GPU Rental PricesGPU Rental Prices
Download data

Sources and update dates

Rental prices come from providers, hardware specifications from manufacturers, model parameters from publisher records, and purchase prices from named retailers. Competitor price tables are not sources.

GPU rental prices

Updated dates belong to each provider. A recent page build does not renew an older observation. Rates are USD per GPU-hour, before taxes; a quoted node total is divided by its published GPU count.

ProviderOfficial pricing sourceUpdatedCollection status
Daytonabilling.app.daytona.io2026-10-12Collected
Liumlium.io2026-10-12Collected
Impossible Cloudconsole.impossiblecloud.com2026-10-10Previous dated prices retained; latest collection failed
Synteqsynteq.com2026-10-12Collected
RunPodwww.runpod.io2026-10-12Collected
Lambdalambda.ai2026-10-12Collected
Hyperstackwww.hyperstack.cloud2026-10-12Collected
CoreWeavewww.coreweave.com2026-10-12Collected
Nebiusdocs.nebius.com2026-10-12Collected
DataCrunchdatacrunch.io2026-10-12Collected
Voltage Parksupport.voltagepark.com2026-10-12Collected
Crusoewww.crusoe.ai2026-10-12Collected
Jarvislabsjarvislabs.ai2026-10-12Collected
OVHcloudwww.ovhcloud.com2026-10-12Collected
Spheronwww.spheron.network2026-10-12Collected
Microsoft Azureprices.azure.com2026-10-12Collected
AWSb0.p.awsstatic.com2026-10-12Collected
Together AIwww.together.ai2026-10-12Collected
falfal.ai2026-10-12Collected
Modalmodal.com2026-10-12Collected
Vultrwww.vultr.com2026-10-12Collected
Massed Computemassedcompute.com2026-10-12Collected
TensorDockwww.tensordock.com2026-10-12Collected
SaladCloudsalad.com2026-10-12Collected
Replicatereplicate.com2026-10-12Collected
Fireworks AIfireworks.ai2026-10-12Collected
Basetenwww.baseten.co2026-10-12Collected
DigitalOceanwww.digitalocean.com2026-10-12Collected
GMI Cloudwww.gmicloud.ai2026-10-12Collected
Prime Intellectwww.primeintellect.ai2026-10-12Collected
Cerebriumwww.cerebrium.ai2026-10-12Collected
Koyebwww.koyeb.com2026-10-12Collected
Thunder Computewww.thundercompute.com2026-10-12Collected
Latitude.shwww.latitude.sh2026-10-12Collected
Hot Aislehotaisle.xyz2026-10-12Collected
Scalewaywww.scaleway.com2026-10-12Collected
Google Cloudcloud.google.com2026-10-12Collected

Availability is reported only when the source explicitly publishes stock. “Not reported” means there is no stock observation, even when the price is current. Reserved rates carry a term, minimum volume or an explicit statement that the commitment was not published. Monthly rates use 720 hours; annual Latitude plans use 8,760 hours. Those conversions do not establish payment timing or cancellation rights.

Retired price sources

SF Compute public price history ends with the last public observation on 24 September 2026. Prices have not been public since 25 September 2026. Collection ended on 4 October 2026; this source is excluded from current provider counts.

Hardware specifications

Updated 18 July 2026. Manufacturer specifications describe the named variant, not measured application performance. A peak compute figure is not a benchmark. Unspecified attributes are left out.

Company facts

Rubin sources and specifications

Updated 2026-10-05. GPU figures refer to NVIDIA's individual-GPU table, not a cloud instance or a complete rack. Peak compute specifications are not measured workload performance.

ClaimOfficial sourceQuoted evidence
Rubin individual GPU specificationswww.nvidia.com288 GB HBM4; 22 TB/s; 3.6 TB/s; 4 PFLOPS; 50 PFLOPS
Vera Rubin NVL72www.nvidia.com72 NVIDIA Rubin GPUs
B200 GPU memorydocs.nvidia.comB200 180 GB product
B300 GPU memorydocs.nvidia.com8 x 288 GB = 2.3 TB total
H200 SXM GPU memorywww.nvidia.com141GB
B200 memory typewww.nvidia.com1,440 GB total, 64 TB/s HBM3e bandwidth
B300 memory typewww.nvidia.comHBM3E
NVIDIA launch and cloud partners, 5 January 2026nvidianews.nvidia.comCoreWeave, Lambda, Nebius and Nscale.
CoreWeave production announcementwww.coreweave.comannounced the availability of NVIDIA Vera Rubin NVL72 on CoreWeave
Lambda Rubin product pagelambda.aiAvailable H2 2026.
Spheron pre-order listingwww.spheron.networkShips H2 2026

NVIDIA publishes two sets of per-GPU Rubin figures. Its HGX table lists 22 TB/s of memory bandwidth and 3.6 TB/s of NVLink bandwidth; its Vera Rubin NVL72 table lists 19.2 TB/s and 3 TB/s, for an at-scale AI factory. Ask the provider which system it rents before comparing performance.

