Serverless GPU Providers

Which serverless GPU providers cost the least?

A serverless GPU is billed while your code holds it and stops costing when it scales to zero. Below are the serverless rates on rent.dev's price board, cheapest first, each with its source and the day it was read.

The lowest serverless rate on rent.dev's price board is $0.58 per GPU hour, for the NVIDIA RTX 2000 Ada Generation at RunPod, read . The board lists 2 serverless GPU providers.

See all GPU prices

Every row on rent.dev's board links its source. GPU pods run on your own provider key, and that provider bills you.

Serverless GPU rates, cheapest first

Up to 6 of each provider's lowest serverless rates, one row per GPU, per GPU hour, from rent.dev's price board, generated . The board marks 64 GPU rows serverless: 29 rows from RunPod and 35 rows from Verda.
GPUMemoryProviderPer GPU hourBasisBilled bySourceRate read
NVIDIA RTX 2000 Ada Generation16 GBRunPod$0.58listedRunPodapi.runpod.io
NVIDIA RTX 4000 Ada Generation20 GBRunPod$0.58listedRunPodapi.runpod.io
NVIDIA RTX A400016 GBRunPod$0.58listedRunPodapi.runpod.io
NVIDIA RTX A450020 GBRunPod$0.58listedRunPodapi.runpod.io
NVIDIA GeForce RTX 309024 GBRunPod$0.75listedRunPodapi.runpod.io
NVIDIA L424 GBRunPod$0.75listedRunPodapi.runpod.io
RTX 6000 Ada 48GB48 GBVerda$1.71375listedVerdaapi.datacrunch.io
A100 SXM4 40GB40 GBVerda$1.841listedVerdaapi.datacrunch.io
L40S 48GB48 GBVerda$2.1975listedVerdaapi.datacrunch.io
A100 SXM4 80GB80 GBVerda$2.5725listedVerdaapi.datacrunch.io
RTX PRO 6000 96GB96 GBVerda$3.027listedVerdaapi.datacrunch.io
H100 SXM5 80GB80 GBVerda$5.595listedVerdaapi.datacrunch.io

Serverless GPU in one paragraph

Serverless GPU products run your code in a container that gets a GPU when a request arrives and lets it go when traffic stops. You pay for the time the container holds the GPU, not for a machine that sits idle. That suits inference with uneven traffic. Its hourly rate is higher than a pod you keep running, so steady work is often cheaper on a pod. rent.dev's GPU API starts pods, not serverless workers: a pod runs on your own key at the GPU provider rent.dev connects to, and that provider bills you.

How to read the table

Each rate is per GPU, per hour, as the provider lists it. A row that prices a group of GPUs is divided by its GPU count, so you compare one GPU with one GPU.

Basis says where a figure comes from. Listed is the provider's own list price; the other words name a figure worked out from the provider's data, such as a median or a floor.

Billed by names who charges you for that row. The read date is the day the board last read that provider's rates, so check the source before you plan a big job.

When serverless is the cheaper choice

Add up the hours your GPU really works in a day. If it works for a small part of the day, serverless usually costs less, even at a higher hourly rate, because you pay nothing while traffic is quiet.

If the GPU is busy most of the day, a pod you keep running usually wins. Cold starts matter too: a large model can take a while to load, and that wait is often on the meter.

Read on: What serverless GPU is, and how its meter runs.

Start with the agent quickstart

Agents can start at rent.dev/llms.txt, which lists every route.