Loading prices…
Lowest on-demand
$1.82per GPU·h · ≈ $1,325/month
at Cudo ComputeMedian on-demand
$3.35/GPU·h
across 30 providers
Lowest spot
$0.74/GPU·h
Reserved from $1.81
FP16/BF16 cost
$1.83/PFLOP·h
#9 of 69 GPUs for value
Providers
30
102 configurations
Price history
Daily H100 SXM prices per GPU-hour. The shaded period is reconstructed from public archives: the version history of SkyPilot's price catalogs, Internet Archive copies of providers' pricing pages and the open gpu-price-feed project (fewer providers than live data). After that, PetaPrice's own tracking.
Lowest on-demandMedian on-demandLowest spot
Price by provider
Cheapest On-demand H100 SXM offer at each provider, per GPU-hour. Shorter and greener is cheaper.
Cudo Compute🇬🇧$1.82best- Horizon$1.95+7%
Voltage Park🇺🇸$1.99+10%
Seeweb🇮🇹$2.12+17%
UpCloud🇫🇮$2.50+37%
Denvr🇨🇦$2.50+38%
Lium$2.50+38%
Koyeb🇫🇷$2.50+38%
Akash Network$2.55+40%
Gcore🇱🇺$2.60+43%
RunPod🇺🇸$2.69+48%
GPU.ai🇺🇸$2.79+54%
Civo🇬🇧$2.99+65%
Massed Compute🇺🇸$3.14+73%
Hyperstack🇬🇧$3.20+76%
Hyperbolic🇺🇸$3.49+92%
Sesterce🇫🇷$3.52+94%
Leafcloud🇳🇱$3.53+95%
Oblivus🇬🇧$4.312.4×
DigitalOcean🇺🇸$4.412.4×
Nebius🇳🇱$4.502.5×
Lambda🇺🇸$4.592.5×
EmpirioLabs AI🇺🇸$4.802.6×
Verda (DataCrunch)🇫🇮$5.092.8×
CoreWeave🇺🇸$6.163.4×
AWS🇺🇸$6.883.8×
STACKIT🇩🇪$9.025.0×
Oracle Cloud🇺🇸$10.05.5×
Google Cloud🇺🇸$11.06.0×
Azure🇺🇸$12.36.8×
Thin vertical line = median across providers ($3.20).
Every provider we track
Pricing
Size
Host
Provider HQ
30 providers · 102 configurations
Sort
- Cudo ComputeOn-demand🇬🇧 HQ in United Kingdomvia Cudo APIfrom · On-demandCheapest$1.82/GPU·hCheapest
- 1× GPU1 vCPU · 4 GB RAM$1.82 /GPU·h$1.82 /hOn-demandno-kristiansand-1○ sold outRent ↗
- HorizonOn-demandHQ unknownvia Shadeformfrom · On-demand+7%$1.95/GPU·h+7%
- 8× GPU · SXM160 vCPU · 1,536 GB RAM · 17,000 GB disk$1.95 /GPU·h$15.6 /hOn-demand🇮🇸 IS, Blonduos 2○ sold outRent ↗
- Voltage ParkOn-demand🇺🇸 HQ in United Statesvia Shadeformfrom · On-demand+10%$1.99/GPU·h+10%
- 8× GPU · NVLink104 vCPU · 1,024 GB RAM · 18,000 GB disk$1.99 /GPU·h$15.9 /hOn-demand🇺🇸 US, Dallas, TX○ sold outRent ↗
- 4× GPU · NVLink104 vCPU · 464 GB RAM · 4,800 GB disk$1.99 /GPU·h$7.96 /hOn-demand🇺🇸 US, Dallas, TX○ sold outRent ↗
- 2× GPU · NVLink52 vCPU · 232 GB RAM · 2,400 GB disk$1.99 /GPU·h$3.98 /hOn-demand🇺🇸 US, Dallas, TX○ sold outRent ↗
- SeewebOn-demandReserved🇮🇹 HQ in Italyvia Seeweb pricing pagefrom · On-demand+17%$2.12/GPU·h+17%
- UpCloudOn-demandSpot🇫🇮 HQ in Finlandvia UpCloud pricing pagefrom · On-demand+37%$2.50/GPU·h+37%
- LiumcommunityOn-demandSpotHQ unknownvia Liumfrom · On-demand+38%$2.50/GPU·h+38%
- DenvrOn-demand🇨🇦 HQ in Canadavia Shadeformfrom · On-demand+38%$2.50/GPU·h+38%
- 8× GPU · NVLink208 vCPU · 940 GB RAM · 20,000 GB disk$2.50 /GPU·h$20.0 /hOn-demand🇺🇸 US, Houston, TX○ sold outRent ↗
- KoyebOn-demand🇫🇷 HQ in Francevia Koyeb pricing pagefrom · On-demand+38%$2.50/GPU·h+38%
What can the H100 SXM run?
