Whisper Batch Cost per Audio Hour: Serverless GPU Guide

If you lead an ML team sizing a transcription backlog, the price tag on a serverless GPU is the wrong number to optimize. Two platforms can quote the same dollars-per-GPU-hour and still differ by 10x on what actually lands in your invoice, because throughput, batching, and the Whisper variant you run move far more money than the rental rate. This guide reframes whisper batch transcription cost per audio hour on serverless gpu around the only metric that maps to spend, then compares the platforms worth shortlisting in 2026.

The metric that matters: $/audio-hour, not $/GPU-hour

Start with the formula. As Lyceum puts it, "Cost per Audio Hour = (Hourly GPU Rental Rate) / (Effective RTF Multiplier)". The denominator is the lever. Real-time factor is the ratio of audio duration to processing time; a 60x real-time factor means one hour of audio transcribes in one minute. Push RTF from 60x to 1,800x and the same GPU rate produces 30x less cost per audio hour.

This is why a headline GPU rate tells you almost nothing about a batch job. Our broader breakdown of GPU inference cost at scale makes the same point for LLMs: utilization, not list price, sets the bill.

How to calculate cost per audio hour for a Whisper batch job

Work two examples with brief figures.

• H100 at high concurrency. Baseten reports an H100 80GiB at $0.10833 per minute (about $6.50 an hour) and, separately, a 1,800x real-time factor (1 hour of audio transcribed in 2 seconds). At 1,800x, $6.50/hr divided by 1,800 is roughly $0.0036 per audio hour. Their faster pipeline hits a 2,400x real-time factor (1.5 seconds per hour of audio), lowering that further.

• Mid-tier GPU, modest batching. Modal demonstrates that batching can achieve a 2.8x increase in inference throughput on an A10G. On a cheaper GPU that 2.8x multiplier is the difference between a viable and a wasteful run. Modal's own worked example transcribes a podcast where "it cost $0.11, so the cost per word was $0.0000105", and on better hardware they note you can process an hour of audio for a cent with L40S GPUs.

The takeaway: always normalize to $/audio-hour before comparing. A step-by-step version of this math for conversational workloads lives in our voice AI agent cost per minute guide.

What drives throughput (and therefore cost)

Three levers dominate the RTF multiplier:

1. Model choice. A turbo or distilled variant trades marginal accuracy for large speedups. Fireworks prices Whisper-v3-large-turbo at $0.0009 per audio minute with 7 sec latency per hour versus $0.0015 and 11 sec for v3-large.

2. Batching. Batching lets a GPU run audio samples in parallel instead of serially, which is where most of the throughput win comes from. Lyceum measured faster-whisper with batch_size=8 and INT8 running about 8.9x faster than the PyTorch baseline, and Mobius Labs reports a 12.5x speed increase over OpenAI's original Whisper and over 3x speed-up compared to Faster-Whisper. The engineering detail behind these wins is in our faster Whisper transcription deep dive.

3. GPU selection. More FLOPs is not automatically cheaper per audio hour. GigaGPU found the RTX 3090 is the lowest cost per audio hour, processing 60 hours of large-v3 transcription per day. Pick the GPU that maximizes RTF-per-dollar for your variant.

Serverless GPU platforms compared

Normalize each quote to a comparable unit. Note that per-audio-minute API prices already bake in throughput, while raw GPU rentals require you to supply the RTF.

Salad's benchmark transcribed 1 million hours of audio transcription for just $5110 on Whisper Large v3, the floor for cost-first batch jobs. For the trade-offs between the two most-shortlisted raw-GPU platforms, see our Modal vs RunPod cost and cold-start comparison, and for managed-endpoint alternatives our Baseten alternatives, Fireworks alternatives, and Replicate alternatives breakdowns.

Which platform wins at your volume

• Spiky, low-volume backlog: a per-minute API (Fireworks turbo, or OpenAI/AWS at $0.006/min) removes all ops overhead and scales to zero.

• Steady high volume: raw serverless GPUs with aggressive batching (Modal, Salad) drop you toward a cent per audio hour, but you own the pipeline.

• Latency-sensitive plus batch: a managed high-RTF endpoint (Baseten) balances both.

The crossover point is simply where your audio-hours per month make the engineering of a batched GPU pipeline cheaper than the API markup. If your workload is conversational rather than offline, concurrency matters more than raw RTF; our concurrent voice sessions per GPU and streaming STT latency guides cover that case.

Self-hosted and API baselines

The managed APIs set a convenient ceiling: OpenAI and AWS both sit at $0.006/min. Anything self-hosted only wins if your effective $/audio-hour beats that after factoring idle time and ops. On infrastructure alone, Deepgram's teardown put Whisper Medium at $0.27 per audio hour on an EC2 L4, which illustrates how quickly a poorly utilized self-host erases the API's apparent premium.

Cerebrium's own voice-agent deployments show how utilization changes the picture: in a production voice pipeline we measured a Deepgram STT service that can run on an A10 GPU and handle roughly 160-180 connections concurrently, where the cost for this on Cerebrium is $1.44 per hour per instance with autoscaling. Spread across that concurrency, the per-stream cost collapses, the same utilization math that governs batch throughput. The full architecture is in our global-scale voice agent write-up, and for offline migration from a managed API our AWS alternatives guide maps the path.

Whichever route you pick, compute $/audio-hour first, then let the RTF multiplier, not the GPU sticker, decide.