01The problem
GPUs ran all night for traffic that stops at six.
The lab's customers are contact centres, and contact centres keep office hours. Almost all audio arrives on weekdays between eight and six, yet the lab rented a fixed fleet sized for the Monday-morning peak.
- Idle capacity most evenings and every weekend
- Queues at peak, because the fleet was still too small for the busiest hour
- Two regions required by customers in different EU countries
02The solution
An endpoint that grows with the queue.
Transcription moved to a Serverless endpoint on H100 workers in EU-AMS-1 and EU-FRA-1. The autoscaler adds workers when the queue grows and removes them when it drains, down to zero overnight.
regions · europe
2 live · 1 planned
EU-AMS-1
EU-FRA-1
EU-HEL-1soon
- Live
EU-AMS-1
Amsterdam
- Live
EU-FRA-1
Frankfurt
- In build
EU-HEL-1
Helsinki
- Billing in EUR for every region
Model weights sit on a volume in each region, so a new worker loads them from local storage instead of the network.
Placeholder
Placeholder quote. Two plain sentences from the customer's engineering lead about what changed, approved before publishing.