H100
A widely deployed accelerator for large-model training and high-throughput inference.
- Memory
- 80 GB HBM3
- Architecture
- Hopper
- Sample listings
- 5 across 5 providers
From $1.56/hrCompare H100 prices →
GPU rental price index
Compare GPU prices by model.
Shortlist cloud accelerators by VRAM, architecture, published hourly rate, and available provider configurations. Every price below is a labeled sample, normalized to USD per physical GPU-hour.
Models tracked6
Provider profiles6
Sample floor$0.32per GPU / hr
Hourly accelerator index
From RTX 4090 experiments to H200 inference.
Sticker price is only the first filter. Open a model page to compare node size, CPU and RAM allocation, location, rental terms, availability, and the total hourly commitment.
H200
A high-memory Hopper accelerator suited to larger contexts and memory-bound inference.
- Memory
- 141 GB HBM3e
- Architecture
- Hopper
- Sample listings
- 3 across 3 providers
From $3.42/hrCompare H200 prices →
B200
A Blackwell datacenter GPU for frontier training and dense inference clusters.
- Memory
- 180 GB HBM3e
- Architecture
- Blackwell
- Sample listings
- 2 across 2 providers
From $5.50/hrCompare B200 prices →
A100 80GB
A mature training and inference option with broad cloud availability.
- Memory
- 80 GB HBM2e
- Architecture
- Ampere
- Sample listings
- 3 across 3 providers
From $0.88/hrCompare A100 80GB prices →
L40S
A flexible accelerator for inference, rendering, and multimodal workloads.
- Memory
- 48 GB GDDR6
- Architecture
- Ada Lovelace
- Sample listings
- 2 across 2 providers
From $0.69/hrCompare L40S prices →
RTX 4090
A cost-focused option for smaller inference jobs, image generation, and experimentation.
- Memory
- 24 GB GDDR6X
- Architecture
- Ada Lovelace
- Sample listings
- 2 across 2 providers
From $0.32/hrCompare RTX 4090 prices →
Selection framework
Match memory first. Optimize price second.
Model weights, context length, batch size, optimizer state, and parallelism determine memory demand. Once the workload fits, compare the delivered node rather than the GPU label alone.
Prototyping and inference
RTX 4090 and L40S listings can suit smaller models, image generation, development, and throughput-focused inference.
Mature datacenter capacity
A100 80GB and H100 are common comparison points for training, fine-tuning, and larger production inference.
High-memory workloads
H200 and B200 reduce memory pressure for large weights, long contexts, and memory-bound inference pipelines.
Buyer questions
GPU rental price FAQ
How much does it cost to rent a GPU per hour?
The hourly cost depends on the accelerator, node configuration, region, and purchase model. The sample offers in this preview range from budget single-GPU listings to multi-GPU datacenter nodes; use each model page for a normalized per-GPU comparison.
Should I compare per-GPU price or total node price?
Use the per-GPU rate to compare accelerator cost, then check total node price before budgeting. A low normalized rate on an eight-GPU node can still require a much larger hourly commitment.
Do you list a B300 GPU price?
Not yet. This preview has no verifiable B300 rental offer, so RackRate does not publish an estimated price. The model will be added when a source-backed listing is available.
Need a stronger shortlist?
Use the buying guide →Compare the complete GPU offer.
Carry the node, region, purchase model, and hidden costs into the decision—not only the accelerator name.