Reserve
Exact maximum credits are reserved after provider tokenization.

OpenAI-compatible inference routed across vetted idle GPUs. Transparent pricing, signed usage and prepaid credit settlement—without opening a single inbound port. Inference, non-refundable USDG/$CGPU funding, and hourly GPU-capacity rewards are live on Solana & pump.fun mainnet.
Keep the SDK you already use. Crow validates the request, reserves spend, ranks eligible offers, and fails over before the first token when a provider drops.
curl https://api.crowgpu.fun/v1/chat/completions \-H "Authorization: Bearer crow_live_••••" \-d '{"model": "crow-qwen3-8b","stream": true,"messages": [{ "role": "user","content": "Route this request." }]}'
Pair a supported NVIDIA host, pass a model-specific benchmark, publish qualified capacity, and keep the agent connected outbound—your router stays closed. Qualified active GPU-hours earn the current versioned $CGPU capacity rate.
View provider setup ↗$ curl -fsSL https://get.crowgpu.fun | sudo sh
✓ signature verified
✓ supported NVIDIA GPU detected
✓ outbound relay connected
✓ runtime isolation available
pairing code CROW-7Q4X-9N2M
expires in 09:42Requests and signed receipts stay in the double-entry credit ledger. Buyer funding and hourly $CGPU capacity-reward batches are the only value movements on Solana & pump.fun.
Exact maximum credits are reserved after provider tokenization.
Provider signs token counts, model, quote and terminal state.
Gateway verifies receipt and releases unused reservation.
USDG swaps through WETH into $CGPU and direct $CGPU enters the same 30% treasury.
Phase 1 isolates runtimes and does not persist prompts by default. But ordinary GPUs are not trusted execution environments. Do not route secrets, regulated data or confidential strategies until approved attested nodes are available.
Read the privacy model →