Buy or rent: purchase price against hourly rental

Break-even at partial utilisation

A card that runs half the time takes twice as long to pay for itself. The same division as above with the hours a card actually runs: months until the rental fee at the median rate card equals the purchase price, at 100% (730 h a month), 70% (511 h a month), 50% (365 h a month), 25% (183 h a month). Rental fee only, as above.

Where the figures come from

Against NVIDIA's own list price, where it publishes one: NVIDIA RTX 5090 retail median 275% of $1,999; NVIDIA RTX 4090 retail median 200% of $1,599. Datacenter cards have no published list price.

About this page

As of November 4, 2026, across the 8 GPUs with both figures, the purchase price equals 5.7 to 15 months of continuous rental at the median rate card (median 9.3 months), and at least 12 months at the cheapest listed rate.

At 50% utilisation (365 hours a month), buying the NVIDIA H200 breaks even in 24 months at the median rate card ($4.54 per GPU-hour), and across the 8 GPUs in 11 to 30 months. Break-even at other utilisation is in the table below.

For every GPU with a retail median from three or more sellers: what the card costs to buy, what it costs to rent by the hour, and how many months of running it around the clock add up to the purchase price. Break-even counts the rental fee only — power, hosting, cooling, idle time and resale value are not in it, so a real payback is longer than the figure shown.

Card by card

• NVIDIA H200: buying costs a median $39,995 (store price) across 6 sellers; renting costs from $2.29 per GPU-hour on Dataoorts (Orion Series) and a median rate card of $4.54 across 56 providers — the purchase price equals about 12 months of continuous rental at that median, 24 at the cheapest.
• NVIDIA H100 PCIe: buying costs a median $30,829 (store price) across 6 sellers; renting costs from $1.98 per GPU-hour on GPU.ai (from) and a median rate card of $2.82 across 10 providers — the purchase price equals about 15 months of continuous rental at that median, 21 at the cheapest.
• NVIDIA RTX PRO 6000 Blackwell: buying costs a median $10,449 (store price) across 3 sellers; renting costs from $0.65 per GPU-hour on Hostinger (from) and a median rate card of $2.20 across 42 providers — the purchase price equals about 6.5 months of continuous rental at that median, 22 at the cheapest.
• NVIDIA L40S: buying costs a median $8,289 (store price) across 5 sellers; renting costs from $0.63 per GPU-hour on CloudRift and a median rate card of $1.42 across 40 providers — the purchase price equals about 8.0 months of continuous rental at that median, 18 at the cheapest.
• NVIDIA A100 80GB: buying costs a median $8,246 (store price) across 4 sellers; renting costs from $0.92 per GPU-hour on Yotta Labs and a median rate card of $1.99 across 63 providers — the purchase price equals about 5.7 months of continuous rental at that median, 12 at the cheapest.
• NVIDIA RTX 6000 Ada: buying costs a median $7,440 (store price) across 18 sellers; renting costs from $0.65 per GPU-hour on GPU.ai (from) and a median rate card of $1.03 across 12 providers — the purchase price equals about 9.9 months of continuous rental at that median, 16 at the cheapest.
• NVIDIA RTX 5090: buying costs a median $5,500 (store price) across 16 sellers; renting costs from $0.50 per GPU-hour on OpenRelay and a median rate card of $0.79 across 23 providers — the purchase price equals about 9.5 months of continuous rental at that median, 15 at the cheapest.
• NVIDIA RTX 4090: buying costs a median $3,195 (store price) across 14 sellers; renting costs from $0.35 per GPU-hour on OpenRelay and a median rate card of $0.48 across 17 providers — the purchase price equals about 9.1 months of continuous rental at that median, 13 at the cheapest.

How to read it

Months to break even = retail median ÷ (rental rate × 730 hours). It answers one question only: at what utilization the rental fee alone exceeds the purchase price. A card that runs a few hours a day never gets there; a card that trains around the clock gets there within the year at median rates. Owning also means power (an H100 draws 350–700 W), a host system, cooling, and the card’s resale value at the end — none of which is in the table. Renting also means the provider’s availability and the tier you actually get: the cheapest rate is usually a community tier, a GPU-only line item or a marketplace floor, each tagged above.

Where NVIDIA publishes a starting price for the card (consumer models only), the last column shows the retail median as a share of it. Datacenter parts are sold through OEMs with no published price, so that comparison does not exist for them. Retail medians come from current listings on tracked sources, each linked on the GPU’s page; a card appears here only with three or more sellers. A store’s own price and a marketplace seller’s asking price are different numbers and are never averaged together; neither is the price a card actually sold for. Rental rates are listed on-demand prices per GPU-hour; see cheapest by provider type for the split.

Cite this page

Published under CC BY 4.0: use and redistribute with attribution and a link. Cite the linked source for a price itself, and cite us for the normalization, the index and the comparison.

Data updated: November 4, 2026