span $2.69 to $14.24 per GPU-hour [3, 13], and H100 rates, posted and reported, span $1.40 to $12.29 by provider, term and pricing type.
| H100 rate | Per GPU-hour |
|---|---|
| Azure ND H100 v5, pay-as-you-go, East US | $12.29 |
| Azure, 1-year reservation | $7.87 |
| AWS p5, on-demand | $6.88 |
| AWS Capacity Block | $5.97 |
| Lambda 8-GPU instance | $3.99 |
| AWS Instance Savings Plan, 3 years, all upfront | $2.59 |
| SemiAnalysis 1-year contract index, March 2026 | $2.35 |
| WhiteFiber 24-month order, March 2026, derived | about $1.40 |
Sources: [1, 3, 4, 6, 9, 12, 13]. Viewed 24 Oct 2026 unless dated. Azure reservations are whole-term prices divided by the term's hours.
For Hopper, the report's HGX H200 case costs $2.09 per GPU-hour to own at full use, a 26% break-even against AWS's p5en on-demand rate [3, 17, 19, 20].
SemiAnalysis reports that the majority of GPU rental capacity trades on contracts of six months or longer, negotiated bilaterally and seldom posted [12]. A break-even against on-demand list prices alone leaves them out.
On-premises versus cloud: costs left out
- Staff. The median wage for network and computer systems administrators was $99,130 in May 2025 [33], or $11.32 per hour of the year, spread over the GPUs one person runs.
- Failures. A 2024 report on a 54-day period of a training run on up to 16,000 H100 GPUs counted 419 unexpected interruptions, 58.7% of them from GPU issues [31].
- Fabric and fit-out. The listing leaves these out [18]. IREN's July 2025 purchase of about 2,400 B200 and B300 GPUs for about $130 million included servers, storage, labor, InfiniBand, racks and cabling [21]: about $54,200 per GPU, below the single-server listing, at bulk volume.
- Tax. IRS Publication 946 reports that P.L. 119-21 reinstated 100% special depreciation for certain qualified property acquired and placed in service after 19 January 2025 [32]. This note works before tax.
For a team that buys, Rillor arranges equipment leases, GPU-backed loans and sale-leasebacks, matched to the deployment and to what lenders need to see.
The Rillor Compute Index measures what GPU compute costs to rent and to own under a published methodology. Its RCI Rental series breaks prices out by GPU class, region, term and pricing type, and RCI Hardware by comparable system grade, the two inputs this method compares [35].
Read the full report
The full report, with an H200 worked case, the formula with power as a variable cost, each input's source type and the complete source list, is available as a PDF.