Provider fees, access and compliance

Updated 2026-07-07. These are documentation observations, not an audit of a provider. Conditional rates and unpublished fields cannot be turned into a complete bill.

Model parameter counts and memory

Parameter counts come from the publisher’s Safetensors metadata. For standard attention, cache bytes = 2 × layers × key/value heads × head dimension × context tokens × batch size × 2 bytes. Weight bytes = parameters × precision bits ÷ 8. Memory tables use decimal GB, with editable overhead assumptions; they do not guarantee that a particular runtime will load a model.

The full-training scenario uses the Transformers mixed-precision AdamW memory accounting: 18 bytes per parameter, plus activations supplied by the reader. Adapter training applies that accounting only to the assumed trainable parameter fraction. These are budget estimates, not throughput measurements.

ModelPublisherUpdatedCache estimate
Llama 3.1 8Bmeta-llama/Llama-3.1-8B-Instruct2026-10-12Configuration restricted or attention type not modelled
Llama 3.1 70Bmeta-llama/Llama-3.1-70B-Instruct2026-10-12Configuration restricted or attention type not modelled
Llama 3.1 405Bmeta-llama/Llama-3.1-405B-Instruct2026-10-12Configuration restricted or attention type not modelled
Llama 3.2 1Bmeta-llama/Llama-3.2-1B-Instruct2026-10-12Configuration restricted or attention type not modelled
Llama 3.2 3Bmeta-llama/Llama-3.2-3B-Instruct2026-10-12Configuration restricted or attention type not modelled
Llama 3.3 70Bmeta-llama/Llama-3.3-70B-Instruct2026-10-12Configuration restricted or attention type not modelled
Qwen2.5 0.5BQwen/Qwen2.5-0.5B-Instruct2026-10-12Published configuration
Qwen2.5 1.5BQwen/Qwen2.5-1.5B-Instruct2026-10-12Published configuration
Qwen2.5 3BQwen/Qwen2.5-3B-Instruct2026-10-12Published configuration
Qwen2.5 7BQwen/Qwen2.5-7B-Instruct2026-10-12Published configuration
Qwen2.5 14BQwen/Qwen2.5-14B-Instruct2026-10-12Published configuration
Qwen2.5 32BQwen/Qwen2.5-32B-Instruct2026-10-12Published configuration
Qwen2.5 72BQwen/Qwen2.5-72B-Instruct2026-10-12Published configuration
Qwen2.5 Coder 0.5BQwen/Qwen2.5-Coder-0.5B-Instruct2026-10-12Published configuration
Qwen2.5 Coder 1.5BQwen/Qwen2.5-Coder-1.5B-Instruct2026-10-12Published configuration
Qwen2.5 Coder 3BQwen/Qwen2.5-Coder-3B-Instruct2026-10-12Published configuration
Qwen2.5 Coder 7BQwen/Qwen2.5-Coder-7B-Instruct2026-10-12Published configuration
Qwen2.5 Coder 14BQwen/Qwen2.5-Coder-14B-Instruct2026-10-12Published configuration
Qwen2.5 Coder 32BQwen/Qwen2.5-Coder-32B-Instruct2026-10-12Published configuration
Qwen3 0.6BQwen/Qwen3-0.6B2026-10-12Published configuration
Qwen3 1.7BQwen/Qwen3-1.7B2026-10-12Published configuration
Qwen3 4BQwen/Qwen3-4B2026-10-12Published configuration
Qwen3 8BQwen/Qwen3-8B2026-10-12Published configuration
Qwen3 14BQwen/Qwen3-14B2026-10-12Published configuration
Qwen3 32BQwen/Qwen3-32B2026-10-12Published configuration
Qwen3 30B A3BQwen/Qwen3-30B-A3B2026-10-12Published configuration
Qwen3 235B A22BQwen/Qwen3-235B-A22B2026-10-12Published configuration
DeepSeek R1 Distill Qwen 1.5Bdeepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B2026-10-12Published configuration
DeepSeek R1 Distill Qwen 7Bdeepseek-ai/DeepSeek-R1-Distill-Qwen-7B2026-10-12Published configuration
DeepSeek R1 Distill Qwen 14Bdeepseek-ai/DeepSeek-R1-Distill-Qwen-14B2026-10-12Published configuration
DeepSeek R1 Distill Qwen 32Bdeepseek-ai/DeepSeek-R1-Distill-Qwen-32B2026-10-12Published configuration
DeepSeek R1 Distill Llama 8Bdeepseek-ai/DeepSeek-R1-Distill-Llama-8B2026-10-12Published configuration
DeepSeek R1 Distill Llama 70Bdeepseek-ai/DeepSeek-R1-Distill-Llama-70B2026-10-12Published configuration
Mistral 7B v0.3mistralai/Mistral-7B-Instruct-v0.32026-10-12Published configuration
Mixtral 8x7B v0.1mistralai/Mixtral-8x7B-Instruct-v0.12026-10-12Published configuration
Mixtral 8x22B v0.1mistralai/Mixtral-8x22B-Instruct-v0.12026-10-12Published configuration
Mistral Small 24B 2501mistralai/Mistral-Small-24B-Instruct-25012026-10-12Published configuration
Mistral Nemo 2407mistralai/Mistral-Nemo-Instruct-24072026-10-12Published configuration
Phi 3 mini 4kmicrosoft/Phi-3-mini-4k-instruct2026-10-12Configuration restricted or attention type not modelled
Phi 3 medium 4kmicrosoft/Phi-3-medium-4k-instruct2026-10-12Configuration restricted or attention type not modelled
phi 4microsoft/phi-42026-10-12Published configuration
Phi 4 minimicrosoft/Phi-4-mini-instruct2026-10-12Configuration restricted or attention type not modelled
gemma 2 2bgoogle/gemma-2-2b-it2026-10-12Configuration restricted or attention type not modelled
gemma 2 9bgoogle/gemma-2-9b-it2026-10-12Configuration restricted or attention type not modelled
gemma 2 27bgoogle/gemma-2-27b-it2026-10-12Configuration restricted or attention type not modelled
gemma 3 1bgoogle/gemma-3-1b-it2026-10-12Configuration restricted or attention type not modelled
gemma 3 4bgoogle/gemma-3-4b-it2026-10-12Configuration restricted or attention type not modelled
gemma 3 12bgoogle/gemma-3-12b-it2026-10-12Configuration restricted or attention type not modelled
gemma 3 27bgoogle/gemma-3-27b-it2026-10-12Configuration restricted or attention type not modelled

