Vast.aiUS-EastOn-demand
$1.73/ GPU / hr
Open provider
Cloud GPU · verified 2026-09-05
NVIDIA H100
Hopper80 GB
The default datacenter hour. 80 GB of HBM3 is enough for most 70B work if you manage the KV cache. It is not a bargain because the FLOPs slide is pretty. It is a bargain when the job already fits.
VRAM80 GB
From /GPU/hr$1.73
ArchitectureHopper
VendorNVIDIA
Sponsored
Featured H100 offers
The four cheapest in-stock on-demand H100 hours in this sample. Each card is a sponsored placement.
RunPodUS-EastOn-demand
$2.39/ GPU / hr
Open provider
LambdaUS-WestOn-demand
$2.49/ GPU / hr
Open provider
FluidstackEU-WestOn-demand
$2.69/ GPU / hr
Open providerSponsored does not change ranking. Commissions never move a row, a GPU, or a displayed price.
In brief
What to know before you rent
- Rent SXM when NVLink or full 3.35 TB/s matters. PCIe and NVL are different SKUs with the same name.
- NVIDIA’s Tensor numbers are usually with sparsity. Dense FP8 on SXM is 1,979 TFLOPS, not 3,958.
- If the model and states fit in 80 GB, H200 is a more expensive H100. If they do not, stop pretending.
Catalog listings
H100 Pricing and Availability
Published on-demand rates first. Spot and waitlists are labeled. Confirm the live rate before you provision.
ProvidersRegionBilling typeInterconnect/GPU/hr
Our reading
Who this GPU is for
I rent H100 when I can fill the SMs and the 80 GB envelope. I do not rent it as a status GPU. Marketplace H100s at meme prices are a different product from a Lambda or CoreWeave SXM node. Compare the SKU, not the three characters on the badge.
Rent it if Fine-tunes and serving that fit 80 GB, single-node training that actually uses NVLink, teams who can checkpoint and who will read SXM vs PCIe before they paste a card.
Skip it if Context windows that blow the KV cache, multi-node jobs that need a reserved fabric, or anyone shopping a 4090 workload with Hopper money.
How we compare hoursDatasheet
Technical specifications
NVIDIA H100 · Per GPU · SXM
Compute · dense
- FP8
- 1,979 TFLOPS3,958 with sparsity
- FP16 / BF16
- 989.5 TFLOPS1,979 with sparsity
- TF32
- 494.5 TFLOPS989 with sparsity
- INT8
- 1,979 TOPS3,958 with sparsity
- FP32
- 67 TFLOPS
- FP64
- 34 TFLOPS67 TFLOPS Tensor Core
- Precision support
- FP8, FP16, BF16, TF32, FP32, FP64, INT8
Memory
- Capacity
- 80 GB HBM3
- Bandwidth
- 3,350 GB/s
- Bus width
- 5,120-bit
- ECC
- Yes
Silicon
- Architecture
- Hopper
- Process
- TSMC 4N
- Transistors
- 80 billion
- Shader cores
- 16,896CUDA cores
- Matrix cores
- 5284th-gen Tensor cores
- Compute units
- 132SMs
Fabric and host
- GPU interconnect
- NVLink 900 GB/s
- Host interface
- PCIe 5.0 x16128 GB/s
Power
- Board power
- 700 WSXM, configurable
Platform
- Partitioning
- MIG, up to 7 instances10 GB each on SXM 80 GB
Table is H100 SXM, the SKU catalogs mean when they say “an H100.” H100 PCIe is ~350 W with slower HBM. H100 NVL is 94 GB and a 600 GB/s NVLink bridge, useful for 70B inference, not a silent upgrade. Confirm the form factor on the listing.
Trade-offs
Pros and cons of H100
What this GPU does well, and when the hour is the wrong buy.
Strengths
- The SKU every catalog actually stocks. You can compare an hour, not a waitlist
- FP8 Transformer Engine is the reason this generation still trains at a sane $/step
- MIG exists when you want to split an 80 GB card instead of overpaying for a second one
- NVLink 900 GB/s on SXM is the difference between 8 GPUs and 8 expensive heaters
Limits
- 80 GB is the wall. Long context and fat Adam states do not care about your TFLOPS screenshot
- PCIe H100 is slower memory and less interconnect, same marketing name
- 700 W SXM needs a real node. A cheap host with a sad PSU is not a discount, it is a crash
- Sparse peak numbers do not show up in your trainer unless the kernel path is sparse
Form factor
What's the difference between the H100 SXM, PCIe and NVL?
Same badge, different memory, bandwidth or interconnect. Confirm the SKU on the listing.
| Variant | Memory | Memory bandwidth | Interconnect |
|---|---|---|---|
| H100 SXM | 80 GB | 3,350 GB/s | NVLink 900 GB/s |
| H100 PCIe | 80 GB | 2,000 GB/s | NVLink 600 GB/s |
| H100 NVL | 94 GB | 3,938 GB/s | NVLink 600 GB/s |
H100 SXM, PCIe and NVL are different products
H100 is three products wearing one badge. SXM is the HGX node: 80 GB HBM3, 3.35 TB/s, NVLink 900 GB/s, up to 700 W. PCIe is the card you can bolt into a random chassis: less bandwidth, less fabric, often a different memory generation. NVL is the 94 GB twin for 70B-class inference in power-capped halls.
