Short answer: for self-hosted Whisper or faster-whisper, a 24–48 GB GPU rented from a GPU cloud by the second or minute is enough. RunPod has the cheapest small cards (RTX 4090 from $0.34 an hour on Community Cloud, $0.74 on Secure Cloud, billed per second), and Thunder Compute has a 48 GB RTX A6000 for $0.35 an hour, billed per minute. Using the faster-whisper project’s own benchmark, one hour of audio needs only a few minutes of GPU time, so the GPU cost is a few cents per hour of audio.
This page contains partner links. See our affiliate disclosure.
Prices come from the providers’ own pricing pages and docs, checked on 7 October 2026. The cost per hour of audio is our estimate from the published benchmark, not a measurement on these exact GPUs.
Table of Contents
GPU prices that matter for Whisper
| Provider | Cheapest suitable GPUs (on demand) | Billing | Storage | Notes |
|---|---|---|---|---|
| RunPod Check price (opens in a new tab) → | RTX A5000 $0.27/h · L4 $0.49/h · RTX 3090 $0.50/h · RTX 4090 $0.74/h (Secure) or $0.34/h (Community) | Per second; needs at least 1 hour of credit to deploy | Volume $0.10/GB/month running, $0.20 stopped; network volume $0.07 | Serverless workers too (24 GB tier $0.69/h, billed per second) |
| Thunder Compute Check price (opens in a new tab) → | RTX A6000 48 GB $0.35/h · L40 $0.79/h · A100 80 GB $1.09/h | Per minute while running | First 100 GB per GPU included; snapshots $0.05/GB/month | No egress fees; CUDA 13 and PyTorch preinstalled; North America |
| Lambda | A100 40 GB $1.99/h | Not stated | Per GB used | Data-centre GPUs only; no egress fees |
| Google Cloud | L4 VM (g2-standard-4) $0.71/h on demand, $0.40 spot | Per second | Separate | A useful reference price |
Prices last checked: 7 October 2026. RunPod’s Secure and Community prices differ; check which one the console shows before you start.
How much GPU time does an hour of audio take?
The faster-whisper README benchmarks the same 13-minute audio file with the large-v2 model on an NVIDIA RTX 3070 Ti 8 GB (CUDA 12.4, beam size 5):
| Implementation | Precision | Time for 13 min of audio | Speed | VRAM |
|---|---|---|---|---|
| openai/whisper | fp16 | 2 min 23 s | about 5.5× real time | 4.7 GB |
| faster-whisper | fp16 | 1 min 3 s | about 12× real time | 4.5 GB |
| faster-whisper, batch size 8 | fp16 | 17 s | about 46× real time | 6.1 GB |
| faster-whisper, batch size 8 | int8 | 16 s | about 49× real time | 4.5 GB |
At about 12× real time, one hour of audio needs about 5 minutes of GPU; batched, about 1.3 minutes. On a RunPod RTX 4090 at $0.74 an hour that is roughly $0.06 per hour of audio unbatched and under $0.02 batched, assuming a 4090 is at least as fast as the benchmark’s 3070 Ti and ignoring model loading. Benchmark your own recordings: phone audio, language and model size all change the numbers.
RunPod: cheapest per-second GPUs
- Per-second billing for pods and serverless workers, no data-transfer fees.
- A 24 GB card (A5000, 3090, L4) runs
large-v3in fp16 or int8 comfortably. - Serverless suits an API with idle periods: you pay from worker start to stop, including cold start.
- Catches: stopped pods still pay $0.20 per GB per month for volume disk; when your balance hits $0, pods stop and pods without a network volume are deleted.
Open RunPod Partner link
Thunder Compute: cheapest 48 GB card
- RTX A6000 48 GB at $0.35 an hour with 6 vCPU and 48–64 GB RAM included; per-minute billing only while the instance runs.
- First 100 GB of disk per GPU included, no egress fees, spending alerts and reminders for running instances.
- Comes with CUDA 13 and PyTorch preinstalled; the docs warn not to reinstall CUDA. faster-whisper uses CTranslate2 CUDA kernels, so run a short test first.
- No T4, L4 or RTX 4090 options; North America only.
Open Thunder Compute Partner link
Running it
- Convert recordings to 16 kHz mono first; it cuts upload and decode time (convert Asterisk recordings).
- Install faster-whisper in a virtual environment and cache the model on a network volume so restarts do not download it again.
- Use batched inference and int8 for long queues; keep beam size 5 for accuracy.
- Delete or stop instances when the queue is empty; set a spending alert.
Our step-by-step setup is in transcribe Asterisk recordings with faster-whisper. For a handful of files, the free call recording transcription tool may be all you need.
Related: Transcribe Asterisk recordings with faster-whisper · Asterisk call recording with MixMonitor · Call recording transcription tool
GPU Cloud at a glance
Frequently asked questions
Which GPU do I need for Whisper large-v3?
faster-whisper runs large models in about 3–6 GB of VRAM in the project’s benchmark, so any 24 GB card (RTX A5000, 3090, 4090, L4) is plenty, with room for batching.
How much does it cost to transcribe an hour of audio on a cloud GPU?
Based on the faster-whisper benchmark (about 12× real time unbatched, about 46× batched) and a $0.74-an-hour RTX 4090, roughly $0.02–0.06 of GPU time per hour of audio, before storage and startup time.
Is a CPU enough?
For small volumes, yes. The README shows the smaller “small” model running about 5–15× real time on an 8-thread Intel i7-12700K. For large models and big backlogs, a GPU is far faster and usually cheaper per hour of audio.
RunPod or Thunder Compute?
RunPod for the cheapest small cards, per-second billing and serverless endpoints. Thunder Compute for a 48 GB card at $0.35 an hour with disk included and no egress fees.
Are call recordings safe on a cloud GPU?
Treat them as personal data: use a provider and region that fit your privacy obligations, encrypt storage, delete audio after processing, and tell callers they are recorded where the law requires it.
Reading a lot of our guides? Go ad-free with srvScripts Pro for $6 a month. It also pays for new guides. Cancel any time.