GPUBeacon
Language

Rent a GPU

Rent a cloud GPU by the hour

Pick the GPU your model actually needs, then rent the cheapest in-stock on-demand hour. No contract, no quote theatre. Every price below is a published rate from the GPUBeacon catalog.

Verified 2026-09-05

GPUs for rent, cheapest hour first

The cheapest in-stock on-demand row for each model. Open a GPU to see every provider and billing type.

GPUVRAMFrom / hr720 h monthBest for
Rent RTX 4090Cheapest on Vast.ai24 GB$0.29$209Image generation, 7B-13B inference, LoRA experiments
Rent RTX 5090Cheapest on Vast.ai32 GB$0.34$24532 GB inference, diffusion and video, LoRA fine-tunes
Rent L4Cheapest on Google Cloud24 GB$0.39$281Low-power inference and video workloads
Rent A10Cheapest on RunPod24 GB$0.45$324Small-model inference and light fine-tuning
Rent RTX PRO 6000Cheapest on Packet.ai96 GB$0.66$47596 GB single-GPU inference and fat adapters
Rent L40SCheapest on RunPod48 GB$0.79$56948 GB inference, FP8 serving, rendering
Rent A100Cheapest on Vast.ai80 GB$1.19$85780 GB training and fine-tuning on Ampere
Rent H100Cheapest on Vast.ai80 GB$1.73$1,246The default datacenter hour for training and 70B serving
Rent MI300XCheapest on Nebius192 GB$1.99$1,433192 GB AMD memory for large models
Rent H200Cheapest on Koyeb141 GB$3.00$2,160141 GB Hopper for long context and big KV cache
Rent B200Cheapest on Packet.ai180 GB$3.75$2,700Blackwell training and FP4 serving
Rent B300Cheapest on Verda288 GB$7.50$5,400288 GB Blackwell Ultra for the largest memory-bound jobs

Full GPU pricing chartBrowse every GPUAll GPU cloud providers

Most requested

Where to rent an H100 right now

In-stock on-demand rows, cheapest first. Confirm SXM vs PCIe and the live rate before you provision.

ProvidersRegionBilling typeInterconnect/GPU/hrOpen provider
Vast.aiIn stockUS-EastOn-demandPCIe$1.73Open Vast.ai
RunPodIn stockUS-EastOn-demandNVLink$2.39Open RunPod
LambdaIn stockUS-WestOn-demandNVLink$2.49Open provider
FluidstackIn stockEU-WestOn-demandPCIe$2.69Open provider
KoyebIn stockEU-WestOn-demandNVLink$2.85Open provider
HyperstackIn stockEU-WestOn-demandNVLink$2.99Open provider

Which GPU should you rent?

Start from the bytes, not the badge. Weights, optimizer states and KV cache decide the GPU. Price comes second.

Not sure the model fits? Use the VRAM fit tool, then price the run with the hour cost calculator. If two cards look close, open a duel: RTX PRO 6000 vs RTX 5090, H100 vs H200 or L40S vs A100.

Where to rent: marketplace, managed cloud or hyperscaler

Marketplaces such as Vast.ai have the cheapest hour and the most variance. Managed clouds such as RunPod trade a little price for a product. Specialist clouds like Lambda and CoreWeave sell clusters and support. Hyperscalers sell the account you already have, at a higher hour. Read on-demand vs spot before you chase the cheapest row.

FAQ

Renting a GPU: questions

How much does it cost to rent a GPU?+

From $0.29 per hour for RTX 4090 to $7.50 per hour for B300, at the cheapest on-demand row in the GPUBeacon sample. A full month (720 hours) of the cheapest card is about $209. Most teams pay for hours, not months.

What is the cheapest GPU to rent for AI?+

RTX 4090 has the lowest hourly rate in this sample at $0.29 on Vast.ai. Cheap is only useful if the model fits: 24 GB of VRAM is the limit. Check the bytes before the price.

Which GPU should I rent for LLM fine-tuning?+

LoRA and QLoRA on 7B-13B models fit 24 GB cards. 30B-class and 70B QLoRA work wants 48-96 GB. Full fine-tunes and multi-GPU jobs belong on H100, H200 or B200 with NVLink. Use the VRAM fit tool to estimate the memory before you rent.

Is it cheaper to rent a GPU or buy one?+

Divide the quote for the card and its server by the hourly rate. That is your break-even in rental hours. Add power, cooling and idle time. If you will not keep it busy past that point, renting is cheaper and removes the capex.

Can I rent a GPU by the hour without a contract?+

Yes. On-demand rows are pay-as-you-go: you pay the published hourly rate and stop whenever you want. Reserved capacity and quotes are different products and are labeled separately in the catalog.