H100 SXM Rental Price: Cheapest Cloud per Hour (2026)
Choose an H100 SXM rental by memory fit, interruption tolerance and the price of an actual launchable configuration. Current rates and reported availability appear below.
For H100 SXM, choose Hyperstack at $3.20 (Not reported); if the workload can restart, use the spot alternative from DataCrunch at $1.93 (In stock), but keep the on-demand choice when interruptions would prevent recovery.
Recommendation source: This on-demand recommendation uses a provider-published list rate; availability can still change.
Spot recommendation source: This spot recommendation uses a provider-published list rate; availability can still change.
Spot evidence: 2 observed spot configurations from 1 provider; this is a configuration count, not a count of independent provider quotes.
H100 SXM is the exact GPU variant compared here. A per-GPU comparison does not establish that a larger node can be rented one GPU at a time.
Before committing, run a bounded test at the confirmed checkout rate. Record peak memory, billed runtime and completed work. Include checkpoint recovery in the test if you intend to use spot.
Current recommendation
For runs that cannot be interrupted, use Hyperstack at $3.20/hr (Not reported).
2 spot offers are observed; the lowest is DataCrunch at $1.93/hr (In stock); use it only for checkpointable work.
The closest on-demand comparison is Hyperstack at $3.20/hr (Not reported) versus RunPod at $3.49/hr (In stock): $0.29/hr, or $2540.40/year.
Today's prices
USD/hr · one observed offer per provider and pricing model
1 GPU
| Provider | Configuration $/hr | Pricing | Price basis | Availability |
|---|---|---|---|---|
| DataCrunch | $1.93 | spot | Published list rate | In stock |
| Hyperstack | $3.20 | on-demand | Published list rate | Not reported |
| RunPod | $3.49 | on-demand | Published list rate | In stock |
| DataCrunch | $3.85 | on-demand | Published list rate | In stock |
| Vast.ai | $3.87 | on-demand | Marketplace quote | In stock |
| Crusoe | $3.90 | on-demand | Published list rate | Not reported |
| Together AI | $3.99 | on-demand | Published list rate | Not reported |
| Lambda | $4.29 | on-demand | Published list rate | Unavailable as of 2026-10-07 |
| Paperspace | $5.95 | on-demand | Published list rate | Not reported |
2 GPUs
| Provider | Configuration $/hr | Per GPU $/hr | Pricing | Price basis | Region | Availability |
|---|---|---|---|---|---|---|
| Lambda | $8.38 | $4.19 | on-demand | Published list rate | us-southeast-1 | In stock |
4 GPUs
| Provider | Configuration $/hr | Per GPU $/hr | Pricing | Price basis | Availability |
|---|---|---|---|---|---|
| Lambda | $16.36 | $4.09 | on-demand | Published list rate | Unavailable as of 2026-10-07 |
8 GPUs
Every row below: Pricing: on-demand · Price basis: Published list rate
What an hour buys
Computed from the recommended on-demand offer: Hyperstack at $3.20/hr for 1 GPU.
Hardware specifications
| Specification | H100 SXM |
|---|---|
| Memory | 80 GB HBM3 |
| Memory bandwidth | 3350 GB/s |
| Dense BF16 | 989 TFLOPS |
| Dense FP8 | 1979 TFLOPS |
| NVLink | 900 GB/s |
| TDP | 700 W |
What your workload needs
VRAM needed is calculated from the formula shown in each row. A listed hourly cost appears only when a currently eligible on-demand offer exists for that exact GPU count.
| Workload | VRAM needed | H100 SXM | Listed hourly cost |
|---|---|---|---|
| 7B Q4 inference | 4 GB 7B × 0.5 B (4-bit) × 1.2 (KV+activations) | 1 GPU | $3.20/hr (Not reported; observed 2026-10-07; price basis: Published list rate) |
| 13B Q4 inference | 8 GB 13B × 0.5 B (4-bit) × 1.2 (KV+activations) | 1 GPU | $3.20/hr (Not reported; observed 2026-10-07; price basis: Published list rate) |
| 70B QLoRA | 46 GB 70B × 0.5 B (4-bit base) × 1.3 (adapters+optimizer) | 1 GPU | $3.20/hr (Not reported; observed 2026-10-07; price basis: Published list rate) |
| 70B FP8 training | 154 GB 70B × 2 B (FP8 mix + master/optimizer) × 1.1 (requires FP8 hardware) | 2 GPUs | $8.38/hr (In stock; observed 2026-10-07; price basis: Published list rate) |
| 70B FP16 inference | 168 GB 70B × 2 B (FP16) × 1.2 (KV+activations) | 8 GPUs | $49.24/hr (Not reported; observed 2026-10-07; price basis: Published list rate) |
What the numbers say
H100 SXM should be selected only after the applicable workload row’s displayed formula and GPU count fit your model and serving plan. Treat each row as a workload-specific estimate, then validate the stated configuration against your actual weights, KV cache, activations, sequence length, batch size, and runtime overhead before committing.
H100 SXM has published specifications of 3350GB/s memory bandwidth, 989 TFLOPS at BF16, and 1979 TFLOPS at FP8, while its 900GB/s NVLink figure describes interconnect rather than single-GPU speed. FP8’s 2x specified throughput relationship and lower bytes per parameter can change both fit and compute-bound speed, but these figures are specifications or modeled bounds, not measured application throughput; HBM generation changes capacity and bandwidth, not compute.
H100 SXM is not the right choice when the workload table requires a GPU count for which no eligible offer is shown or when a representative run fails the stated fit assumptions. In those cases, change the configuration or evaluate another GPU with verified capacity before comparing rental prices.
Turn the workload estimate into a rental decision
- Measure the memory peak
Use the intended precision, batch size and sequence length. KV cache stores attention state during serving; activations and optimizer state depend on the training setup. A weight-only estimate leaves these out.
- Match the sold configuration
Read the GPU count next to the hourly cost. If no eligible offer exists for that count, the memory estimate is not a launchable rental. Check topology before splitting a job across GPUs.
- Bound the experiment
Set a test budget and stop condition. Record billed runtime and completed work at the confirmed checkout rate, including recovery time for a spot test.
How Marlin helps
Marlin matches your workload requirements to the lowest-priced suitable GPU option across supported providers.
- Broader coverage: Marlin matches across every supported provider, not only the providers priced on this page.
- Unified comparison: Marlin compares supported providers in one matching process, reducing the need to check prices manually across separate sites.
- Fit-aware choice: Marlin returns the cheapest option that satisfies your workload requirements, not simply the cheapest row regardless of fit.
Manually revisiting provider pages to compare rates and verify the $3.20 figure takes repeated effort; Marlin handles that comparison when matching your workload to the lowest-priced suitable option.
Before you rent
Workload memory and cost projections use the stated assumptions; validate them with measured memory and billed runtime. Prices were collected from provider pricing sources, while fixed hardware specifications came from the published GPU datasheet. Re-check current rates and availability on the linked provider pages and hardware figures in the manufacturer sources, then confirm the checkout terms. https://vast.ai/pricing · https://www.hyperstack.cloud/gpu-pricing · https://datacrunch.io/pricing · https://www.runpod.io/pricing · https://crusoe.ai/cloud/pricing · https://www.together.ai/pricing · https://www.paperspace.com/pricing · https://resources.nvidia.com/en-us-tensor-core/nvidia-tensor-core-gpu-datasheet · https://lambda.ai/service/gpu-cloud · https://www.coreweave.com/pricing
Marlin beta
Stop comparing. Start running.
Stop tracking prices manually and use Marlin to match your requirements to the lowest-priced suitable option across supported providers.
Get started with Marlin