# DCP — Datacenter Compute Platform

> DCP is the Kingdom of Saudi Arabia's sovereign AI compute platform. It exposes an OpenAI-compatible inference API, on-demand whole-GPU rental ("pods"), persistent storage volumes, and an autonomous agent layer — all served from Saudi-owned hardware inside the Kingdom (full PDPL / data-residency compliance). DCP is built to be used by AI agents and software, not only humans: the inference endpoint is a drop-in replacement for `api.openai.com`, and an official MCP server lets any agent rent GPUs, run inference, and manage workloads through tool calls.

## For agents: start here

- **Agent guide (start here):** [/agents](https://dcp.sa/agents) — the zero-human flow, the MCP catalog + config, the GPU price table, and the machine-readable money signals, written for both humans and agents.
- **Zero-human onboarding:** `POST https://api.dcp.sa/api/renters/agent-register` (no auth needed; optional JSON body `{email,label,use_case}`) → `201 { "api_key": "dcp-renter-…", "trial_credit_sar": 20, "balance_sar": 20 }`. A real, immediately-usable key + a 20 SAR trial credit, no human and no email click. (The full 100 SAR trial is the human email-verified signup.)
- **Inference (OpenAI-compatible):** Base URL `https://api.dcp.sa/v1`. Use any OpenAI SDK by setting `base_url` to it and `Authorization: Bearer <renter_api_key>` (also accepts `x-renter-key`). Endpoints: `POST /v1/chat/completions`, `GET /v1/models` (each entry has an `available` flag). Both `dcp-renter-` and `dc1-sk-` key prefixes work.
- **MCP server:** [dcp-mcp](https://github.com/dhnpmp-tech/dcp-mcp) — install with `npx -y github:dhnpmp-tech/dcp-mcp` (the npm package `@dcp/mcp` is coming soon). Eleven tools: `register_agent`, `list_models`, `chat`, `get_balance`, `list_gpus`, `create_pod`, `get_pod`, `extend_pod`, `stop_pod`, `rent_volume`, `get_volume`. `register_agent` is unauthenticated (it mints the first key); the rest take a renter key. The native way for an MCP-capable agent (Claude, etc.) to use DCP.
- **Safe money:** money routes (`POST /api/pods`, `POST /api/pods/{id}/extend`, `POST /api/volumes/rent`) accept an `Idempotency-Key` header (a retry with the same key returns the SAME pod/charge), and return a machine-readable `HTTP 402 { code:"insufficient_balance", required_sar, balance_sar, currency:"SAR", topup_url, retryable:true }` (creating no pod) when the wallet is short.
- **Agent manifest:** [/.well-known/ai-plugin.json](https://dcp.sa/.well-known/ai-plugin.json)
- **OpenAPI spec:** [/docs/openapi.yaml](https://dcp.sa/docs/openapi.yaml)

## Inference

- Models available right now: query `GET https://api.dcp.sa/v1/models` (returns OpenAI-style `data[]` with an `available` flag — only `available: true` models are currently serveable). Catalog is Arabic-and-long-tail focused (Qwen family, ALLaM-class Arabic models).
- Auth: a renter API key (`Bearer` token). Get one at https://dcp.sa (renter signup) and fund the wallet (SAR, prepaid).
- Pricing: per-token, cost-plus, billed in SAR. See `GET https://api.dcp.sa/api/pricing`.

## GPU compute (pods)

- Rent a whole GPU with root + Jupyter + SSH. `POST https://api.dcp.sa/api/pods` `{ "duration_minutes": N }` (Bearer renter key). Returns a pod id; poll `GET /api/pods/{id}` for `access_url` (Jupyter), `ssh_command`, `ends_at`, `seconds_remaining`.
- Extend without restart: `POST /api/pods/{id}/extend` `{ "extend_minutes": N }`.
- Stop early (prorated refund): `DELETE /api/pods/{id}`.
- Billing is prepaid per GPU-second in SAR; unused time is refunded on stop.

