NVIDIA H100 PCIe rental pricing
Hopper architecture · 80 GB memory · 2.0 TB/s bandwidth · 350 W board power · Released 2022
Price history by provider
Every provider's cheapest daily on-demand rate for the NVIDIA H100 PCIe, last 30 days
One line per provider, reconstructed from recorded price changes for listings currently live. Click a name in the key to hide or show it.
At a glance
prices checked 5h ago- Cheapest in stock
- $1.83/hr
- On-demand · 1× H100 PCIe · Beam
- Median price
- $2.70/hr
- On-demand · 6 providers
- 90-day trend
- Flat
- Weekly median on-demand, 4 weeks
- Coverage
- 10 providers
- 19 price points
- $ per FP8 PFLOP-hour
- $1.21
- Rank 13 of 31 priced GPUs · lower is better
Why the PCIe costs what it does
Against the other priced NVIDIA H100 variants: the price floor, what a GB-hour and a PFLOP-hour cost, and the specs that explain the gap.
The PCIe rents from $1.83 per GPU-hour, 132% more than the SXM5 at $0.79.
The SXM5 is an SXM module, rated at 700 W against 350 W for the PCIe card. Hosting a module with that power and interconnect budget costs more than hosting a card, and the hourly rate carries it.
Memory bandwidth is 3.4 TB/s on the SXM5 against 2.0 TB/s on the PCIe, 68% more, which is what sets decode throughput for inference.
On dense FP8 the SXM5 rates 1,979 TFLOPS against 1,513 TFLOPS, 31% more.
32 providers list the SXM5 against 10 for the PCIe; a deeper market tends to price lower.
Per FP8 PFLOP-hour and per GB-hour the SXM5 is the better buy: $0.40 against $1.21 per PFLOP-hour, and $0.010 against $0.023 per GB-hour.
Floors are the cheapest live listing of any pricing type; medians are of each provider's cheapest on-demand rate. Per GB-hour divides the floor by memory, per PFLOP-hour by dense throughput on the family's basis.
Every provider we track
Cheapest listing per provider and pricing type. Sold-out rows stay so you can see who normally carries it.
| Provider | Configuration | Price/GPU-hr | Checked | |
|---|---|---|---|---|
 Beam | 1× H100 PCIe | $1.83 | 9d ago | View → |
| 1× H100 PCIe Community | $1.90 | 5h ago | View → | |
| 1× H100 PCIe Community | $1.99 | 5h ago | View → | |
 Civo | 1× H100 PCIe Reserved | $1.99 | 5h ago | View → |
| 1× H100 PCIe Spot | $2.00 | 5h ago | View → | |
| 1× H100 PCIe Community | $2.00 | 5h ago | View → | |
| Civo | 1× H100 PCIe | $2.49 | 5h ago | View → |
| 1× H100 PCIe | $2.50 | 5h ago | View → | |
 Oblivus | 1× H100 PCIe Reserved | $2.56 | 5h ago | View → |
| 1× H100 PCIe | $2.89 | 5h ago | View → | |
 Thunder Compute | 1× H100 PCIe | $3.20 | Stale · 42d ago | View → |
| Oblivus | 1× H100 PCIe | $3.50 | 5h ago | View → |
 Shadeform Latitude.sh capacity | 1× H100 PCIe | $1.66 | 5h ago | View → |
| 1× H100 PCIe | $3.29 | 5h ago | View → |
Per GPU-hour. Spot, reserved and community pricing differ from on-demand; click through for the provider's current figure.
What can the NVIDIA H100 PCIe run?