LLM API rates

Serverless billing

Runpod billing documentation. Updated 27 September 2026. Startup, execution and idle timeout are billed, with whole-second rounding; active-worker discounts require a sales inquiry.

Purchase prices

A quote-only listing does not establish a purchase price. Expired or unconfirmed prices are excluded, not assigned a new date.

B200, quote only: Lenovo ThinkSystem NVIDIA HGX B200 180GB 1000W GPU from Eton Technology. Updated 2026-09-27. NVIDIA B200 Tensor Core GPU from Eton Technology. Updated 2026-09-27. Supermicro HGX B200 8U Air-Cooled System from Eton Technology. Updated 2026-09-27. Supermicro HGX B200 4U Liquid-Cooled System from Eton Technology. Updated 2026-09-27. GIGABYTE AI-DLC-Rack NVIDIA-HGX-B200 from Eton Technology. Updated 2026-09-27. GIGABYTE AI-DLC-POD NVIDIA-HGX-B200 from Eton Technology. Updated 2026-09-27. GIGABYTE AI-AC-POD NVIDIA-HGX-B200 from Eton Technology. Updated 2026-09-27. Supermicro SRS-48UAC-B200SX from Eton Technology. Updated 2026-09-27.

GB200, quote only: Supermicro SRS-GB200-NVL72 from Eton Technology. Updated 2026-09-27.

Credits

Price history and corrections

History starts with the first dated observation. Daily records are append-only; identified errors are excluded or reclassified through a separate correction record. Statistical pages state whether they compare the same provider and configuration or a changing market sample.

Read the calculation methods. To report a price, use its row’s email link; it includes the provider, GPU and update date. Nothing submitted through that link is stored by this site.

Data attribution and reuse

The current rental-price dataset is offered under CC BY 4.0, with attribution to GPU Rental Prices. Historical-data licensing is separate. The dated archive has the identifier 10.5281/zenodo.21435394.

Compare your offers

Saved prices keep their original dates. Confirm the configuration, terms and availability at the source.

Select up to three offers to compare their rates and terms.