Saudi Arabia's sovereign GPU compute & AI-inference marketplace.

Run AI inference and rent GPUs by the minute — on verified in-Kingdom GPUs, PDPL-compliant, billed in Saudi Riyal.

Platform · Docs · Provider setup · Renter app · For agents

---

Overview

DCP (by DC Power Solutions, CR 7053667775) is Saudi Arabia's sovereign GPU compute and AI-inference marketplace. Workloads run on verified in-Kingdom GPUs, data stays in the Kingdom by default, and everything is billed in Saudi Riyal (SAR) with PDPL-compliant handling.

Two products are live today:

• AI Inference — "AI by the token." An OpenAI-compatible API (api.dcp.sa/v1) for chat/completions, billed per token in SAR. Sovereign round-trip: an Arabic prompt is served by a Saudi GPU and answered in Arabic, never routed to a foreign cloud (cross-border frontier models are strict opt-in, per tenant).
• GPU Pods by the minute. Launch a dedicated GPU pod (Jupyter + SSH) in minutes, billed per-second in SAR, extend it on the fly, and attach persistent in-Kingdom storage.

Settlement is fiat SAR via Moyasar. Top up by card (3DS), pay as you go, and — as a provider — get paid in real Saudi Riyal. Not crypto.

Built to be operated by humans and autonomous agents: an agent can register itself, get a trial balance, run inference, rent a GPU pod, and manage storage end-to-end with zero humans in the loop.

---

Features

🧠 AI inference

• OpenAI-compatible POST /v1/chat/completions + in-browser playground; per-token SAR billing with prepaid balances.
• Multi-engine provider routing (llama.cpp · Ollama · vLLM) with live health probes, capability verification, single-in-flight gating per provider, and automatic fallback.
• Sovereign round-trip — served by an in-Kingdom GPU by default; cross-border/frontier models are opt-in per tenant.
• ELM (Expert Language Models) — fine-tuned 4–8B domain models that beat frontier APIs on a specific task at a fraction of the cost ("build a Camry for the job, don't drive a Porsche everywhere").
• Arabic-first, with a reproducible Arabic customer-service benchmark (harness + dataset + baselines).

⚡ GPU pods

• Renter-launchable dedicated GPU pods (Jupyter + SSH), billed per-GPU-second in SAR, prepaid with early-stop refunds.
• Extend on the fly — +30m / +1h / +2h, no restart, incremental SAR charge, 24h ceiling.
• Persistent workspaces — rentable in-Kingdom volumes (10 / 20 / 30 GB); /workspace reattaches across pods, with honest ephemeral-vs-persistent signalling and an upsell when ephemeral.
• Live rental countdown + sub-5-minute expiry warning; fast teardown frees the GPU (and restores inference) within ~7s.
• Container templates (PyTorch, …) on Docker + NVIDIA Container Toolkit.

💾 Persistent volumes

• Paid, exclusive, in-Kingdom persistent storage (S3-compatible, mesh-only), per-renter quota, monthly billing in advance with a 7-day suspend-but-keep-data grace on lapse. Cross-provider restore/snapshot on pod launch/teardown.

🤖 Agent-first (built for autonomous agents)

• Model Context Protocol (MCP) server (@dcp/mcp) exposing tools — register_agent, list_models, chat, get_balance, list_gpus, create_pod, get_pod, extend_pod, stop_pod, rent_volume, get_volume — so an MCP-capable agent (Claude, Cursor, custom) runs inference, rents GPUs, and manages storage via native tool calls.
• Self-serve agent onboarding — POST /api/renters/agent-register → API key + 20 SAR trial, zero humans to a running pod.
• Agent discoverability / AEO — llms.txt, .well-known/ai-plugin.json, sitewide schema.org JSON-LD, a dedicated /agents page, and machine-readable HTTP 402 (insufficient_balance → topup_url) before any unpaid work.

💳 Payments & settlement — fiat SAR (live)

• Moyasar card top-up (hosted invoice + 3DS — no card data touches DCP), SAR/halala ledger, idempotent webhook crediting.
• Per-second / per-token metering, prepaid quotes, refunds on early stop or failure, monthly volume billing.
• Provider payouts in SAR via Moyasar pay-out rails.
• On-chain escrow settlement is roadmap, not live — see Roadmap.

🖥️ Provider platform

• Self-serve onboarding (web wizard · CLI · desktop app), a lightweight auto-updating daemon, GPU capability reporting + heartbeats, self-serve pod pricing, and an earnings dashboard.
• Connect from home/residential networks over a self-hosted WireGuard mesh — no public IP or port-forwarding required (works behind CGNAT), which is what makes consumer GPUs viable.

🛠️ Control plane & ops

• Admin command center: pricing, payment reconciliation, incident command, provider approval, fleet/launch-readiness, deployment templates, and health checks.

🔒 Security & compliance

• PDPL-ready, sovereign data residency, HMAC-signed daemon/webhook auth, scoped API keys, audit logging, secret scanning (CI), and hardened container isolation.

