Inference and agents, in the KingdomPay per token · Saudi RiyalDCP-Agent for Saudi business · agents.dcp.saAgents can rent a GPU · npx -y github:dhnpmp-tech/dcp-mcpEarn Riyal from your GPUPDPL · Saudi data residencyInference and agents, in the KingdomPay per token · Saudi RiyalDCP-Agent for Saudi business · agents.dcp.saAgents can rent a GPU · npx -y github:dhnpmp-tech/dcp-mcpEarn Riyal from your GPUPDPL · Saudi data residency
DCP
DC Power Solutions · CR 7053667775Start free →

Illustrative network footprint

§ DCP · Sovereign Arabic AI Runtime · KSA

Saudi Arabia's open GPU cloud.

AI by the token. Whole GPUs by the minute. Verified live, billed in Riyal — your data never leaves the Kingdom.

Proof you can touch — ask a Saudi GPU right now
🇸🇦 Inference · KSA🇸🇦 Agents · KSA🇸🇦 GPUs · KSA🌐 Frontier · opt-in only
DCPSovereign Arabic AI · built in the Kingdom
🇸🇦 Saudi Arabia · data does not cross this line
you sendYour Arabic prompt
verified · in-KingdomA Saudi GPU answers
you receiveYour Arabic answer

Your prompt and its answer stay inside the Kingdom — on a verified GPU, never a foreign cloud. Cross-border frontier models are available by explicit per-tenant opt-in only. See the full architecture →

Capacity is published only after live provider verificationNo simulated telemetry
Live availability
Gated by /status
Serving right now — live from /v1/modelsquerying…
No verified capacity is serving right now, so nothing is listed. That is the honest state — not an error.Watch live status →
Prefer the whole card? RTX-class · 24 GB · dedicated — rent it by the minute, in SARLaunch a pod →
Rent one of these GPUs — spin up your own podOpen the launch console →

Your own GPU. By the minute.

Spin up a dedicated pod — root, Jupyter and SSH — in about two minutes. OpenAI-compatible inference, agent-ready via MCP, billed in Saudi Riyal, sovereign and in-Kingdom.

  • ~2-min launch
  • per-minute, cost-plus
  • root + Jupyter + SSH
  • MCP + OpenAI API
  • in-Kingdom
What the public marketplace means

No provider is listed until the inference path itself is proven.

Most GPU lists are typed in by hand — and go stale. This one cannot be typed in: a machine appears only after our backend has reached it, asked it a real question, and verified the answer. The moment any check fails, the machine disappears from the list instead of rotting on it.

01We can reach it endpoint_reachable

Our backend connects to the machine over the private mesh — right now, not at sign-up time.

02It really answers verified_online

We send the machine a real question and verify a real answer comes back. A heartbeat alone earns nothing.

03It serves what it claims model_coverage

A model is listed only while a verified machine is actually serving that exact model.

If those checks fail, the right public state is empty capacity plus a route to status, not a stale GPU list.Check live status →
Tokens for answers — or the whole GPU for control. The capacity gates above govern real inference. If you want raw compute instead, GPU pods are one section down. And if you want to bring hardware as a provider, start with the provider path.
Inference and raw compute

Buy tokens when you want answers. Rent the whole GPU when you want control.

A DCP pod is a dedicated GPU container on the same verified mesh: pick an image, get Jupyter over TLS and root SSH in about a minute, train or fine-tune, tear it down. The same health gates that protect inference decide which machines may host your pod.

01zero_setup · jupyter + root_ssh ≤ 60s

From idea to training in about a minute. Open Jupyter in your browser or SSH straight in — nothing to install, no ticket queue, no GPU waitlist.

02--gpus all · pinned driver

The whole card is yours. No sharing, no throttling, no noisy neighbors — your benchmarks run at bare-metal speed and reproduce tomorrow, because we even freeze driver updates mid-rental.

03hard 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 — so a forgotten pod can never squat a GPU or surprise you later.

04wireguard mesh · live docker/cuda/nvml gates

Verified Saudi machines only. Your pod lands exclusively on hardware that just passed live Docker, CUDA, and GPU-health probes — the same earned-online discipline behind our inference catalog. Your data stays in the Kingdom.

