Real workloads. Real numbers.
Every story below was written with the team that runs the workload. The figures come from their dashboards, over at least 90 days in production, and each story keeps the section on what did not go smoothly.
-
Media
Media group, archive team
900 hours of audio per night · 14 languages · 3 regions
transcribing 900 hours of audio a night on capacity that goes to zero by morning
How a media group moved its overnight transcription batch onto Terrane, cut the cost per audio hour by more than half and stopped babysitting failed jobs.
- Cost per audio hour
- 100 43
- Nightly backlog cleared in
- 9.5 h 3.1 h
- GPU hours per night
- 210 96
- Jobs re-run by hand
- 6% / night 0
-
Fintech
Payments provider, risk team
2,100 authorisations/s · 6 regions · card data residency in the EU and Brazil
fraud scoring on every card authorisation, inside a 40 ms budget
How a payments provider moved its fraud model from one region to six, kept every decision under 40 ms and stopped failing open when the model was slow.
- p99 decision time
- 71 ms 34 ms
- Decisions failed open
- 0.8% 0.02%
- Regions serving traffic
- 1 6
- Pages to on-call per month
- 14 2
-
E-commerce
Marketplace search team
38M listings · 14,200 requests/s at peak
a semantic search API at 14,000 requests per second
How a marketplace team moved embeddings, reranking and their public search API onto Terrane, cut p99 from 340 ms to 98 ms and stopped running a control plane.
- p99 latency
- 340 ms 98 ms
- GPU utilisation
- 38% 79%
- Cost per 1M queries
- 100 61
- Deploy pipelines
- 3 1
How these are written.
Facts before interpretation
Tables and configuration come from the customer's own systems. Sentences about what it meant for them are theirs, quoted, or ours, marked as such.
Ninety days minimum
Nothing is published before the workload has run a full quarter in production. Launch-week numbers are not results.
The part that went wrong
Every story has a section on what did not go smoothly. If a migration had no such section, we would not believe it either.
Your workload, same treatment.
Send us the shape of what you run. An engineer replies with regions, a hardware tier and a capacity estimate, and tells you if it is not a fit.