---

Architecture

Renter · SDK · IDE · Agent (MCP)
        │
        ▼
   dcp.sa            Next.js web app, dashboards, docs, agent surfaces
        │
        ▼
 api.dcp.sa          Express API
        ├─ SQLite          SAR/halala ledger · jobs · providers · volumes
        ├─ Moyasar         fiat SAR pay-in / pay-out   ← live settlement
        ├─ Routing         multi-engine health, verification, in-flight gating, fallback
        └─ WireGuard mesh  outbound-only gateway for residential providers
        │
        ▼
 Provider daemon + GPU runtime   (verified in-Kingdom GPUs)
        +  partner-backed on-demand capacity for additional GPU types

---

Repository structure

dcp-platform/
├── app/                Next.js app — dashboards, docs, agent surfaces, API proxies
├── backend/            Express API, services, migrations, provider installers, tests
├── components/         Shared React UI
├── contracts/          On-chain settlement workspace — ROADMAP, dormant (see Roadmap)
├── docker-templates/   GPU workload + pod template definitions
├── docs/               Public product, API, compliance, provider, and architecture docs
├── e2e/                Playwright end-to-end tests
├── infra/              Deployment and runtime configuration
├── integrations/       Agent integrations (MCP server)
├── ops/                Operator scripts
├── orchestration/      Health, monitoring, failover, checkpoint modules
├── packages/           Shared packages
├── public/             Static assets, llms.txt, .well-known/ai-plugin.json
├── scripts/            Repo-level smoke, release, maintenance scripts
├── sdk/                Node + Python SDK sources
├── security/           Security test harnesses and policy
└── vscode-extension/   VS Code / Cursor extension source

See REPO_MAP.md for the maintained directory map.

Long-term source-of-truth repos	Purpose
DCP-SA/dcp-agent	Provider machine agent
DCP-SA/dcp-desktop	Desktop provider app
DCP-SA/dcp-contracts	Shared API contracts + generated client types

---

Tech stack

Layer	Technology
Frontend	Next.js 14, React 18, TypeScript, Tailwind CSS
Backend	Express.js, SQLite (better-sqlite3), Zod
Inference	llama.cpp · Ollama · vLLM, OpenAI-compatible API
Pods & jobs	Docker, Dockerode, NVIDIA Container Toolkit
Networking	self-hosted WireGuard mesh (residential-NAT friendly)
Payments	Moyasar, SAR/halala accounting
Agents	Model Context Protocol (MCP) server + AEO discovery
Testing	Jest, Playwright, template/contract validation
Deployment	Vercel (frontend), VPS (backend), GitHub Actions

---

Getting started

Renter (API)

# 1. Get a key (self-serve / agent)
curl -X POST https://api.dcp.sa/api/renters/agent-register
# 2. Run inference (OpenAI-compatible)
curl https://api.dcp.sa/v1/chat/completions \
  -H "Authorization: Bearer <dcp-renter-key>" \
  -d '{"model":"<model>","messages":[{"role":"user","content":"مرحبا"}]}'

Rent a GPU pod

dcp --base-url https://api.dcp.sa pod create --image pytorch --duration 60 --token <key>
dcp pod list | get | extend | stop

Provider — install the daemon / desktop app, join the WireGuard mesh, report your GPU, set your price, earn SAR. See Provider onboarding.

Agent — point an MCP client at @dcp/mcp; tools auto-register.

---

Local development

npm install            # root deps
npm run dev            # web app (Next.js)

cd backend
npm install
node src/server.js     # API

npm --prefix backend run templates:validate   # validate deploy templates
npm run build          # production build

Configuration

Start from the example env files; enable only the integrations you need locally.

cp .env.example .env.local
cp backend/.env.example backend/.env

Variable	Purpose
BACKEND_URL	Backend API origin used by the frontend
NEXT_PUBLIC_DC1_API	Public API URL consumed by client code
MOYASAR_SECRET_KEY	Payment API key for SAR top-ups and payouts
MOYASAR_WEBHOOK_SECRET	Webhook HMAC validation secret

Secrets stay in local env files or the deployment secret store. Never commit real keys, databases, logs, or operator notes.

---

Documentation

• API reference · OpenAPI
• Provider onboarding · Dashboard · Earnings
• Architecture overview · Container jobs
• Compliance · Security · Privacy · Terms
• Changelog

---

Roadmap

• On-chain agent-to-agent settlement rail — Solidity escrow / provider-staking / job-attestation contracts (Base testnet). Not live; kept for when autonomous agents settle payments directly with each other. Today, all settlement is fiat SAR via Moyasar.
• ELM registry + autoscale-by-domain — replicate in-demand domain models across providers; renters pick a model + SLA, not a provider.
• Managed in-Kingdom object storage — swap the mesh MinIO store for a managed KSA S3-compatible backend.

---

License

Proprietary — © DC Power Solutions. All rights reserved.