The NVIDIA H100 price for a single 80GB card runs from about $25,000 to $30,000 for the PCIe version and $35,000 to $40,000 for the SXM version. A full DGX H100 system with eight GPUs costs several hundred thousand dollars. Renting is much cheaper to start: one verified cloud listing shows an H100 80GB from about $2.40 per GPU hour.
The H100 is NVIDIA's Hopper-generation data center GPU. The common version carries 80GB of memory and is built for training and serving large AI models. NVIDIA does not publish one fixed list price. Resellers and system builders set the price, so quotes differ by form factor, vendor, volume and supply.
This article covers what drives the cost, what each configuration costs, how to compare buying with renting using a worked example, and what to check before you accept a quote. Geodd offers H100 capacity by the hour, so we know the rent side of that math well.
Why the H100 price varies so much
The H100 price varies because the card is a supply-constrained part sold in several form factors through resellers, not at a single list price. Two buyers asking for an "H100" in the same week can get quotes $10,000 apart and both be correct.
What makes the H100 expensive
The card combines a large die, high-bandwidth memory and advanced chip packaging, and demand from AI labs and cloud providers has outrun supply for much of its life. Market reports put street prices above $40,000 at the peak. Reports on competing accelerators put their prices at $10,000 to $15,000 each, which shows how much NVIDIA's software ecosystem and availability are worth to buyers.
Memory is a large part of the value. Fitting a large model on one card, or across a few cards linked by fast interconnects, saves you engineering time and extra servers. That is why buyers pay for the 80GB version instead of cheaper cards with less memory.
What moves your quote
Five factors explain most of the gap between quotes:
- Form factor: PCIe or SXM, covered in the next section.
- Sales channel: an authorized partner, an OEM system builder, a broker or a used-hardware seller.
- Volume: bulk orders of several cards or full servers usually get lower per-unit pricing.
- Supply timing: lead times and stock levels shift the price you can negotiate.
- Taxes and currency: a European listing may show a price before VAT, which can add roughly a fifth to the total.
The H100 has no single price. It has a price per configuration, per channel and per quantity.
H100 price by configuration
Expect about $25,000 to $30,000 for an H100 80GB PCIe, $35,000 to $40,000 for an SXM card, and several hundred thousand dollars for a DGX H100. Published ranges overlap, and one data center reseller quotes $27,000 to $40,000 across all single-unit configurations.
| Configuration | What you get | Typical price |
|---|---|---|
| H100 80GB PCIe | One card for a standard server slot | $25,000 to $30,000 |
| H100 80GB SXM | One module for HGX-style boards | $35,000 to $40,000 |
| DGX H100 | Complete system with 8 SXM GPUs | Roughly $300,000 to $400,000, by quote |
| Used H100 | Single card from a resale vendor | Listings start around $30,000 |
| Cloud H100 80GB | Rented by the hour | From about $2.40 per GPU hour |
PCIe versus SXM
The PCIe card fits a standard server and draws about 350W. The SXM module draws up to 700W, sits on a dedicated board and links to its neighbors over NVLink at 900GB/s. SXM also uses faster HBM3 memory at about 3.35TB/s, against roughly 2TB/s for PCIe.
If you train across many GPUs, the SXM interconnect and memory bandwidth justify the extra $10,000 or so per card. If you serve a model that fits on one or two cards, PCIe usually gives the better cost per dollar.
DGX H100 price and the cost beyond the card
The DGX H100 is a finished 8-GPU server with 640GB of total GPU memory. NVIDIA sells it through partners and quotes it individually, and reported figures cluster between $300,000 and $400,000. The eight SXM cards alone would cost about $280,000 to $320,000 at the prices above. You are paying for the host CPUs, system memory, networking and support on top.
Budget for the room too. A DGX H100 can draw around 10kW at full load, which means rack power, cooling and high-speed networking are line items of their own.
A DGX H100 is mostly the cost of eight SXM cards, plus the server and networking around them.
How to decide between buying and renting an H100
Rent if your GPU use is uneven or you need capacity within minutes. Buy only if you can keep the card busy for roughly two years or more. The break-even math below shows why.
A worked break-even example
Take a $30,000 PCIe card and add a 30% allowance for the host server, networking and installation. That is an illustrative figure, not a quote, and it brings the cost to $39,000. Compare it with renting at $2.40 per hour.
Break-even is $39,000 divided by $2.40, or 16,250 GPU hours. How long that takes depends on how busy the card is:
| Daily use | Days to break even | Years |
|---|---|---|
| 24 hours | 677 | 1.9 |
| 12 hours | 1,354 | 3.7 |
| 6 hours | 2,708 | 7.4 |
This ignores electricity, staff time and resale value. It also uses a low rental rate. At a higher hourly rate, break-even arrives sooner. Newer GPUs keep arriving, though, which can cut what your hardware is worth. For most teams under steady 24/7 load, buying only wins after about two years.
Renting an H100 from Geodd
Geodd's dedicated GPU access gives you single-tenant bare metal with SSH access and hourly billing. H100 is one of the available GPU models, and capacity is ready in 60 seconds to 5 minutes depending on the GPU type. Virtual machine options include a 1x H100 80GB type with 20 vCPUs, 128 GiB of memory and 1200 GB of storage. The 2x type has 40 vCPUs, 256 GiB and 2500 GB. Check the current hourly price before you plan, since it can change.
Skipping the GPU with per-token pricing
If you only need model output, you may not need a GPU at all. NVIDIA cites inference on an H100 at about $0.09 per million tokens for GPT-OSS-120B using vLLM, as of April 2026. That is a vendor figure for a well-tuned, busy card, but it shows the point: high utilization spreads the card's cost over a huge number of tokens. An API gives you that utilization without the hardware, and you pay only for what you use.
Compare the cost per useful hour, not the sticker price. An idle card costs the same as a busy one.
What to check before you buy an H100
Before you buy, confirm the exact model, the seller's authorization and the warranty. Most bad H100 deals come from mismatched specs or unverifiable sellers, not from the headline price.
New, used and marketplace listings
Buy new cards from an authorized NVIDIA partner or an established OEM. Used cards exist, with one vendor listing them from around $30,000, but they need a clear warranty and a way to verify the serial number. Treat very low marketplace prices, including those on large retail sites, as a warning sign. Also confirm you are getting an 80GB card in the form factor your servers support, since PCIe and SXM are not interchangeable.
Is the H100 end of life?
H200 and Blackwell-generation GPUs now exist, but the H100 is still widely sold, deployed and rented. Newer chips can put downward pressure on H100 pricing over time, which is one more reason to rent when your needs are uncertain. Ask any vendor for the lead time, the warranty length and what support covers before you commit.
Budgeting for an H100
Plan around roughly $25,000 to $40,000 per card, several hundred thousand dollars for a DGX H100, and an hourly rate if you rent. Add 30% or more for the server and networking when you buy. Then run the break-even math against your real daily usage before you commit.
For most teams, renting wins until demand is steady and high. If you want to test your workload on real hardware first, you can rent dedicated H100 GPUs by the hour and compare the result with the quotes you have collected.
