The only metric: cents per audio hour
Whisper fits on anything — large-v3 is about 3 GB in float16 — so VRAM never decides. What decides is throughput × price: how many hours of audio a card transcribes per hour of rent. With faster-whisper's batched pipeline the spread between a T4 and an H100 is roughly 10× in speed but only about 2× in cost per audio hour, because the fast cards cost more per hour. Pick by the last column below, then by whether the job is a batch or a service.
Cards ranked by cost per audio hour
Indicative speeds for faster-whisper large-v3, float16, batched pipeline (batch 16, VAD on), expressed as multiples of real time; your audio mix, batch size and int8 settings move them. Prices are today's on-demand rates.
| Card | On-demand | Speed (× real time) | Audio hours per GPU-hour | $ per audio hour | Note |
|---|---|---|---|---|---|
| Tesla T4 16 GB | $0.104 | ~18× | 18 h | $0.0058 | Cheapest datacenter card; int8 helps |
| RTX 3060 12 GB | $0.041 | ~28× | 28 h | $0.0015 | Cheapest consumer card that fits large-v3 easily |
| L4 24 GB | $0.225 | ~40× | 40 h | $0.0056 | 72 W, single-slot — the always-on API card |
| RTX 4070 12 GB | $0.066 | ~45× | 45 h | $0.0015 | Ada speed at a low rate for batch jobs |
| RTX 4090 24 GB | $0.262 | ~95× | 95 h | $0.0028 | Batch transcription workhorse |
| A100 PCIE 80 GB | $0.662 | ~110× | 110 h | $0.0060 | Many parallel streams, 80 GB for big batches |
| H100 PCIE 80 GB | $2.147 | ~170× | 170 h | $0.0126 | Highest throughput per card |
Interruptible capacity halves every number in the price column — transcription queues are the ideal interruptible workload, since every file is an independent, restartable item.
Which card for which job
| Job | Rent | Because |
|---|---|---|
| A podcast archive, thousands of hours, once | RTX 4090, interruptible | Lowest $/audio-hour of the consumer cards; batch tolerates pauses |
| An always-on transcription API | L4, reserved | 72 W datacenter card, −35% reserved, ~40× real time is plenty for streaming |
| Meetings for a small team, evenings only | RTX 3060 | Cheapest card that runs large-v3 comfortably; stop it when idle |
| Thousands of concurrent streams | A100 PCIE or H100 PCIE | 80 GB for big batches, highest aggregate throughput |
| Diarization + transcription pipelines | RTX 4090 or L4 | Pyannote models add ~2 GB; both cards have the room |
Setup: WebUI and API in 30 seconds
powergpu launch --gpu l4 --template whisper-webui-api --disk 40 --volume audio:/data
# ✓ instance i-5d0a91f2 running (27.3s) · $0.225/hr
# https://i-5d0a91f2.powergpu.io:7860 (Web UI) · :8000/transcribe (REST)
curl -F 'file=@call.mp3' -F 'model=large-v3' https://i-5d0a91f2.powergpu.io:8000/transcribeThe Whisper WebUI & API template runs faster-whisper behind a browser UI and a REST endpoint with word timestamps, language detection and subtitle export. Mount a volume for input files and transcripts so instances stay disposable.
Squeezing more out of any card
- Batch. The batched pipeline is the difference between 8× and 90× real time on the same 4090; single-file sequential decoding wastes the card.
- int8 on small cards. On the T4 and 3060, int8 weights run faster with negligible accuracy loss; keep float16 on Ada and newer.
- VAD first. Voice activity detection skips silence — 20–40% less compute on meeting audio.
- Right-size the model. turbo and distil-large-v3 run 4–6× faster than large-v3 for English at a small accuracy cost.
- Interruptible for queues, reserved for APIs. Batch jobs pay half; a 24/7 endpoint pays −35% and never sleeps.
Related: serving LLMs on the same cards and the L4 vs T4 comparison.
Put the numbers to work
Every price in this guide is our live rate — fixed, ≥30% under the market median, billed per second. Deploy the exact setup above from the console in about 30 seconds, paid in crypto, no card and no KYC.
