NVIDIA H100 price: buying vs renting per hour

The honest answer is $25,000 to $36,000 for one H100 card, and NVIDIA does not publish a list price. The low end is the street price Tom's Hardware reported for the PCIe version in August 2023, and the high end is what CNBC said some retailers had asked before April 2023, without naming the version. A distributor still listed it at $35,315.93 on 2026-09-28, and eBay listings went past $40,000 in 2023. Renting the same GPU on QuantaCloud costs $2.59/GPU-hr. At the $2.59 per GPU-hour it cost on 2026-09-27, a card at $25,000 to $35,316 equals 9,653 to 13,635 hours of rental, which is 13 to 19 months of round-the-clock use (our calculation, below). The SXM H100 has no card price I can quote: it ships on 4- and 8-GPU HGX boards inside complete servers.

How much one H100 card costs to buy#

The number that matters is the date on the price, because the same card has been quoted anywhere from $25,000 to $46,000 since 2023. Tom's Hardware noted in February 2024 that NVIDIA does not officially disclose H100 pricing, so every figure below comes from a named, dated source instead.

The sources disagree because they measure different things. The August 2023 figure is a reported street price, the eBay numbers are asking prices from the H100 shortage of 2023 and early 2024, and the 2026 figure is a list price for a card the distributor did not have in stock. All of them are for the card alone, with no server to put it in. My rule is to plan with the current list price and treat anything lower as the result of a negotiation you have not had yet.

What the SXM premium buys#

The SXM H100 costs more and does more. The same February 2024 Tom's Hardware report said the SXM version tends to cost more than the PCIe card, and NVIDIA's specs show where the money goes.

For one GPU serving a model, the PCIe card has the same 80 GB as the SXM part, and the SXM part is faster: 3.35 TB/s of memory bandwidth against 2.0, and 3,958 FP8 TFLOPS against 3,026. The bigger difference is the NVSwitch fabric that lets eight GPUs work as one training node, and you only get it by buying the whole server. If the job fits on one or two GPUs, I would not buy SXM, because the smallest SXM purchase is a 4-GPU board. The H100 NVL sat between the two with 94 GB per card, and CDW shows its PNY listing as discontinued on June 9, 2026. The InfiniBand vs NVLink guide explains what the fabric does inside a server and between servers. SXM vs PCIe and NVLink covers what the form factor means for multi-GPU work.

HGX H100 and DGX H100 prices#

The HGX H100 has no public price I could verify. NVIDIA does not disclose H100 pricing, and the 8x H100 SXM5 servers that distributors list, such as a Supermicro system at SabrePC, show no price, so the number comes from a quote. I am not going to print a number I cannot source. A server quote depends on the CPUs, memory, storage, network cards and support term you choose, which is why an HGX or DGX H100 has no single price.

The market is also moving past H100 servers. Cisco announced the end of sale of its UCS C885A M8 configurations with H100 GPUs on April 24, 2025, set July 19, 2025 as the last order date, and listed an H200 configuration of the same server as a replacement. B200 vs H100 covers what Blackwell changes.

If you are pricing an HGX H100 server, get a written quote for a reserved build too. QuantaCloud builds HGX H100 servers to order: we order and build the hardware to your spec, and the configuration, lead time and terms come back in writing. Send a capacity brief with the GPU count, network and term, and put the written quote next to the server quote you already have.

What renting an H100 costs on QuantaCloud#

An H100 PCIe rents for $2.59/GPU-hr on QuantaCloud, billed by the hour. These are the live configurations:

Prices checked 5 Oct 2026, 04:49 UTC

The hourly price covers a whole VM, not only the GPU. On 2026-09-27 the single-GPU H100 VM came with 20 vCPU, 128 GB of RAM and 1,250 GB of disk in us-midwest-2, running Ubuntu 22.04 with the NVIDIA driver and Docker. Billing runs from a prepaid balance with a $5 minimum deposit. The first hour is charged at launch, each further hour is charged when the previous one is used up, and the unused seconds of the current hour are refunded when you stop. There are no egress or ingress charges. Stopping terminates the VM and deletes its disk, so copy your results off first. The pricing page and the billing docs have the full rules, and the H100 page covers the specs and what fits in 80 GB. SXM H100 servers are not on demand: they are the reserved builds described above.

Break-even: how many rented hours one card buys#

The rule I follow is to divide the purchase price by the hourly rental price, then check how many hours a year the GPU will really be busy. At $2.59 per GPU-hour, the H100 PCIe price on 2026-09-27 (live now: $2.59/GPU-hr), our calculation gives:

The arithmetic for the first row: $25,000 / $2.59 = 9,653 hours. A year has 8,760 hours, so that is 1.1 years at 100% use, 2.2 years at 50% (4,380 busy hours a year) and 4.4 years at 25% (2,190 busy hours). If the live price differs from $2.59, divide by the live number instead.

Those hours are where rent paid equals the price of the card alone, and every cost in the next section pushes the real break-even later. Time matters too. The A100 went from full production in May 2020 to a server maker's end-of-life notice in January 2024, less than four years, and a card used 25% of the time needs 4.4 to 6.2 years to match what renting would have cost. The A100 price guide covers that lifecycle in detail.

What the card price leaves out#

The card price leaves out everything the card needs to run.

The 3,066 kWh is our calculation: 350 W x 8,760 hours. Multiply it by your electricity rate, then add the host server and the cooling, which draw power too. I have not found a reputable, dated source for used H100 prices, so treat resale value as unknown rather than as money you will get back.

Buy, rent or reserve#

My rule is to rent until you can prove the GPU will be busy, then choose between buying and reserving. Renting fits bursty or uncertain work: most single-GPU VMs are running in about 3 minutes (median), and you pay only for the time you use because unused seconds are refunded when you stop. If you are weighing a consumer card instead, the H100 vs RTX 5090 comparison covers when 80 GB matters.

Buying fits when you already have the rack, power, cooling and people, and your measured use keeps the card busy past the break-even in the table above, well inside its useful life.

Reserving fits sustained work where you want dedicated hardware without owning it, or where you need SXM, 8-GPU HGX servers or several nodes on InfiniBand. We order and build the hardware to your spec, and the configuration, lead time and terms are quoted in writing. The reserved capacity page explains the process, and GPU clusters covers multi-node builds.

The decision comes down to one measured number. Rent an H100 PCIe for a week of real work, count the hours it is busy, and put that number into the break-even table. If 80 GB is your constraint, the H200 price guide runs the same arithmetic for 141 GB, and if budget is, the A100 price guide does it for the older generation. If the numbers say you should own the hardware, send a capacity brief before you sign a server quote.