Same mesh, same verification, same KSA residency — applied to raw compute. Pods launch only on providers that pass live Docker + CUDA + GPU-health probes.dcp pod create →

Every idle GPU in the Kingdom is a data center.

Hyperscalers build walls around compute. We build rails between the GPUs the Kingdom already owns — gaming rigs, workstations, university clusters — verify each one live, and put it to work serving Arabic-first AI and raw compute, paid in Riyal. The numbers below are the mesh as it exists this minute, not a projection.

models serving this minute
models in the catalog
2products: tokens + whole GPUs
NOW · live today

The verified mesh

WireGuard-meshed providers, earned-online catalog, token billing in SAR, interactive GPU pods, Arabic-first models.

NEXT · building now

Production hardening

Pod billing by the minute in SAR, VM-grade isolation (gVisor), card payments switched on, a larger verified fleet.

THEN · the bet

Sovereign AI at scale

ALLaM at scale, a datacenter tier, and the long tail of Saudi expert models — trained, served, and paid for inside the Kingdom.

01 · Inference

One API. Per-million-token billing.

OpenAI-compatible chat, embedding, and rerank endpoints, served from KSA-resident GPUs. Arabic-first, open-source model lineup. Frontier models stay off unless you opt in.

  • OpenAI SDK · no rewrite needed
  • Streaming · function calling · JSON mode
  • Halala-grained billing · SAR + USDC
api.dcp.sa / v1See rates →
02 · Agents

DCP-Agent. Live for SMB.

The Arabic AI agent for Saudi small & mid-size businesses. Already in production at agents.dcp.sa. Free personal version for every Saudi is coming.

agents.dcp.saVisit →
03 · Providers

Earn SAR with your GPU.

A 4 MB desktop app for Windows, macOS Apple Silicon, and Linux. Auto-detects your GPU, installs the inference engine (Ollama or MLX), downloads a model, and reports measured throughput after verification. Joins a self-hosted WireGuard mesh — no port forwarding.

  • Windows · macOS Apple Silicon · Linux
  • 4 MB app · zero config · WireGuard mesh
  • 85% provider · 15% platform · monthly SAR payout
dcp.sa / v2 / provider-setupRegister a GPU →

Three lines from nothing to an Arabic answer.

Drop your existing OpenAI SDK in — only the base URL and key change. The official dcp CLI ships with a one-shot inference command and per-tenant config. Python and Node SDKs are thin wrappers; you can swap in for OpenAI’s any time.

OpenAI-compat · v1Streaming · SSEFunction callingJSON mode200k context · frontier
# Arabic chat completion · OpenAI-compatible · KSA-resident
$ curl https://api.dcp.sa/v1/chat/completions \
  -H "Authorization: Bearer $DCP_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3-4b",
    "stream": true,
    "messages": [
      { "role": "user",
        "content": "ما حكم زكاة الراتب الشهري إذا لم يبلغ النصاب إلا بعد جمعه لسنة؟" }
    ]
  }'

# → streams Arabic answer with cited sources (verifier-checked)
# → settled in halala against your wallet at end of stream
# One-line install · macOS · Linux · Windows
$ npm install -g @dcp/cli

# Authenticate once · stores key in ~/.dcprc (encrypted at rest)
$ dcp login
  → Opens console.dcp.sa to fetch a tenant-scoped key

# Run inference · streams Arabic answer to stdout
$ dcp run "ما حكم زكاة الراتب الشهري؟" \
        --model qwen3-4b \
        --cite

# Pin sovereign-only · frontier blocked even if router prefers it
$ dcp config set sovereign_only=true

# Show last 24h spend · per-model breakdown
$ dcp usage --since 24h
# pip install openai · standard OpenAI SDK works as-is
import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.dcp.sa/v1",
    api_key=os.environ["DCP_KEY"],
)

stream = client.chat.completions.create(
    model="qwen3-4b",
    messages=[{
        "role": "user",
        "content": "ما حكم زكاة الراتب الشهري؟",
    }],
    stream=True,
    extra_body={"cite": True, "sovereign_only": True},
)

for chunk in stream:
    print(chunk.choices[0].delta.content or "", end="", flush=True)
// npm install openai · standard OpenAI SDK works as-is
import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.dcp.sa/v1",
  apiKey:  process.env.DCP_KEY,
});

const stream = await client.chat.completions.create({
  model: "qwen3-4b",
  messages: [{
    role:    "user",
    content: "ما حكم زكاة الراتب الشهري؟",
  }],
  stream: true,
  // DCP extensions
  cite: true,
  sovereign_only: true,
});

for await (const chunk of stream) {
  process.stdout.write(chunk.choices[0]?.delta?.content ?? "");
}
api.dcp.sa / v1 · OpenAI-compatible · Arabic-first defaultFull reference →

Your agent can rent a whole GPU from DCP.

