Start
Concepts
The vocabulary and the mental model: instances, models, GPUs, volumes, scale to zero and credits.
What ClockingCloud is
ClockingCloud is a GPU cloud built only for image and video generation. You choose a model, ClockingCloud picks and tunes the GPU, caches your weights and data next to it, and bills per second of GPU time. There are no reservations, and an idle instance costs nothing.
Vocabulary
| Term | Meaning |
|---|---|
| Instance | A named, long-lived definition (model, GPU, scaling, volumes) that runs replicas on demand. Created with clockingcloud deploy or clockingcloud up. |
| Replica | One running copy of an instance, on one or more GPUs. Instances scale between scale.min and scale.max replicas. |
| Run | A one-off generation started with clockingcloud run. It uses a short-lived replica and is not an instance. |
| Request | One generation sent to an instance or run. Every request has an ID like req_01JA7X9K2M. |
| Model | A supported image or video model, referenced by ID such as z-image-turbo, or by hf://org/model. See Models. |
| GPU | L40S, H100, H200 or B200. gpu: auto picks the smallest that fits the model. See GPUs. |
| Volume | Named storage mounted into replicas. Either a ClockingCloud Global Volume or a connected S3, Azure Blob or Google Cloud Storage bucket. |
| Harness | The tuning, caching and storage that ClockingCloud wraps around a model so the model itself needs no changes. |
| Secret | A write-only value (such as WEBHOOK_TOKEN) stored on your account and injected as an environment variable. |
| Credit | Prepaid US dollars. GPU time is deducted from your balance by the second. |
Instance lifecycle
| State | Meaning | Billed |
|---|---|---|
| provisioning | A GPU is being allocated and the image and setup steps are running. | Yes, from GPU attach |
| running | At least one replica is ready for requests. | Yes |
| scaling | Replicas are being added or removed. | Yes, per active replica |
| stopping | GPUs are being released. | Until released |
| stopped | No GPUs attached. Definition and volumes are kept. | No GPU charge |
| failed | Setup failed or the instance could not start. The GPU is released automatically. | No, once released |
Scale to zero and idle timeout
With scale.min: 0 an instance holds no GPU while it has no work. The first request starts a replica from cached weights. After idle_timeout seconds with no work the replica is stopped and billing ends. Raise scale.min to 1 if you need a replica always warm and are willing to pay for it.
What happens during a request
- Queue. The request waits for a free replica. Dynamic batching shares one pass between compatible requests.
- Cold start. If no replica is running, a GPU attaches and the instance starts. Skipped when warm.
- Weight load. Model weights are read from the node cache. The first use of a model on a new node pulls them from the registry.
- Denoise. The model runs its steps. Multi-GPU video models split the sequence across GPUs.
- Decode and write back. Output is decoded and written to your volume or returned to the caller.
clockingcloud trace <request-id> shows the seconds and cost of each stage.
Credits and billing
You prepay by buying credit. Usage draws it down by the second. If your balance reaches zero, running instances are stopped and new requests fail with insufficient_credits. Volumes and definitions are kept. See Billing.
Something unclear or missing? Email hello@clockingcloud.com.