GPN — Technical Overview

A short, site-friendly technical summary. See GPN-WHITEPAPER.md for the full paper.

The model

A GPN node is one entity — one GPU, one identity/wallet — that is always both a miner and an AI provider. It mines whenever the GPU is free; when a job arrives it pauses mining, frees VRAM, runs the job to completion, then resumes mining. There is no provider-only or mine-only mode, and no profitability switch. (A separate Staking role is planned.)

"One GPU, one job" note: "one job at a time" is mutual exclusion — the GPU never runs two jobs concurrently — and is NOT a restriction on job types. Every node is equipped by design to serve every job type (LLM · image · video · ML inference · fine-tune) and advertises all of them.

The chain is thin: it coordinates jobs, holds escrow, enforces eligibility/capability gates, and settles rewards. Compute runs off-chain.

Consensus

Ethash-derived GPU proof-of-work; anyone can mine; no validator set.
15-second blocks; self-tuning difficulty (millisecond resolution, runaway-proof).
Bitcoin-style rewards: the block producer keeps the block reward (account bound via pre-runtime digest).
Per-node wallets: each miner pays out to its own account.
Fork tolerance: explicit network identity (chain-id + genesis hash + version); divergent chains warned + starved, not banned. No finality gadget → confirmation is probabilistic by depth.

Eligibility (GPU-only)

Proof-of-Staked-GPU attestation (GPU count + measured GFLOPS) → tiers (Bronze/Silver/Gold) → reward multiplier (staking raises your reward, optionally).
Stake-free path: producing a seal-verified block is the proof of GPU work; the reward goes to the producer after a small number of authored blocks.
Capability gates: min_gflops, min_vram_mb, min_ram_mb — the router won't send a job to a device that can't meet it.
Capacity challenge: random compute challenges make throughput proven, not self-reported.
Liveness: always-on, publicly-reachable full nodes earn credit and gate top tiers.

AI marketplace

Job lifecycle: Pending → Accepted → Running → AwaitingVerification → Completed (+ Cancelled/Failed).
Per-token escrow: cap = rate × max tokens; charge = rate × actual; surplus auto-refunds; dollar-stable, GPN-settled.
Cost-derived pricing: a roofline of the provider's true throughput × margin, clamped ≥ cost.
Routing: capability filter → require idle → lowest-latency (least-ms) → dispatch the assignment. Providers register + advertise endpoints on-chain for discovery; models can be pulled on demand.
Off-chain escrow: payment channels lock funds once and settle each job off-chain → no per-job block delay; either side can force-settle on-chain.
Safety: permissionless expiry, per-job failure refunds, a bounded on-chain expiry sweep, and a per-requestor cap.

Tokenomics

Fair launch via mining; halving every ~2 years; 50 GPN/block @ 15 s; hard supply cap ≈ 520.5 M GPN (≈ 420.48 M emitted + 100 M treasury); the 100 M treasury is held by a keyless pallet account (70 M genesis-unlocked, 30 M vested over 10 yr) and spendable only by a 13-of-25 governance roster; 5% fee split (40% staking / 30% treasury / 20% burn / 10% oracle).

Privacy

A shielded pool (Sapling-style commitments + nullifiers): deposit hides a note; spend proves ownership in zero knowledge, publishing only an unlinkable nullifier; supports shielded↔shielded and shielded→transparent with value conservation. Production security relies on a multi-party trusted-setup ceremony (trapdoor destroyed by independent contributors).

Governance & upgrades

Committee-root governance after a timelock (no single key acts alone); immutable genesis; runtime upgrades forklessly via set_code; review-first.

Network

Permissionless, redundant access points; Bitcoin-style peer discovery (persistent peer store + DNS/HTTP seeds + on-chain registry + a compiled-in bootnode fallback).

Status: live on a public testnet (chain, marketplace, channels, shielded spends, discovery, governance). Next: public CRS ceremony, mainnet, bridge.