Velar Cloud
A serverless GPU cloud we built and ran in 2025–26. What it was, what it taught us, and where it stands.
What we built
Velar Cloud let a Python function run on a GPU with one decorator. Under it: a per-second billing engine that metered every GPU job to the exact second and scaled to zero between calls; a content-hash image cache that cut redeploys to under 15 seconds; parallel fan-out of one Python call across 50 GPUs; and persistent endpoints that kept models warm in VRAM, so production traffic never paid a cold start.
What running it taught us
GPU idle time is the bill. Most of what customers paid for was capacity waiting for work, and per-second metering made that waste visible for the first time.
Rate limits and cold starts are the outages. Almost every incident we handled was one of the two, not a hardware failure. The tooling that mattered was the boring kind: utilization dashboards, spend attribution per job, and runbooks for the failure modes that kept coming back.
Where it stands
Velar Cloud is paused in 2026 while the team does reliability and cost engineering for clients. Its successor, Velar Gateway — cost and reliability controls for LLM traffic — is in early access.