Cloud GPUs · Models, providers & rental costs

Compare GPU rental prices

One GPU for an experiment, or several for production? Compare cloud GPU prices and memory across providers, then find a configuration for inference, training, or video generation.

32 GB · VRAM per GPU

RTX 5090

$0.47/GPU-hr

Compare 3 providers ↗

24 GB · VRAM per GPU

RTX 4090

$0.14/GPU-hr

Compare 6 providers ↗

80 GB · VRAM per GPU

H100 SXM

$1.60/GPU-hr

Compare 15 providers ↗

288 GB · VRAM per GPU

NVIDIA B300

$7.40/GPU-hr

Compare 4 providers ↗

Starting rates use recent quotes from listed providers. Check the model page for configuration, minimum GPU count, and availability.

How much does a cloud GPU cost per hour?

Filter by model and memory, then compare provider quotes for the same GPU.

Sources & methodology ↗
98 of 98 GPUs

Swipe the table to see providers, quote dates, and comparison links →

GPU modelVRAM / GPUFrom / GPU-hourStarting-rate sourceQuote observed (UTC)Compare
H100 SXM5NVIDIA H100 SXM5 80GB80 GB$1.60Confirm availabilityGenesis Cloud15 quoted providers03:00 UTCCompare →
H100 PCIeNVIDIA H100 PCIe 80GB80 GB$2.49Confirm availabilityCrusoe Cloud3 quoted providers03:00 UTCCompare →
A100 80GBNVIDIA A100 SXM4 80GB80 GB$0.89Confirm availabilityJarvis Labs13 quoted providers10:26 UTCCompare →
L40SNVIDIA L40S 48GB48 GB$0.69Confirm availabilityShadeform5 quoted providers03:00 UTCCompare →
RTX 4090NVIDIA RTX 4090 24GB24 GB$0.141-GPU node · Texas, US · Host verification revoked$0.14 / minimum rental hourRentable when checkedvast.ai6 quoted providers02:59 UTCCompare →
H200 SXM5NVIDIA H200 SXM5 141GB141 GB$2.25Confirm availabilityShadeform8 quoted providers03:00 UTCCompare →
L40 48GBNVIDIA L40 48GB48 GB$0.94Rentable when checkedQuantaCloud1 quoted provider03:00 UTCCompare →
RTX 6000 AdaNVIDIA RTX 6000 Ada 48GB48 GB$0.69Confirm availabilityLambda Labs2 quoted providers02:59 UTCCompare →
A40 48GBNVIDIA A40 48GB48 GB$0.15Confirm availabilityShadeform3 quoted providers03:00 UTCCompare →
RTX 3090NVIDIA RTX 3090 24GB24 GB$0.121-GPU node · Quebec, CA · Host verification revoked$0.12 / minimum rental hourRentable when checkedvast.ai4 quoted providers22:17 UTCCompare →
V100 32GBNVIDIA V100 SXM2 32GB32 GB$0.39Confirm availabilityShadeform4 quoted providers03:00 UTCCompare →
RTX A6000NVIDIA RTX A6000 48GB48 GB$0.42Secure · per GPU/hour · min 1 GPU, $0.42/hour total · 1-GPU node · datacenter US · reported memory 48GB$0.42 / minimum rental hourRentable when checkedLium8 quoted providers20:11 UTCCompare →

USD quotes are used where available; CNY quotes are converted at CNY 7.25 per USD. Historical quotes are excluded from recent-rate comparisons. Provider checkout determines final pricing and extra charges.

What are you renting a GPU for?

Start with a shortlist, then verify it against your workload.

01

Image & video generation

Check your workflow’s image, memory, and runtime requirements. For batch image or video jobs, compare the total cost of single-GPU and multi-GPU configurations.

RTX 5090 →RTX 4090 →
02

LLM inference & fine-tuning

Model size, precision, context length, and concurrency all affect memory demand. Shortlist configurations that fit, then measure throughput and cost with your own requests.

A100 80GB →H100 SXM →
03

Large-memory & multi-GPU jobs

For B200, B300, and similar configurations, compare the required GPU count, whole-instance price, and interconnect. Multiple GPUs do not automatically provide a single shared memory pool.

B200 →B300 →

GPU rental budget

What could your next workload cost?

Choose a GPU and provider to estimate the cost of the recorded whole instance or minimum rental. A two-GPU node must be budgeted as two GPUs, even if you only need one. Quotes without a recorded rental configuration need checking first.

Estimate for the recorded minimum rental$46.89$0.4689 × 100 hours × 1 units1 GPU(s) per rental unit. Multiple units are a budget scenario, not confirmed capacity.1-GPU node · Slovakia, SK · Host verification revokedRentable when checkedQuote source: vast.ai · 2026-10-27 02:59 UTCUSD 46.89 · Native-currency estimateExcludes storage, transfer, taxes and discounts. Conversion uses a fixed display rate of USD 1 = CNY 7.25, not a live exchange rate. Quotes with unconfirmed stock are budget references only.Compare quotes for this GPU →

Three things to check before you rent

One GPU or an entire instance?

A low per-GPU rate may require a four- or eight-GPU instance. Check minimum GPU counts and whether CPU, RAM, and disk are included.

Can your workload and data run there?

Check the image, drivers, framework, region, and data transfer requirements. A ready-made template still needs the right version for your workflow.

What stops billing when work ends?

Check stop versus terminate rules, storage retention, and minimum billing periods. Compare total bills for short jobs and committed rates for sustained workloads.

From GPU selection to deployment

All GPU guides →

RTX 5090 prices: buy or rent?

Compare a dated hardware listing and recorded rental rates for occasional jobs, monthly use and purchase budgets.

Read the guide →

Choosing GPUs for enterprise LLMs

Choose memory and concurrency capacity, then compare hourly and monthly rental.

Read the guide →

Rent a GPU for MiniMax H3

Inspect the image and GPU configuration before setting up a cloud deployment.

Read the guide →

Video production: API or rented GPUs?

Compare Seedance generation costs with rented GPUs for H3 batch production.

Read the guide →

Buying vs. renting GPUs for teams

Compare purchasing and rental budgets using procurement examples and expected runtime.

Read the guide →

GPU rental questions

How is cloud GPU rental billed?

Providers may bill by runtime, offer daily or monthly plans, or use spot pricing. This directory shows rates per GPU-hour. Open a model to inspect the source, and check whether the offer requires a whole multi-GPU instance. Storage and data transfer may cost extra.

How fresh are these GPU prices?

Collectors run on a schedule; each quote shows its actual observation time. We prioritize the latest provider quotes observed within 48 hours. A model with only two-to-seven-day-old quotes is marked as historical. Models without a valid observation in seven days are temporarily omitted.

Should I rent an RTX 4090, RTX 5090, or H100?

Start with the memory required by your model and workflow, then compare runtime and rental cost for the same task. RTX 4090 and 5090 are candidates for generation and inference. For larger-memory workloads or distributed training, check data-center GPU variants, interconnects, and the software environment. Memory capacity alone does not establish compatibility.

Will I still be charged after stopping a GPU instance?

Stopping, pausing, and terminating an instance can have different billing effects. Persistent volumes may continue to incur charges. Check the provider’s minimum billing period, stop-billing rules, and storage retention policy before starting.

Is buying a GPU cheaper than renting long term?

Compare expected runtime, a matched whole-machine configuration, and operating costs. For intermittent experiments, start with pay-as-you-go costs. For sustained workloads, compare monthly and dedicated-instance quotes with an actual equipment quote. Future resale values need not be assumed.

ComputeUnion aggregates GPU rental and model API quotes for comparison. Compare model API prices →