12 entries · Updated 2026-10-05
Rented GPUs
Where to rent the card you do not own: hourly, per-second and reserved GPU providers, with the billing unit, the regions and the price reference read from their own pages.
12 providers listed
RunPod
Secure Cloud and a community pool of hosts, both billed by the second, plus serverless endpoints and multi-node clusters.
Read on 2026-10-05: per-second prices span about $0.58 to $7.89 per GPU-hour across the catalogue. Community hosts are cheaper and appear and disappear.
runpod.io/pricing per-secondVast.ai
A marketplace rather than a datacenter: hosts rent out their own cards, rates follow supply and demand and are refreshed hourly.
Medians read on 2026-10-05 run from about $0.45/GPU-hour on consumer cards to $7.81 on the newest datacenter ones. Your host is an unknown third party: the machine holds your weights and reads your prompts.
vast.ai/pricing per-secondModal
Serverless GPUs: nothing runs between requests, the container starts with the call and the meter stops when it exits.
Read on 2026-10-05: B300 at $0.001972 per second, which is about $7.10 per GPU-hour. Cheapest per-second rate of the set, and the one that punishes a job that idles.
modal.com/pricing per-minuteHyperstack
H100, A100 and L40 virtual machines billed by the minute, with reserved private clusters for anything larger.
Read on 2026-10-05: H100 SXM $3.20/GPU-hour, H100 NVLink $2.60, A100 SXM 80 GB $1.60, A100 NVLink $1.40; reserved H100 from $1.75. B300 was announced as coming, not listed.
hyperstack.cloud/pricing per-secondOVHcloud
European public cloud with GPU instances, Bare Metal GPU and job-based AI Training, on a per-second or per-minute grid depending on the product.
The prices page is a comparison tool: pick the instance there. The same account can hold your VPS and its mail, so one invoice covers both.
ovhcloud.com/en/public-cloud/prices/ per-hourScaleway
French cloud renting GPU instances from a single L4 up to eight H100 in one instance, with the hourly price and a monthly estimate on the same line.
Read on 2026-10-05: L4 from €0.79/hour, L40S €1.58, H100 from €3.15, up to 8 GPUs per instance.
scaleway.com/en/pricing/gpu/ per-hourVerda
Finnish AI cloud, formerly DataCrunch: single GPUs up to 8-GPU NVLink instances, spot and reserved rates, and instant clusters on InfiniBand.
Read on 2026-10-05: on-demand from about $1.15/GPU-hour up to $10.32, spot below on-demand, and clusters up to 144 GPUs over InfiniBand. The site now brands itself Verda.
datacrunch.io/pricing per-hourNebius
European cloud listing H100, H200, B200 and GB200 with managed Slurm and Kubernetes, and smaller prices on committed capacity.
The prices page shows the GPU set and the discount structure; the per-hour figures sit behind the configuration picker.
nebius.com/prices per-hourTogether AI
Rents GPU clusters by the hour alongside its hosted inference, including GB200 and GB300 NVL72 racks.
Two separate bills on one page: GPU time per GPU-hour, and tokens for the hosted models. Read on 2026-10-05 the hardware table listed GB200, GB300 and H-series racks.
together.ai/pricing per-hourLambda
Instances, 1-Click Clusters and Superclusters, with the plainest hourly list prices of the set and reserved capacity on request.
Read on 2026-10-05: B200 SXM6 180 GB $6.69/GPU-hour, H100 SXM 80 GB $3.99, A100 SXM 80 GB $2.79, A100 SXM 40 GB $1.99.
lambda.ai/pricing reservedDigitalOcean
GPU Droplets next to the ordinary cloud furniture: H100, H200, B300 and MI300X, on-demand or on a twelve-month reserved plan.
Reserved prices read on 2026-10-05: B300 $7.94/GPU-hour, H200 $3.40, H100 $3.26. Spot droplets are interruptible with roughly two hours' notice.
digitalocean.com/pricing/gpu-droplets reservedCoreWeave
Large NVIDIA clouds with on-demand, spot and reserved capacity, GB200 NVL72 racks included, and managed Kubernetes that expects to be the whole estate.
Built for long commitments: the interesting prices are the reserved ones, and the page leads with the storage, networking and Kubernetes lines that come with them.
coreweave.com/pricingWhat renting is for
A local rig answers most of the questions this site is about, and none of the ones that need more memory than you can buy. A 1M-token context on a 30B-class mixture-of-experts is about 90 GB of cache before the weights; a frontier MoE checkpoint is several hundred gigabytes. Both are rented problems, not upgrade problems.
Three things change when the card is someone else’s:
- The data leaves the machine. Weights and prompts travel over the network and land on a machine you do not control. On a marketplace host you also do not know who that is. That is a decision, not a detail.
- The meter runs on time, not on work. A per-second provider bills the container’s life, so a job that idles costs the same as a job that computes. A reserved contract bills whether you are there or not.
- Storage and egress are separate lines. A 1 TB checkpoint that stays resident on their NVMe is billed monthly; pulling it out again is billed by the gigabyte. The 12 cards below list compute only.
How to read this page
Each card is a provider, and the two filter rows do the work: the first by billing unit (how fine the meter is), the second by which GPU families that provider actually lists. Prices are the providers’ own, read on the date inside each note — they move, and no aggregator figure is kept here. “Reserved” means the published headline price is a commitment, not a card you can have in an hour.