DCP ships an official Model Context Protocol server. Wire it into any MCP-capable agent — Claude, Cursor, or your own — with one line: npx -y github:dhnpmp-tech/dcp-mcp. The agent can then run sovereign in-Kingdom inference, rent a whole GPU, and keep persistent storage — all prepaid in Riyal, through native tool calls.

Inference · OpenAI-compatRent a GPUPersistent storagePrepaid · SAR
01chat · list_models

Sovereign, in-Kingdom inference. The agent lists what is serving this minute and runs OpenAI-style chat completions — Arabic-first, billed per token.

02create_pod · get_pod · extend_pod · stop_pod

A whole GPU on demand — root, Jupyter, and SSH in about a minute. Prepaid per minute; unused time is refunded when the agent stops the pod.

03rent_volume · get_volume

Exclusive, in-Kingdom persistent storage. With an active volume the pod's /workspace survives across pods and providers — the agent's work is never lost.

04get_balance · prepaid in SAR

One renter wallet funds everything. The agent checks its Riyal balance before it spends, so it never overcommits — inference, pods, and volumes all settle against the same balance.

Further still: an agent can mint its own key — POST /api/renters/agent-register returns a real key plus a 20 SAR trial, no human and no email. Money routes accept an Idempotency-Key and return a machine-readable 402 when the wallet is short.

npx -y github:dhnpmp-tech/dcp-mcp
// .mcp.json · claude_desktop_config.json · Cursor
// One line of npx — no install, no SDK to learn.

{
  "mcpServers": {
    "dcp": {
      "command": "npx",
      "args": ["-y", "github:dhnpmp-tech/dcp-mcp"],
      "env": { "DCP_API_KEY": "dcp-renter-…" }
    }
  }
}

# Then just ask the agent, in plain language:
» "Rent a GPU on DCP for 30 minutes and open Jupyter."
» "ما رصيدي بالريال على DCP؟"

// 11 native tools: register_agent · list_models · chat
//   get_balance · list_gpus · create_pod · get_pod
//   extend_pod · stop_pod · rent_volume · get_volume
npx -y github:dhnpmp-tech/dcp-mcp · 11 tools · prepaid in SARAgent playbook →
Saudi Data & AI Authority

ALLaM-7B-Q4

Arabic generation · 32k ctx

Tuned for Modern Standard Arabic with Saudi domain context — Sharia, ZATCA, GoSi, business law, Tadawul filings.

🇸🇦 KSA
Pricing on servingregistered · not yet online
DeepSeek · cross-border

DeepSeek V4 Flash

Frontier · fast · 128k ctx

Frontier reasoning for hard cases. Off by default — opt-in per tenant, cross-border marker on every call, separate invoice line.

🌐 Cross-border · opt-in
SAR 1.10/M in · 3.40/M outdefault off
DeepSeek · cross-border

DeepSeek V4 Pro

Frontier · max · 200k ctx

Top-tier reasoning, max context. For long-doc analysis, complex agent planning, advanced code generation. Cross-border, audit-tagged.

🌐 Cross-border · opt-in
SAR 4.20/M in · 12.60/M outdefault off
On the roadmap

Arabic-first sovereign model · in evaluation

Arabic generation · 64k ctx

A bigger Arabic-first sovereign model is in evaluation. We’ll bring it up alongside ALLaM as a router option once it’s online.

🇸🇦 KSA · pilot
Pricing TBDjoin waitlist

The Arabic AI agent built for Saudi business.

DCP-Agent is live at agents.dcp.sa for Saudi small & mid-size businesses. The agent works in Arabic end-to-end — reply to customers, draft documents, run back-office tasks. Free personal version for every Saudi is coming.

