01The problem
Every new model waited for hardware.
The studio trains a new base model each quarter and fine-tunes it for clients in between. Their previous provider could hand out single 8-GPU machines on short notice, but anything bigger meant a sales call and a waiting list.
That wait ended up setting the release calendar. The team needed:
- Multi-node capacity running on the day it was requested
- One EU region, because client footage is not allowed to leave it
- Per-second billing, so evaluation jobs between runs cost only what they use
02The solution
One cluster for training, single servers for the rest.
Training now runs on an Instant Cluster of eight B300 nodes in EU-AMS-1: 64 GPUs on one fabric. Between runs the team deletes the cluster and keeps two single B300 servers for evaluation and client fine-tunes.
nexctl — train
Checkpoints live on a shared volume in the same region. A run can stop on Friday and resume on a fresh cluster on Monday without moving a byte.
Placeholder
Placeholder quote. Two plain sentences from the customer's engineering lead about what changed, approved before publishing.