# GPU Rent or Buy?

> A free, static, browser-based calculator that compares the full economic cost of owning an NVIDIA GPU against renting comparable cloud compute — including the market return forgone on committed cash, and the fact that both rental rates and resale values fall over time.

GPU Rent or Buy? was created by [Aaron Ring](https://github.com/cytokineking) and is published under The Compute Desk, the site's editorial persona rather than a news organization. It compares ownership and rental over a user-selected horizon and reports which is cheaper, the break-even month, and the three assumptions that would flip the answer. Every input is adjustable and the whole calculation runs client-side.

Market data comes from a dated, manually reviewed snapshot — currently **2026-08-14**. GPU prices move fast in both directions; if that date is more than a few months old, treat the numbers as historical. The snapshot file carries its own date and methodology note, so it is the authoritative source for freshness.

What the model accounts for: acquisition price, sales tax, shipping, setup, supporting hardware, financing as real monthly cash flow, opportunity cost on every outlay at the month it occurs, workload and idle electricity, seasonal heating credit and air-conditioning penalty, maintenance and repairs scaled per GPU, resale value on a depreciation curve with a floor, rental repricing over time toward a floor, optional commitment discounts, fixed cloud storage, egress, and cloud setup/idle/failed-job overhead.

What it deliberately does not do: predict availability or capacity, use spot or preemptible pricing, apply reserved or committed rates by default, cover non-NVIDIA accelerators, cover markets outside the United States, or benchmark performance. It is not financial advice, a purchase appraisal, or a guarantee that any listed price is currently obtainable.

## Main pages

- [GPU Rent or Buy?](https://www.gpu-rent-or-buy.com/): Interactive calculator, verdict, cost-over-time chart and sensitivity range.
- [Methodology and sources](https://www.gpu-rent-or-buy.com/#methodology): Calculation details, data scope, uncertainty and disclosures.
- [Privacy policy](https://www.gpu-rent-or-buy.com/privacy/): First-party disclosures for calculator data, Cloudflare Web Analytics, advertising, consent storage and visitor choices.

Privacy questions and rights requests: [privacy@gpu-rent-or-buy.com](mailto:privacy@gpu-rent-or-buy.com)

## GPU coverage

Fourteen NVIDIA models, and only these. Anything else — AMD, Intel, TPUs, older NVIDIA cards — is out of scope.

- Data center / SXM: B300 SXM, B200 SXM, H200 SXM, H100 80GB, A100 80GB, L40S, A40
- Workstation: RTX PRO 6000 Blackwell, RTX 6000 Ada, RTX A6000
- Consumer: RTX 5090, RTX 4090, RTX 3090
- Integrated system: DGX Spark

## Public data

- [GPU catalog](https://www.gpu-rent-or-buy.com/data/gpu-catalog.json): Reviewed GPU identities, memory, power draw, installation requirements and relative performance index.
- [Market snapshot](https://www.gpu-rent-or-buy.com/data/market-snapshot.json): Dated U.S. purchase-price bands by condition, plus normalized per-GPU-hour cloud rental observations with provenance. Contains its own snapshot date and methodology note.
- [Electricity lookup](https://www.gpu-rent-or-buy.com/data/electricity-by-zip.json): U.S. ZIP-to-utility average residential rates, from the NREL/Department of Energy 2024 dataset, CC BY 4.0.

## Prefilling the calculator

Assumptions are encoded as URL query parameters, so a scenario can be linked directly. Unspecified parameters use the defaults below.

Example — a used H100 at 30 useful hours per week on a managed cloud:
`https://www.gpu-rent-or-buy.com/?gpu=h100-80gb&condition=used&tier=managed&hours=30&years=3`

- `gpu` — one of: `b300-sxm`, `b200-sxm`, `h200-sxm`, `h100-80gb`, `a100-80gb`, `l40s`, `rtx-pro-6000-blackwell`, `rtx-6000-ada`, `rtx-a6000`, `a40`, `dgx-spark`, `rtx-5090`, `rtx-4090`, `rtx-3090` (default `rtx-6000-ada`)
- `condition` — `new`, `refurbished`, `used` (default `used`)
- `tier` — `any`, `managed`, `hyperscaler` (default `managed`)
- `hours` — useful GPU hours per week, 0–168 (default `24`)
- `years` — comparison horizon (default `3`)
- `qty` — number of GPUs (default `1`)
- `buy` — override purchase price per GPU in USD; setting it switches to manual purchase pricing
- `rent` — override rental rate per GPU-hour in USD; setting it switches to manual rental pricing
- `power` — marginal electricity price per kWh (default `0.19`)
- `rentchange` — annual rental price change, percent; negative falls, positive rises (default `-20`)
- `rentfloor` — rental price floor as a percent of today's rate (default `25`)
- `resalechange` — annual resale value change, percent; negative depreciates, positive appreciates (default `-15`)
- `resalefloor` — resale floor as a percent of purchase price (default `10`)
- `return` — expected annual return on uncommitted cash, percent (default `5`)
- `sell` — `1` to sell at the horizon, `0` to keep the GPU (default `1`)
- `commit` — committed-use discount, percent off on-demand (default `0`)
- `heating` — `resistance`, `heat-pump`, `gas-oil`, `none` (default `gas-oil`)
- `climate` — `cold`, `mixed`, `mild`, `hot` (default `mixed`)

Both price-trend parameters use the same convention: an annual percentage change where negative means the price falls and positive means it rises.

The in-app Share button emits the complete parameter set, including advanced power, financing and supporting-hardware selections. ZIP code and the optional domestic-tranquility reserve are deliberately excluded from shared links.

## Important context

- Purchase prices are reviewed asking-price observations, not completed sales or appraisals. Rental defaults use a mid-range eligible ordinary on-demand observation within the selected provider class, with a p10–p90 sensitivity band.
- Rental observations span community marketplaces, managed GPU clouds, specialty managed providers and hyperscalers. These are kept as distinct operating models, not averaged together.
- Chart shading is deterministic sensitivity analysis across observed price ranges. It is not a confidence interval, error bar or probability forecast.
- Commercial relationships do not determine source inclusion, defaults, rankings or calculator verdicts. The site may use clearly disclosed affiliate links when an independently recommended product or provider offers a suitable program. ZIP codes, quotes and calculator answers are never submitted to an application database. Cloudflare Web Analytics measures aggregate visits and page performance without cookies, local storage, fingerprinting or calculator query strings. Google Analytics is not used. AdSense is not configured, so no advertising tag or consent banner loads. If ads are enabled later, a Cloudflare Worker will return only a regional consent regime without exposing or storing the visitor's country, state or IP in the app; regulated traffic will use the appropriate Google-certified or opt-out flow with Global Privacy Control support.

## Optional

- [Inspiration](https://www.nytimes.com/interactive/2024/upshot/buy-rent-calculator.html): This project is a playful, silicon-flavored reference to The New York Times's "Is It Better to Rent or Buy?" calculator. It is not affiliated with or endorsed by The New York Times.