Raw compute · interactive GPU podsLive capacity on /status

A whole Saudi GPU. Yours in about a minute.

Pick an image, get Jupyter in your browser and root SSH on a dedicated GPU — train, fine-tune, or serve, then tear it down. No queue, no commitment, data in the Kingdom.

Launch a pod →See live capacity

§ 01 · The 60-second pathCLI · API · web console — same rails

From your terminal

Three commands. No account manager, no sales call.

# install the dc1 SDK + CLI
curl -sL https://api.dcp.sa/installers/dc1-sdk.tar.gz | tar xz && cd dc1-sdk && ./install.sh

# launch: dedicated GPU, PyTorch image, 2 hours
dcp pod create --image pytorch --duration 120

#   → access_url:  https://api.dcp.sa:41xxx/?token=…   (Jupyter, TLS)
#   → ssh_command: ssh -p 42xxx root@api.dcp.sa

Prefer clicking? The web console launches the same pod. Prefer raw HTTP? POST /api/pods with your renter key. All three paths hit the same verified scheduler.

01choose_image

PyTorch, vLLM, CUDA, Ubuntu — or any public Docker image. Unknown images get SSH bootstrapped automatically.

02verified_machine_boots

The scheduler only considers providers that just passed live Docker, CUDA, and GPU-health probes.

03jupyter_tls + root_ssh

Your access URL and SSH command arrive in about a minute. One-time credentials, shown once, stored nowhere.

§ 02 · Why rent hereWhat you actually get

zero_setup · ≤ 60s

From idea to training in a minute

Open Jupyter in the browser or SSH straight in. Nothing to install, no ticket queue, no GPU waitlist — the machine is already provisioned, probed, and waiting.

--gpus all · pinned driver

The whole card is yours

No sharing, no throttling, no noisy neighbors. Benchmarks run at bare-metal speed and reproduce tomorrow — we freeze driver updates mid-rental so the ground doesn't move under you.

hard deadline · restart-proof reaper

It ends when you said it ends

The host machine itself enforces your rental's deadline — even across crashes and reboots. A forgotten pod can never squat a GPU or surprise you later.

wireguard mesh · nvml gates

Verified Saudi machines only

Pods land exclusively on hardware that just passed live health probes — the same earned-online discipline behind our inference catalog. Your data stays in the Kingdom.

§ 03 · What runs well on 24 GBConsumer GPUs, honestly framed

unsloth · lora · qlora

Fine-tuning 7–13B models

LoRA and QLoRA runs on RTX-class cards are the sweet spot — hours, not days, at a fraction of hyperscaler cost.

pytorch · cuda 12 · shm tuned

Research & experiments

A real CUDA box you fully control — profile kernels, pin versions, break things, reset in seconds.

vllm · llama.cpp · ollama

Private model serving

Serve your own checkpoint behind your own endpoint — quantized 7–32B models run well on a dedicated 24 GB card.

any public docker ref

Anything Docker runs

Rendering, scientific compute, CI with GPU tests — if it's a container that wants a GPU, it boots here.

§ 04 · Honest engineeringWhat we promise — and what we don't, yet

Promised today

Every claim on this page is enforced by code you can audit.

Health-gated scheduling, TLS on the Jupyter relay, deadline enforcement on the host, server-measured usage, one-time credentials. Capacity numbers on this site come from live probes — never from static copy.

!1host_pinned

Pods live and die with their host machine. There is no live migration or cross-host failover — checkpoint your work.

!2container_isolation

Isolation is hardened Docker today; VM-grade sandboxing (gVisor) is on the roadmap before general availability.

!3small_verified_fleet

The mesh is young. Capacity is small and stated honestly — check /status before planning a large run.

Ready? Launch your first pod from the console, or start with the docs. And if you have an idle GPU — the provider path pays you for its time.

Questions about the rails? The renter docs cover pods, keys, and billing.Read the docs →