What It Costs to Run Whisper Yourself
MAI Transcribe 2 · MAI Transcribe 2
GPU time, electricity, and extra tools — the real bill for local Whisper versus a hosted transcript.
Accuracy tables belong on MAI Transcribe 2 vs Whisper. This article is the invoice you do not see on that page: what it costs to keep Whisper on your own machine, when that bill is worth it, and when a hosted pass is cheaper even if you already “have Whisper installed.”
The numbers below are order-of-magnitude estimates from public cloud GPU list prices and a laptop run in 2026, not a quote.
The local bill has four lines
1. Hardware that is actually busy
Whisper large-v3 on a consumer GPU is fine for a 20-minute file. A week of daily two-hour interviews is a different machine. If you rent an NVIDIA L4 or similar in the cloud so the laptop can sleep, you are paying by the hour whether the job is 8 minutes or you left the instance on. A hosted API only charges the audio you send (on this site, 1 credit ≈ 1 minute, failed jobs free).
2. Electricity and heat you stop noticing
A 200–300 W GPU running three hours in the evening is not “free” because the software is open source. It is a small number until you do it every weekday. Hosted inference moves that power bill to the provider.
3. The tools Whisper does not include
Out of the box, Whisper gives you words. Speaker labels, tidy timestamps for captions, and a phrase list for drug names are extra jobs: pyannote, your own VAD, a subtitle muxer. Budget a day to wire them, then an afternoon every time a dependency breaks. MAI Transcribe 2 on this site returns speakers and timestamps in the same request on paid plans. That is an ops comparison, not a flex.
4. The person who restarts the job at 11 p.m.
Local queues fail silently. Disk fills up. CUDA versions drift. Someone on the team becomes the transcription sysadmin. Hosted tools fail in public (you see an error and retry; this site does not charge a failed paid job). Pick which kind of failure you can afford.
When local Whisper still wins
- The file cannot leave the building. Healthcare, legal holds, a customer contract that forbids processors.
- You already run Whisper at volume and the GPU is amortized. The marginal cost of one more file is close to zero.
- You need a model file you can air-gap.
If that is you, stay local. Do not paste those files into a browser, including this one.
When the hosted pass is the cheaper hour
- You have one ugly file and you need notes before the next meeting.
- You do not want to maintain speaker diarization yourself.
- Your “Whisper setup” is a notebook you last ran six months ago.
In that case, run a short clip on the homepage, then decide. Privacy of the hosted path (no permanent media library) is in the privacy policy. How to edit the draft is in How to transcribe audio with AI.
A worked comparison (illustrative)
Suppose 10 hours of interviews this month.
- Local: one evening installing or updating the stack, 10 hours of GPU time, plus labeling speakers in a second pass. If your time is worth anything, the install evening dominates.
- Hosted on this site: 10 × 60 ≈ 600 credits. Plus includes 300 credits / month; Pro includes 1,200. Or a Work pack (500 credits, never expires) if you do not want a subscription.
If the audio must stay local, those 600 credits are irrelevant. If it can leave, the subscription is often cheaper than the install evening.
What this article is not
It is not a word-error bake-off. FLEURS numbers and same-file clips are on the vs Whisper tool page. It is not legal advice about processors. If you need a DPA, email hello@maitranscribe2.com before you upload anything sensitive.
About the operator: About.