Open agents.dcp.sa →Personal AI · waitlist soon
Live in production
agents.dcp.sa
Visit →

Generic models speak Arabic like a tourist.

ALLaM-7B was tuned on Modern Standard Arabic with Saudi domain context. The verifier hook is locked against an Arabic-negation regression test that runs on every commit. The combination is what banks and regulators actually need from Arabic AI.

Prompt · MSA
ما حكم زكاة الراتب الشهري إذا لم يبلغ النصاب إلا بعد جمعه لسنة كاملة؟
Grounded · cited
يجمع الراتب الشهري ويُحسب الحول من يوم بلوغ النصاب لا من يوم استلام أول راتب. إذا بلغ مجموع المدّخر النصاب وحال عليه الحول، وجبت فيه الزكاة بنسبة ٢٫٥٪ [١]. إن لم يبلغ النصاب فلا زكاة حتى بلوغه [٢].
🇸🇦 Generation · KSA🇸🇦 Verified · 2 citations
A · I want to use DCP

Build with Arabic AI.

For founders, banks, hospitals, regulators, agencies. You ship the product; we serve the inference and the agents. SAR billing, halala-grained, no rental contracts.

  • 01Sign up free · no card, no minimum. Tenant created in < 30s.
  • 02Grab an API key in console.dcp.sa. Drop it into your OpenAI SDK — only base URL changes.
  • 03Try sovereign-only if you’re regulated — frontier stays off, free.
  • 04Spin up agents from the Playground or via the agent SDK. 5 personal + 3 project on Team tier.
SAR + USDC accepted · mada · Apple Pay · bank transfer
B · I have idle GPUs

Earn SAR on your hardware.

For studios, labs, universities, family offices, anyone with consumer or workstation GPUs sitting idle. We handle orchestration, customers, and the SLA. Provider earnings use the published 85/15 platform split, paid monthly to a Saudi bank account.

  • 01Submit your rig — hardware profile, RAM, region, uplink. Approval inside 48h.
  • 02Install the agent · one bash line on Linux, one MSI on Windows. Auto-quarantines on tamper.
  • 03Start serving inference — workload is mixed across customers; we won’t saturate you past 90% by default.
  • 04Get paid in SAR · monthly payout to your bank, with a per-rig dashboard showing utilisation, errors, and earnings.
85% provider share · monthly SAR payout · KSA bank account required
ALLaM-7B
Pricing on serving
Arabic generation · sovereign · registered, not yet online
DeepSeek V4
SAR 1.10/M in
SAR 3.40/M out · flash
Frontier · cross-border · opt-in
Model classTypical modelsPAYG · halala / 1MBest fit
Embeddingbge-m35Retrieval · RAG indexing
TinyTinyLlama 1B · Gemma-2B15Classification · extraction
Smallqwen3:8b · Mistral-7B · ALLaM-7B30Production chat · summaries
MediumQwen 3.6-27B · Coder-32B150Coding · long-context reasoning
Large70B class400High-end reasoning, when online

Inference runs on the OpenAI-compatible API and is billed per million tokens in Saudi Riyal, cost-plus by model class. Pay-as-you-go from a prepaid balance, or take a monthly tier (Starter / Growth / Scale) for a uniform 15–30% discount on every class. New renter accounts start with 100 SAR of credit — no card required — and the API returns a machine-readable 402 insufficient_balance before any unpaid work starts, so there is never a silent negative balance.

GPUVRAMfrom SAR / hour≈ USD / hour
NVIDIA H200141 GBfrom SAR 23.05$6.15
NVIDIA H10080 GBfrom SAR 17.27$4.61
NVIDIA A10080 GBfrom SAR 7.30$1.95
NVIDIA L40S48 GBfrom SAR 5.20$1.39
NVIDIA RTX 509032 GBfrom SAR 5.20$1.39
NVIDIA RTX 409024 GBfrom SAR 3.62$0.97
NVIDIA RTX 309024 GBSAR 0.50$0.13

Every pod is a whole, dedicated NVIDIA GPU with root, Jupyter over TLS and SSH in about a minute — served from verified, in-Kingdom Saudi hardware (PDPL-compliant). You can run up to two pods at once, and each on-demand pod scales onto its own dedicated GPU — no sharing across your pods. On-demand types (H200, H100, A100, L40S, RTX 5090, RTX 4090) are priced cost-plus from the live market, so each is a 'from' floor that floats with the upstream rate; the native RTX 3090 is a Saudi-owned community card and Apple Silicon (M2-class) is live for inference-only workloads. USD figures are indicative conversions at the SAMA peg (1 USD ≈ 3.75 SAR); billing is in Saudi Riyal.