Memory for the weights plus one sequence of context, how many GPUs hold it, and what that replica costs at the cheapest in-stock on-demand rate.
| Model | Params | Memory | GPUs | $/hr |
|---|---|---|---|---|
| gpt-oss-20b OpenAI | 20.9B· 3.6B active | 20 GiB | 1 | $1.83 |
| Gemma 3 27B Google | 27.4B | 26 GiB | 1 | $1.83 |
| Qwen3-32B Alibaba | 32.8B | 32 GiB | 1 | $1.83 |
| Llama 3.3 70B Instruct Meta | 70.6B | 67 GiB | 1 | $1.83 |
| gpt-oss-120b OpenAI | 117B· 5.1B active | 109 GiB | 2 | $3.66 |
| Qwen3-235B-A22B Alibaba | 235B· 22.0B active | 220 GiB | 4 | $7.32 |
| DeepSeek-R1 DeepSeek | 684B· 37.0B active | 638 GiB | 16 | $29.28 |
| Kimi K3 Moonshot AI | 2.78T· 104B active | 2.53 TiB | Over 16 | — |
Sized for one sequence at 4K context with 16-bit KV cache and 10% of memory held back. Serving many users at once needs more; the LLM GPU calculator sizes for concurrency and throughput.
NVIDIA H100 PCIe pricing questions
How much does it cost to rent a NVIDIA H100 PCIe per hour?
The cheapest NVIDIA H100 PCIe in stock right now is $1.83 per GPU-hour on-demand from Beam for the 1× H100 PCIe configuration. The median on-demand rate across 6 providers is $2.70 per GPU-hour.
How much does a NVIDIA H100 PCIe cost per month?
Running one NVIDIA H100 PCIe around the clock for a month (730 hours) costs about $1,336 at the cheapest in-stock rate of $1.83 per hour. At the median on-demand rate it is about $1,967 a month. Reserved terms usually come in below that; ask the provider for a quote.
Which provider has the cheapest NVIDIA H100 PCIe?
Beam has the cheapest NVIDIA H100 PCIe in stock right now at $1.83 per GPU-hour on-demand for the 1× H100 PCIe configuration. Rates change often, so check the provider table for today's figure before you commit.
Are NVIDIA H100 PCIe rental prices going up or down?
The weekly median on-demand rate for the NVIDIA H100 PCIe has been flat, moving 0.0% over the last 4 weeks, from $2.70 to $2.70 per GPU-hour.
What is the difference between on-demand, spot and reserved NVIDIA H100 PCIe pricing?
On-demand is pay-as-you-go with no commitment. Spot is spare capacity sold at a discount that the provider can reclaim at short notice. Reserved is a committed term, usually months, in exchange for a lower rate. Community is capacity from individual hosts on a marketplace, cheaper and less uniform. This week's median rates for the NVIDIA H100 PCIe are on-demand $2.70 (8 providers), spot $2.00 (1 provider), reserved $2.28 (2 providers), community $1.99 (3 providers), per GPU-hour.
How many cloud providers rent the NVIDIA H100 PCIe?
We track 10 providers with 19 live prices for the NVIDIA H100 PCIe. Sold-out listings stay in the table so you can see who normally carries it, but never set the cheapest figure.
What LLMs can you run on one NVIDIA H100 PCIe?
With 8-bit weights and a single sequence of context, one NVIDIA H100 PCIe holds gpt-oss-20b (20.9B), Gemma 3 27B (27.4B), Qwen3-32B (32.8B), Llama 3.3 70B Instruct (70.6B). See the model fit table for the full ladder.
How does the NVIDIA H100 PCIe compare on price per PFLOP-hour?
At its cheapest in-stock on-demand rate the NVIDIA H100 PCIe delivers dense FP8 compute at $1.21 per PFLOP-hour, which ranks 13 of 31 priced GPUs we cover on that basis. Lower is better: it is the rental price divided by peak FP8 throughput.
Where does this NVIDIA H100 PCIe pricing data come from?
Prices are read daily from each provider's public price list or API and recorded whenever they change. Listings are shown per GPU-hour on-demand unless marked otherwise, and the chart covers the last 30 days. The newest listing was checked 5h ago.
Why does the PCIe cost more than the SXM5?
The PCIe rents from $1.83 per GPU-hour, 132% more than the SXM5 at $0.79. The SXM5 is an SXM module, rated at 700 W against 350 W for the PCIe card. Hosting a module with that power and interconnect budget costs more than hosting a card, and the hourly rate carries it.