### GPU types + indicative price (per hour)

Whole, dedicated NVIDIA GPUs, in-Kingdom (PDPL-compliant). On-demand types are priced cost-plus from the live market, so each is a "from" floor (it floats with the upstream rate, refreshed about every 4 minutes); the native RTX 3090 is a Saudi-owned community card at its configured rate. USD ≈ at the SAMA peg (1 USD ≈ 3.75 SAR); billing is in SAR. Live availability per type: `GET https://api.dcp.sa/api/renters/available-providers`.

| GPU | VRAM | Type | from SAR/hr | ≈ USD/hr |
|-----|------|------|-------------|----------|
| NVIDIA H200 | 141 GB | on-demand | 23.05 | 6.15 |
| NVIDIA H100 | 80 GB | on-demand | 17.27 | 4.61 |
| NVIDIA A100 | 80 GB | on-demand | 7.30 | 1.95 |
| NVIDIA L40S | 48 GB | on-demand | 5.20 | 1.39 |
| NVIDIA RTX 5090 | 32 GB | on-demand | 5.20 | 1.39 |
| NVIDIA RTX 4090 | 24 GB | on-demand | 3.62 | 0.97 |
| NVIDIA RTX 3090 | 24 GB | native | 0.50 | 0.13 |

Apple Silicon (M2-class) is also live in-Kingdom for inference-only workloads.

## Persistent storage (volumes)

- Rent an exclusive, in-Kingdom persistent volume: `POST https://api.dcp.sa/api/volumes/rent` `{ "size_gb": 10|20|30 }`. A pod's `/workspace` then persists across pods and providers.
- `GET /api/volumes/me` (status + usage + pool), `DELETE /api/volumes` (release).

## Wallet & account

- Balance: `GET https://api.dcp.sa/api/renters/balance` (Bearer). Top up in SAR via the wallet at https://dcp.sa.

## What DCP is (positioning)

- Sovereign, in-Kingdom AI compute: data, models, storage, and control plane inside Saudi Arabia, under Saudi law — a structural advantage no foreign API can match.
- A two-sided marketplace of Saudi-owned GPUs (RTX-class) and Apple-Silicon Macs, unified by a self-healing mesh and a smart routing layer that presents one clean OpenAI-compatible endpoint.
- Not a token pipe: you can run inference, fine-tune and host your own models, rent whole GPUs, attach persistent storage, and deploy agents — all in one place.

## Use cases (what people ask for, and what DCP answers)

- "Rent an H200 / H100 / A100 / L40S / RTX 5090 / RTX 4090 on demand" → whole-GPU pods, per-second billing in SAR, root + Jupyter + SSH in ~1 minute. See GPU table above.
- "OpenAI-compatible inference API alternative" → drop-in `base_url = https://api.dcp.sa/v1` with an OpenAI SDK and a DCP renter key. No rewrite.
- "MCP server to rent a GPU / run inference" → the official DCP MCP server (`npx -y github:dhnpmp-tech/dcp-mcp`) exposes self-registration (`register_agent`), inference, GPU-type listing (`list_gpus`), pods, volumes, and balance as tools for MCP-capable agents (Claude, etc.).
- "AI agent that buys its own compute with no human" → `register_agent` (or `POST /api/renters/agent-register`) mints a real key + 20 SAR trial with no human; idempotent money routes + the HTTP 402 funding signal let the agent rent and pay safely. See https://dcp.sa/agents.
- "Sovereign / in-Kingdom AI compute in Saudi Arabia / the Gulf" → data, models, storage, and control plane all inside KSA under PDPL; cross-border frontier models only by explicit per-tenant opt-in.
- "Fine-tune & host your own model" → rent a whole GPU, attach a persistent in-Kingdom volume, train/fine-tune, serve.

## How DCP compares (truthful framing)

- vs US-based GPU clouds (RunPod / Vast / Lambda): DCP keeps data, hardware, and settlement inside Saudi Arabia under Saudi law (PDPL), billed in SAR — a data-residency guarantee a foreign API cannot match for KSA/Gulf workloads.
- vs hosted inference APIs (Together / DeepInfra / Modal / Replicate): DCP's `/v1` is OpenAI-compatible (drop-in base_url) and runs in-Kingdom; it also offers whole-GPU pods and an agent-native MCP server, not only a token pipe.
- Honest catalog: only models with `available: true` from `GET /v1/models` are serveable right now; capacity is published only after live provider verification, never simulated.

## Docs

- Agent guide (zero-human onboarding, MCP, money signals): https://dcp.sa/agents
- Human docs: https://dcp.sa/docs
- Quickstart: https://dcp.sa/quickstart
- Pricing (GPU/hr + per-token, in SAR): https://dcp.sa/pricing
- This file: https://dcp.sa/llms.txt