DarkGPU Documentation

Distributed inference marketplace for NVIDIA GPUs. OpenAI-compatible API, 40-60% cheaper than OpenRouter, operators keep 95%.

Authentication

Sign up at console.darkgpu.ai ($10 free credits). Create an API key on the API Keys page. Use Authorization: Bearer dgpu-....

First request

curl https://api.darkgpu.ai/v1/chat/completions \
  -H "Authorization: Bearer $DGPU_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"qwen3.5-7b","messages":[{"role":"user","content":"Hi"}]}'

Python (OpenAI SDK):

from openai import OpenAI
client = OpenAI(base_url="https://api.darkgpu.ai/v1", api_key="dgpu-...")
r = client.chat.completions.create(
    model="qwen3.5-7b",
    messages=[{"role":"user","content":"Hello"}],
)
print(r.choices[0].message.content)

POST /v1/chat/completions

OpenAI-compatible chat completions. Supports streaming via "stream": true.

Response headers include x-darkgpu-provider (provider ID that served the request) and x-darkgpu-request-id.

GET /v1/models

Returns the enabled model catalog with pricing. See pricing reference.

Pricing reference

Model ID	Input /1M	Output /1M	Min VRAM
qwen3.5-7b	$0.015	$0.06	8 GB
llama-3.3-8b	$0.02	$0.08	8 GB
mistral-nemo-12b	$0.030	$0.12	12 GB
gemma-4-12b	$0.035	$0.14	12 GB
qwen3.5-14b	$0.030	$0.15	16 GB
qwen3.5-27b	$0.060	$0.30	32 GB
llama-3.3-70b	$0.100	$0.50	48 GB
deepseek-v3	$0.270	$1.10	160 GB

Rate limits

Free tier: 60 req/min, 200K tokens/day
Paid: 600 req/min, unlimited tokens
Enterprise: custom — email [email protected]

Errors

Errors follow the OpenAI shape: {"error":{"message":"...","type":"..."}}.

HTTP	type	Meaning
400	invalid_request	Bad input
401	auth	Missing / invalid API key
402	billing	Out of credits
404	not_found	Unknown route or model
429	rate_limit	Too many requests
503	service_unavailable	No provider online for this model

Provider install

Linux with NVIDIA driver 535+, CUDA 12+, Python 3.10+:

curl -fsSL https://darkgpu.ai/install.sh | bash
darkgpu auth
darkgpu serve

The installer detects your GPU, installs vLLM if missing, writes a config to ~/.config/darkgpu/provider.toml, and optionally installs a systemd unit.

Configuration

api_url         = "https://api.darkgpu.ai"
price_floor_pct = 0        # 0 = accept catalog prices
idle_detection  = true     # pause when GPU is in active use

[[schedule]]
days  = ["mon", "tue", "wed", "thu", "fri"]
start = "22:00"
end   = "08:00"

Provider earnings

Three tiers — the calculator on the landing page defaults to conservative. Actual revenue depends on demand for your model and your provider reputation.

Payouts are weekly, every Monday UTC, via Stripe Connect (USD). Minimum payout: $10.

Provider security

The agent is open source Rust. Only outbound WebSocket — no inbound ports, works behind NAT/CGNAT.

Each agent generates an X25519 keypair on first run, stored at ~/.config/darkgpu/x25519.key. The public key is advertised via /v1/providers/:id/attestation.

Trust tiers:

basic — consumer GPU, software-attested
verified — H100/H200 with NVIDIA Confidential Computing (Phase 5)

Economics

Revenue split: 95% provider, 5% platform. We don't pretend 0% is sustainable — 5% funds Stripe fees, Cloudflare infra, and the free-tier subsidy.

Consumer free tier: $10 of credits on signup, no card required.

See economics.md for the full breakdown.

FAQ

See the FAQ on the landing page for common questions: earnings reality, hardware wear, crypto, trust, latency.