Ombra renders avatar video in under 400 ms per frame

Background

Ombra builds on top of real-time models and needed infrastructure that could follow its traffic without a platform team.

Before the move, the team ran reserved GPU nodes in one region and paid for idle time between peaks.

The problem

Traffic changed tenfold within minutes. Reserved nodes were either idle or saturated, and adding a region took weeks of work.

1. Burst capacity. New replicas within seconds during spikes.
2. Low network latency. Model calls close to users.
3. Data residency. EU traffic on EU hardware.
4. Fast iteration. New model versions without infra work.

Why Axonfield

Per-request autoscaling removed the capacity planning. Memory snapshots brought cold starts under three seconds, so the team no longer kept warm replicas running.

Region pinning let them serve EU customers from EU hardware with one config change.

"We stopped talking about capacity in planning meetings. It just follows the traffic now."

- Head of engineering

Results

Within two months Ombra ran all production inference on Axonfield, with lower cost and a shorter release cycle.