Cold starts in seconds
Time to first response
Concept figures
Warm before you ask.
Containers come back from a memory and GPU snapshot, so a scale-up from zero answers in about three seconds, not a minute.
AI infrastructurewith aclean edge
Run voice agents, video models and LLMs on serverless GPUs. Cold starts in seconds, scaling that follows your traffic, nothing to babysit.
Concept customers
- Parlance
- hollowbell
- NORTHFIELD
- Quill & Co.
- Tandem
- Ostara Labs
- brightwater
Why Selvage
The speed of a prototype with the discipline of production. GPUs that are ready when the request is.
The platform
Elastic GPU scaling
Regions online
us-east1,240 GPUs11 ms to the nearest user
Capacity when you need it.
Draw on thousands of GPUs across regions and clouds. Scale from zero to a burst and back, and pay for the seconds you use.
Bring your own stack
$
ok Packed 2 files (5.0 KB)
ok Built the image in 6.4 s
ok Snapshot ready, 0 replicas idle
ok Live at https://voice-agent.example/v1
$
One command to production.
Ship a Python file or a Dockerfile. Any framework, any weights, no platform lock-in. Your code stays yours.
See every request
voice-agent, last 24 hours
requests per minute
- p95 latency
- 212 ms
- GPUs in use
- 38
- Cost today
- $412
Every token and frame, traced.
Live traces, GPU use and cost per request in one view, with alerts before your users feel a slowdown.
Use cases
Built for workloads that do not wait.
Answer in under 300 ms
Speech in, speech out, with turn-taking that feels like a person on the line.
212 ms p95Transcribe an hour in a minute
Fan a long recording out across GPUs and get the transcript back while the kettle boils.
60x real timeJournal
From the workshop
How we restore a GPU in three seconds
What a cold start actually costs