If a marketplace row says H100 and the price looks like a consumer card, you are probably not buying the SXM row Lambda would sell you. Compare hours inside a SKU. Mixing them is how people save twenty cents and lose a training run.
80 GB is the H100 decision, not 1,979 TFLOPS
Most people landing here are fitting a 70B, serving a LoRA, or trying to keep a KV cache in HBM. Hopper’s SM count is the same family as H200. The upgrade is memory. If your batch already fits, you are paying H200 rent for headroom you will not touch.
Adam states, gradients, and long context are what blow 80 GB, not the marketing FLOPs. If the optimizer does not fit, extra TFLOPS are a very expensive fan. Buy memory first. Then fill the SMs.
H100 on-demand prices vs waitlisted Blackwell
On-demand H100 is the workhorse of this index: enough providers that you can walk away. That is the product. A waitlisted Blackwell at a fantasy FLOPs per watt is not a cheaper H100.
I still open Vast.ai for unverified floor prices and RunPod Secure when I want the job to finish. Lambda when the interconnect is the job. Hyperscalers when the data cannot leave. The GPU name stays H100. The product does not.
How I actually rent it
- 01
Name the SKU
SXM, PCIe, or NVL. GPU count. NVLink or not. If the listing will not say, leave.
- 02
Fit the 80 GB
Model + states + KV cache. If it spills, you want H200 or more GPUs, not a denser kernel fantasy.
- 03
Compare on-demand only
Spot and waitlists are labeled on this site. Do not mix them into the hour you budget.
FAQ
H100 questions
Who has the cheapest H100?+
As of September 5, 2026, the cheapest verified in-stock on-demand H100 tracked by GPUBeacon is $1.73 per GPU per hour from Vast.ai.
How much does the H100 cost per month?+
As of September 5, 2026, at 720 hours per month, one H100 costs an estimated $2,153 at the median on-demand price in this sample. Cheapest verified in stock: $1,246 per month on-demand.
Where can I rent an H100?+
GPUBeacon currently tracks H100 configs from 15 providers. The cheapest verified in-stock on-demand configs come from Vast.ai, RunPod, Lambda and Fluidstack. See the full price comparison above for every provider and config.
Are H100 prices going up or down?+
GPUBeacon publishes a catalog sample, not a 90-day price index. As of September 5, 2026, the median on-demand H100 in this sample is $2.99 per GPU per hour. The cheapest verified in-stock hour is $1.73.
How many H100 GPUs do I need?+
One H100 runs models up to roughly 120B parameters at 4-bit quantization or 31B at 16-bit, assuming a 32K context. Larger models run across multiple GPUs.
Why choose the H100?+
80 GB HBM3 with 900 GB/s NVLink for multi-node training. FP8 Transformer Engine for improved training throughput over A100. The standard choice for large-scale AI training and high-throughput inference.
When is the H100 not a good fit?+
May be more than needed for single-GPU inference on models under 30B parameters. Consider the L40S or A10 for cost-effective inference workloads.
What's the difference between the H100 SXM, PCIe and NVL?+
The H100 ships in 3 versions with different memory, bandwidth or interconnect. The figures at the top of this page are for the H100 SXM.
Is H100 still worth renting in 2026?+
Yes, when the job fits 80 GB and you need a SKU that is actually in stock. Blackwell is faster. Blackwell is also a waitlist with a power bill. H100 is the hour you can buy this week.
H100 or H200?+
H200 if the extra 61 GB of HBM3e shortens the job enough to beat the H100 hour, usually KV cache and fat states. If you already fill 80 GB with compute-bound work, stay on H100.
Do NVIDIA’s TFLOPS numbers include sparsity?+
The datasheet Tensor rows are with sparsity. Dense FP8 on H100 SXM is 1,979 TFLOPS. If your stack is not sparse, believe the dense column.
Can I MIG an H100 on a cloud?+
The silicon supports up to 7 instances. The provider may not expose it. If you need a slice, confirm the product page, do not assume a marketplace pod is MIG.
Keep comparing
Alternatives to NVIDIA H100
Same job, different envelope. Compare VRAM, bandwidth and the hour before you lock a badge.
NVIDIA B300
288 GB HBM3e · 2.4x the bandwidth of the H100
From$7.50/ GPU / hr
CompareNVIDIA H200
141 GB HBM3e · 43% more bandwidth than the H100
From$3.00/ GPU / hr
Read the verdictNVIDIA RTX 5090
32 GB GDDR7 · 47% less bandwidth than the H100
From$0.34/ GPU / hr
CompareNVIDIA A100
80 GB HBM2e · 39% less bandwidth than the H100
From$1.19/ GPU / hr
Read the verdict
AMD MI300X
192 GB HBM3 · 58% more bandwidth than the H100
From$1.99/ GPU / hr
CompareSources
Where these numbers come from
TFLOPS and TDP are from the vendor datasheet. Hourly prices are from the GPUBeacon catalog sample. Open the originals, then confirm the live SKU with the provider before you rent.
- DatasheetNVIDIA H100 Tensor Core GPU datasheet
Official Hopper SXM figures on this page: 80 GB HBM3, 3,350 GB/s, NVLink 900 GB/s, 700 W.
resources.nvidia.comOpen - MethodHow GPUBeacon ranks an hour
On-demand first, waitlists labeled, commissions never reorder a row. The rules behind the table above.
GPUBeaconRead