Owning a GPU costs the purchase price, the power it draws and the rack it sits in, less what you sell it for. Renting costs the hourly price times the hours you use. This calculator puts the two side by side using today's cheapest and median rental price for the GPU you choose, and tells you the month in which owning catches up, if it ever does.
Last reviewed . Prices on this page are read live and are not part of the review date.
| Option | Rate per GPU-hour | Up front | Per month | Total over 36 months | Owning breaks even |
|---|---|---|---|---|---|
| Buy NVIDIA H100 | $1.75 effective | $32,412 | $202power $66, colocation $137 | $31,589after $8,103 resale | n/a |
| Rent at the cheapest live price (Hyperstack) | $2.50 | $0 | $1,250 | $45,000 | After 24 months |
| Rent at the median live price | $3.29 | $0 | $1,645 | $59,220 | After 17 months |
Over 36 months at 500 hours a month, owning NVIDIA H100 costs $31,589 after resale, against $45,000 to rent at today's cheapest price. Owning breaks even after 24 months and ends $13,411 ahead.
See every current offer on the NVIDIA H100 rent page.
One NVIDIA H100 bought at the CDW reference of $32,412 (retrieved 21 Sep 2026), used 500 hours a month for 36 months, drawing its 700 W TDP at a PUE of 1.4, with electricity at $0.1341 per kWh, colocation at $195 per kW per month, and sold at the end for 25% of what it cost. About that reference price: PNY H100 80GB PCIe card. Marked Discontinued, so this is the last listed price, not an offer. The rental figures in the table are live.
| Option | Rate per GPU-hour | Up front | Per month | Total over 36 months | Owning breaks even |
|---|---|---|---|---|---|
| Buy NVIDIA H100 | $1.75 effective | $32,412 | $202power $66, colocation $137 | $31,589after $8,103 resale | n/a |
| Rent at the cheapest live price (Hyperstack) | $2.50 | $0 | $1,250 | $45,000 | After 24 months |
| Rent at the median live price | $3.29 | $0 | $1,645 | $59,220 | After 17 months |
Over 36 months at 500 hours a month, owning NVIDIA H100 costs $31,589 after resale, against $45,000 to rent at today's cheapest price. Owning breaks even after 24 months and ends $13,411 ahead.
Rental prices observed , USD per GPU-hour, on-demand, in stock.
NVIDIA publishes no list price for data center GPUs, so every figure here is a reseller listing, an analyst estimate or a server price divided by its GPUs. Each carries its seller and date. Out-of-stock and discontinued listings are last-seen prices, not offers. Use them to sanity-check a quote, not as one.
| GPU | Price per GPU | Seller or publisher | Date | What was priced |
|---|---|---|---|---|
| NVIDIA H100 | $32,412 | CDW | retrieved 21 Sep 2026 | PNY H100 80GB PCIe card. Marked Discontinued, so this is the last listed price, not an offer. |
| NVIDIA H100 | $25,000 to $40,000 | Silicon Data | 30 Jan 2026 | Retail listings from mid-2024 to early 2026, from an analysis of thousands of listings. |
| NVIDIA H100 | $23,750 | SemiAnalysis | 20 Aug 2025 | Per-GPU share of an 8x H100 server at about $190,000, hyperscaler volume pricing. Analyst estimate. |
| NVIDIA H200 | $33,732 | CDW | retrieved 21 Sep 2026 | PNY H200 NVL 141GB PCIe card, lead time 4 to 6 weeks or more. |
| NVIDIA H200 | $39,928 | Supermicro eStore | retrieved 21 Sep 2026 | Per-GPU share of an 8x HGX H200 system at its starting price of $319,421.64. Includes the server. |
| NVIDIA H200 | $49,725 | Lenovo Press LP2368 | 24 Jul 2026 | Per-GPU share of an 8x H200 ThinkSystem at Lenovo's usual customer sale price of $397,801.60 (as of 15 June 2026). Includes the server. |
| NVIDIA B200 | $48,029 to $49,301 | Newegg | retrieved 21 Sep 2026 | Per-GPU share of two Supermicro 8x HGX B200 servers listed at $384,228 and $394,406. Both out of stock, so last listed prices. Includes the server and its network cards; B200 modules are not sold alone. |
| NVIDIA B200 | $68,809 | Lenovo Press LP2368 | 24 Jul 2026 | Per-GPU share of an 8x B200 ThinkSystem SR680a V3 at Lenovo's usual customer sale price of $550,475.10 (as of 15 June 2026). |
| NVIDIA B300 | $67,096 | Supermicro eStore | retrieved 21 Sep 2026 | Per-GPU share of an 8x HGX B300 system at its starting price of $536,771.05. Includes the server. |
| NVIDIA B300 | $82,856 | Supermicro eStore | retrieved 21 Sep 2026 | Per-GPU share of a fixed-configuration HGX B300 server at $662,845.95, shown in stock. |