Subscription tiers · Starter SAR 375/mo · Growth SAR 1,500/mo · Scale SAR 5,625/mo · Enterprise on request.Start with 100 SAR credit →
How do I rent an H100 (or other GPU) on demand on DCP?

Sign up for a DCP renter account at dcp.sa, fund your wallet in Saudi Riyal, then launch a pod from the console or via the API — POST https://api.dcp.sa/api/pods with a Bearer renter key. You get a whole NVIDIA GPU (H200, H100, A100, L40S, RTX 5090 or RTX 4090) with root, Jupyter over TLS and SSH in about a minute. Billing is prepaid per GPU-second in SAR, with a prorated refund when you stop early.

Is DCP an OpenAI-compatible inference API?

Yes. DCP exposes an OpenAI-compatible API at https://api.dcp.sa/v1 (POST /v1/chat/completions, GET /v1/models). Point any OpenAI SDK at it by setting base_url to https://api.dcp.sa/v1 and using your DCP renter key as the Bearer token — no code rewrite needed. Inference is billed per token in Saudi Riyal.

Can an AI agent rent a GPU on DCP via MCP?

Yes. DCP ships an official Model Context Protocol (MCP) server. An MCP-capable agent (such as Claude) can list models, run inference, list available GPU types, create and extend GPU pods, rent storage volumes, and check wallet balance through tool calls. See dcp.sa/docs for the MCP setup and tool reference.

What is sovereign / in-Kingdom AI compute in Saudi Arabia?

Sovereign AI compute means your data, the models, the storage, and the control plane all stay inside Saudi Arabia, under Saudi law. DCP runs on Saudi-owned hardware in the Kingdom with full PDPL data-residency compliance, so prompts and answers never leave the country unless a tenant explicitly opts in to cross-border frontier models.

How much does it cost to rent a GPU on DCP?

GPU rental is billed prepaid per GPU-second in Saudi Riyal, cost-plus from the live market. On-demand types and indicative hourly rates: NVIDIA RTX 4090 from about 3.62 SAR/hr, RTX 5090 from 5.2 SAR/hr, L40S from 5.2 SAR/hr, A100 (80 GB) from 7.3 SAR/hr, H100 (80 GB) from 17.27 SAR/hr, and H200 (141 GB) from 23.05 SAR/hr. The native in-Kingdom RTX 3090 is 0.5 SAR/hr. New renter accounts start with 100 SAR of credit and no card is required to begin.

Where does my data live when I use DCP?

Inside Saudi Arabia. Inference, GPU pods, agents, and persistent storage volumes all run on in-Kingdom, Saudi-owned hardware under PDPL data-residency rules. Cross-border frontier models are available only by explicit per-tenant opt-in.

PDPLAlignedSaudi residency
SettlementIn-KingdomHalala · SAR
HostingSelf-hostedIn-Kingdom infrastructure
ZATCAVAT-registered311102233400003
CR7053667775DC Power Solutions Co.
§ Enterprise

Run it in your own VPC.

For banks, hospitals, regulators, the bigger gov programmes. Bring your own keys, your own corpora, your own audit pipeline. We sign the DPA, the MSA, the data-flow appendix, and we sit on the call with your CISO.

  • In-Kingdom settlement · SAR + halala
  • Dedicated tenancy · isolated control plane
  • Private peering · IPsec or DirectConnect
  • Dedicated CSM · onboarding sprint
  • Customer data-classification workbook
  • SLA per enterprise contract · credits + escalation path
§ Ready when you are

Sovereign Arabic AI. Run it.

Eight minutes from this page to a ready renter workspace. First inference is enabled by the catalog only when a verified serving model is online. No procurement. No data-egress conversation. No flat GPU rental.