Ceilings that refuse
$0.50 per instance and $3.00 per pool by default. Both are enforced before any provider API is called, so an offer above either is never bought.
Run inference on the cheapest GPU anywhere
gpuscale shops 7 clouds and marketplaces, serves your model on whichever is cheapest right now, and releases the hardware the moment demand stops.
helm install gpuscale deploy/helm/gpuscale --set providers.vastai.enabled=true- vast.ai · fi-helprovisioned0.284$0.284 per hour. Clears your ceiling.
- verda · nl-amsprovisioned0.331$0.331 per hour. Clears your ceiling.
- runpod · us-orprovisioned0.390$0.390 per hour. Approaching.
- tensordock · denext in line0.472$0.472 per hour. At the ceiling.
- gcp · us-central1-aabove ceiling0.593$0.593 per hour. Refused — never bought.
- aws · us-east-1above ceiling0.718$0.718 per hour. Refused — never bought.
Cheapest first. Nothing right of the dashed line is ever bought, however deep the queue — that is what the ceiling means. A failed offer is held out for 60 seconds; 5 attempts, 3 offers each.
7Providers shopped per request
$0.284Cheapest 4090 in this snapshot
$0.00Cost of an idle pool
10sBatch window before dispatch
30sSpot reclaim detection
Built for a bill you control
Ceilings that refuse, not ceilings that warn
Every pool carries a per-instance and a per-pool maximum. An offer above either is passed over, however deep the queue gets.
Scale to zero
Set minNodes: 0 and an idle pool costs nothing. A 2m cooldown stops a gap between requests from provisioning again.
Spot preemption handled
Provider APIs are polled every 30s. A reclaimed instance is replaced while demand remains, and queued requests wait rather than fail.
No Kubernetes on the GPU
The rented machine runs one agent and one engine. It opens an outbound tunnel, holds no credential of yours, and needs no inbound port.
Bin-packing
When the same models fit on fewer GPUs the controller consolidates them and drains what is left over.
Metrics included
Each worker's engine is scraped and re-exposed, so one endpoint carries control-plane and per-GPU series together.
Engines: vLLM, SGLang, Ollama. Providers: AWS, Azure, GCP, RunPod, Vast.ai, TensorDock, Verda.
The product
Everything on one screen
Capacity, models, cost and chat. The running total never leaves the header, because it is the only number that can hurt you.
gpuscale · gpu-workloadsaccruing $1.005/hr of $3.00
| Claim | Provider | GPU | $/hr | Status | Age | Model |
|---|---|---|---|---|---|---|
| clm-8f31a2 | vast.ai · fi-hel | RTX 4090 · 24 GB | 0.284 | Ready | 41m | glm-4-7 |
| clm-1c07de | verda · nl-ams | A100 · 80 GB | 0.331 | Ready | 12m | qwen3-coder-30b |
| clm-b45119 | runpod · us-or | RTX 4090 · 24 GB | 0.390 | Bootstrapping | 80s | glm-4-7 |
| clm-6ae802 | vast.ai · fi-hel | RTX 3090 · 24 GB | 0.211 | Draining | 2h 06m | — |
| clm-33f9c1 | vast.ai · pl-waw | RTX 4090 · 24 GB | 0.298 | Reclaimed | — | Replaced |
Active requests by model
gpu_api_active_requests_by_model — the series that triggers dispatch and, on the way down, release.
Chat
YouWhich model is cheapest per token today?
glm-4-7 · vast.ai · fi-helglm-4-7 on the 4090 at $0.284/hr, and it is serving you now. qwen3-coder-30b costs 17% more per hour and only wins above roughly 30B.
0
No GPUs running, nothing billing. The pool released its last node six minutes ago, after the 2m cooldown expired. This is the pool working, not the pool broken.