| NVIDIA A100 | $13,224 | CDW | retrieved 21 Sep 2026 | PNY A100 80GB PCIe card. Marked Discontinued, so this is the last listed price. |
| NVIDIA A100 | $19,789 | Newegg marketplace | retrieved 21 Sep 2026 | New PNY A100 80GB PCIe from a third-party marketplace seller. An asking price. |
| NVIDIA A100 | $7,800 to $18,900 | Hashrate Index | 8 Apr 2026 | Used and refurbished listings in early 2026, from an A100 40GB refurbished to an A100 80GB PCIe. Listings, not sales. |
| NVIDIA L40S | $8,500 | CDW | retrieved 21 Sep 2026 | PNY L40S 48GB card, lead time 9 to 11 days or more. |
| NVIDIA L40S | $7,072 | ShopBLT | retrieved 21 Sep 2026 | NVIDIA part 900-2G133-0080-000. Availability not confirmed. |
| NVIDIA RTX PRO 6000 Blackwell | $13,825 | CDW | retrieved 21 Sep 2026 | PNY RTX PRO 6000 Blackwell Server Edition 96GB, the variant clouds rent. Lead time 7 to 9 days or more. |
| NVIDIA RTX PRO 6000 Blackwell | $16,000 | Tom's Hardware | 12 Aug 2026 | Workstation Edition on NVIDIA's US Marketplace after the August 2026 increase. The launch price of $8,565 no longer exists. |
| NVIDIA GeForce RTX 5090 | $6,389 | Tom's Hardware | 14 Sep 2026 | Best US price in the Tom's Hardware tracker. The $1,999 launch price is no longer available at retail. |
| NVIDIA GeForce RTX 5090 | $6,395 to $9,500 | Tom's Hardware | 14 Sep 2026 | Third-party seller asking prices after first-party US retail stock ran out. |
| NVIDIA GeForce RTX 4090 | $3,220 | Tom's Hardware | 14 Sep 2026 | Best US price in the Tom's Hardware tracker (launch price was $1,599). |
Defaults and where they come from:
What is left out, and which way it leans. The host server, CPUs, memory, network cards, switches, storage, spares, support contracts, staff and the cost of capital are all missing from the owning side. A GPU bought inside an 8-GPU server can cost noticeably more per GPU than a bare card, as the reference table shows. On the renting side, storage and egress are missing; see data egress costs. On balance the calculator flatters owning, so a result that favours renting is a firm one.
Rental prices are read live when the page renders and are never typed into the page: the cheapest and the median on-demand price per GPU-hour among offers in stock from secure providers, seen in the last 15 minutes. GPU power figures are in the GPU specs chart. The calculator runs in your browser and sends nothing.
It depends almost entirely on how many hours a month the GPU is busy. A card that runs around the clock for years usually pays for itself; one that runs a few hours a day rarely does, because colocation is billed whether or not the GPU is working. The calculator finds the month where the two lines cross for your utilisation, using today's rental prices.
The purchase price minus what you expect to sell the card for, plus electricity (GPU TDP times PUE times the hours used times your rate) and colocation (GPU TDP in kW times a monthly rate per kW). It deliberately stops there. The host server, networking, spares, staff time and the cost of capital are not included, and all of them make owning more expensive than the figure shown.
NVIDIA does not publish list prices for data center GPUs. What a buyer pays depends on volume and on whether the GPU comes inside a server from Dell, Supermicro, Lenovo or NVIDIA itself. That is why the reference prices next to the purchase field are reseller listings, analyst estimates, or a server price divided by eight, each with its seller and date. Type in your own quote when you have one.
Be honest about idle time. A month has 730 hours. Research clusters and inference fleets with steady demand can stay busy most of those hours; a team that trains in bursts often uses far fewer. Renting only bills the hours you use, so every idle hour moves the break-even further out.
Two live figures for the GPU you pick: the cheapest current on-demand price per GPU-hour, and the median across all current offers. Both count only offers that are in stock, from secure providers, seen in the last 15 minutes. You can also type a quote of your own, for example a reserved rate.
Slightly. TDP is the GPU's power ceiling and real draw under load is usually a little lower. But TDP covers the GPU alone, and the server around it draws more, so as an estimate of what your meter sees it is on the low side overall. You can overwrite the wattage with a measured figure.