Memory for the weights plus one 32K-token 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 | $/mo |
|---|---|---|---|---|---|
gpt-oss-20b OpenAI · 20.9B | 20.9B · 3.6B active | 20 GiB | 1 | $2.12 at Seeweb | $1,547 |
Gemma 3 27B Google · 27.4B | 27.4B | 28 GiB | 1 | $2.12 at Seeweb | $1,547 |
Qwen3-32B Alibaba · 32.8B | 32.8B | 39 GiB | 1 | $2.12 at Seeweb | $1,547 |
Llama 3.3 70B Instruct Meta · 70.6B | 70.6B | 76 GiB | 2 | $4.24 at Seeweb | $3,094 |
gpt-oss-120b OpenAI · 117B | 117B · 5.1B active | 110 GiB | 2 | $4.24 at Seeweb | $3,094 |
Qwen3-235B-A22B Alibaba · 235B | 235B · 22B active | 225 GiB | 4 | $8.48 at Seeweb | $6,189 |
DeepSeek-R1 DeepSeek · 684B | 684B · 37B active | 640 GiB | 16 | $33.91 at Seeweb | $24,755 |
Kimi K3 Moonshot AI · 2.78T | 2.78T · 104B active | 2.53 TiB | 64 | $135.64 at Seeweb | $99,018 |
Assumes 90% of each GPU's memory is usable and rounds up to 1, 2, 4, 8… GPUs for tensor parallelism (up to 64). $/hr is the GPU count times the cheapest in-stock on-demand rate per GPU; check that the provider rents instances that large. Serving many requests at once needs more memory for context.
Alternatives to the H100 SXM
GPUs with similar compute and memory, and their lowest on-demand price right now.
- H100 NVL94 GB · 0.835 PF FP16$2.59/h$3.10/PFLOP·h
- H100 PCIe80 GB · 0.756 PF FP16$1.66/h$2.20/PFLOP·h
- GH20096 GB · 0.989 PF FP16$2.59/h$2.62/PFLOP·h
- H200 SXM141 GB · 0.989 PF FP16$2.92/h$2.95/PFLOP·h
- H200 NVL141 GB · 0.835 PF FP16$3.36/h$4.03/PFLOP·h
TPU v6e (Trillium)32 GB · 0.918 PF FP16$2.70/h$2.94/PFLOP·h
H100 SXM pricing questions
How much does it cost to rent an H100 SXM?
On-demand H100 SXM prices start at $1.82 per GPU-hour (Cudo Compute). The median across 30 providers is $3.35, and the most expensive is $12.3. Running one GPU around the clock costs about $1,325 per month at the lowest price.
Is there cheaper spot or interruptible H100 SXM capacity?
Yes. Spot (interruptible) H100 SXM capacity starts at $0.74 per GPU-hour, 59% below the cheapest on-demand price. Spot instances can be preempted at any time, so use them for checkpointed training or batch jobs.
How much compute do you get per dollar with the H100 SXM?
At $1.82 per hour and 0.989 PF of dense FP16/BF16, the H100 SXM costs $1.83 per PFLOP-hour. That ranks 9 of 69 GPUs with on-demand prices on PetaPrice. Figures are datasheet peaks; real utilization is typically 30–60%.
What are the H100 SXM's key specs?
80 GB HBM3 with 3,350 GB/s of memory bandwidth, Hopper architecture (2022), SXM form factor, 700 W. Dense peak: 0.989 PF FP16 / BF16, 1.98 PF FP8, 1.98 PF INT8, 0.495 PF TF32, 0.067 PF FP32, 0.067 PF FP64.
How current are these prices?
Prices were collected on Nov 11, 2026, 09:01 UTC from provider APIs and public catalogs and refresh about every 2 hours. They are indicative list prices; always confirm on the provider's website before booking.