ClockingCloud provisions GPUs optimized for image and video generation. Install it, run your model, ship.
One command. Starts in seconds. Pay per second.
Video
Image
Image
ImageIllustrative examples of what you will generate.
01The platform
ClockingCloud sits between your code and the GPUs. You send a request. We handle everything in the middle.
Your workloads
One layer. Everything below is handled.
GPUs
L40S
48 GB
Fast images
H100
80 GB
Images and video
H200
141 GB
Large video
B200
192 GB
The biggest jobs
0
No drivers, containers, clusters or autoscalers. You bring the model. We run everything under it.
100%
Warm pools, autoscaling, retries and tracing from the very first request.
1
Pay per second of GPU. It scales to zero, so idle time costs nothing.
Targets, not measured results.
02Why ClockingCloud
General-purpose clouds rent you a machine and leave the rest to you. ClockingCloud is tuned for one job: generating pixels, fast and cheap.
Setup
Drivers, CUDA, containers and configs.
Cold start
Minutes to load model weights.
GPU choice
You guess, then overpay.
Idle cost
You pay for boxes doing nothing.
Setup
One command.
Cold start
Under 2 seconds, with warm pools.
GPU choice
Picked for every job.
Idle cost
Scales to zero.
Cold start and scaling figures are launch targets.
03Config as code
Everything is a YAML file. Check it in, review it, and deploy it from CI/CD. Use the API to provision, start, stop and manage instances.
# clockingcloud.yaml
name: product-shots
model: z-image-turbo
gpu: auto # L40S, H100, H200 or B200
scale:
min: 0
max: 8
idle_timeout: 60s # stop the GPU when it goes quiet
setup:
- pip install -r requirements.txt
- python warm_cache.py
storage:
- mount: /data
source: s3://brand-assets
- mount: /out
volume: global://renders# provision from CI/CD
curl -X POST https://www.clockingcloud.com/v1/instances \
-H "Authorization: Bearer $CLOCKINGCLOUD_API_KEY" \
-H "Content-Type: application/yaml" \
--data-binary @clockingcloud.yaml
POST /v1/instances/{id}/start
POST /v1/instances/{id}/stop
GET /v1/instances/{id}clockingcloud dev
Run and fix your setup scripts in a CPU sandbox with the same image. A GPU only attaches when your job really needs it.
clockingcloud plan
Validate the YAML and see the GPU and the estimated cost before anything starts.
idle_timeout: 60s
If setup breaks or the instance goes quiet, the GPU is released. No idle bill while you debug.
04The harness
ClockingCloud wraps the model with the serving, caching and storage around it. Connect the drives you already use, or use native ClockingCloud Global Volume storage.
clockingcloud drive connect s3://bucket
clockingcloud drive connect azure://container
clockingcloud drive connect gs://bucket
clockingcloud volume create renders --size 500
Native
ClockingCloud harness
Open models run with optimized settings for each GPU.
Datasets and weights are cached next to the GPU.
Checkpoints and renders land back on your drive.
Your models, your data, one config.
05Models
Not a general-purpose cloud. For each launch model you see what it is best at, what it outputs, how fast we aim to run it and which GPU we pick for it.
Photos, product shots, portraits and art.
Fast photoreal images and product shots
Editing an image from a text instruction
Images with accurate text in them
Clips from a prompt or from a single image.
Text to video and image to video
Cinematic motion and detail
The fastest drafts and previews
Point the CLI at a folder. We handle the rest.
clockingcloud deploy ./my-modelSpeeds are targets, not measurements. Images are illustrative.
06How it works
brew install clockingcloudOne command sets up everything.
clockingcloud run <model>Pick a model. We pick the GPU.
clockingcloud topWatch it scale up, then down to zero.
$ brew install clockingcloud ==> Installing clockingcloud ==> clockingcloud is ready $ clockingcloud login -> signed in $ clockingcloud run z-image-turbo --prompt "a neon city at dusk" --gpu auto -> provisioning H100 80GB (us-east) ... warm in 1.8s -> compiled kernels, fp8, cached weights -> image ready in 0.9s $0.0021 $ clockingcloud run wan-2.2 --video --seconds 5 -> scaling to 4x H200 ... 24 fps ... done in 38s
07Performance
Every layer is tuned for image and video generation.
Launch targets. Goals, not measured results.
torch.compile and custom kernels, tuned per model.
FP8 and INT8 quantization that keeps quality.
Cached weights and warm pools skip the wait.
Dynamic batching keeps every GPU busy.
Multi-GPU sequence parallelism splits long clips.
Picks L40S, H100, H200 or B200 for each job.
08Observability
Built in from the first request. No setup, no extra tools. Four questions, answered live.
p50 and p95 latency for every model.
Utilization and queue depth, live.
Cost per image and per second of video.
A trace for every request, from queue to upload.
clockingcloud logs --followclockingcloud topOr watch it all from your terminal. Mockup with simulated numbers.09Pricing
Pay for GPU seconds, not idle hours.
No traffic means nothing on the bill.
No reservations and no minimums.
The platform is opening to the public. Create your account, add funds, and run your first GPU for pixels.