Réseaux de calcul décentralisé et DePIN : places de marché GPU pour la pile souveraine

Réponse rapide

Vous n'avez pas besoin de posséder le silicium pour faire tourner de l'IA. Une vague de réseaux DePIN (infrastructure physique décentralisée) permet désormais de louer du GPU et du calcul auprès de places de marché sans permission, incitées par des jetons, plutôt qu'auprès des trois hyperscalers — la même logique « une couche de plus décentralisée » que Bitcoin a appliquée à la monnaie, appliquée au calcul. Cette référence en cartographie 12 (dont 7 entièrement open source) : ce que chacun fournit réellement, son jeton et sa chaîne, comment on le consomme en pratique, son degré d'ouverture et son modèle de décentralisation. C'est le complément « louer décentralisé » à notre base de données GPU (posséder le silicium) et à nos runtimes auto-hébergés — trois réponses honnêtes à « où tourne mon modèle ? »

Utilisez-le pour choisir une couche de calcul décentralisée, pas pour choisir un jeton — nous ne stockons AUCUNE donnée de prix, de capitalisation ou de ROI, et les lignes d'économies sont les allégations marketing des projets eux-mêmes, signalées comme telles. Lisez les catégories honnêtement : Akash/io.net/Nosana/Spheron/Flux/Golem sont des places de marché de GPU/calcul à la tâche ; Bittensor est un marché d'incitation à l'intelligence (on achète/vend de la sortie de modèle, pas des heures-GPU) ; Gensyn est un protocole d'ENTRAÎNEMENT décentralisé arrivé en mainnet en 2026, tandis que Prime Intellect est encore émergent (pas encore de jeton officiel). Les tickers, chaînes et statuts de mainnet évoluent vite — vérifiez à la source de chaque projet avant de vous engager. Rien ici n'est un conseil financier ni un aval de jeton. CSV/JSON gratuits sous CC BY 4.0.

Télécharger le CSV Télécharger le JSON API REST →

Use case

Status

Réseau	Ce que vous obtenez et comment	Jeton / chaîne	Open source	Modèle de décentralisation
Akash Network Mainnet Decentralized cloud (GPU + CPU) marketplace GPU rentalAI inferenceGeneral compute	On-demand GPU and CPU containers; deploy any Docker/OCI image via reverse-auction bidding (you set max price, providers bid). Pre-built templates for Llama 3, DeepSeek, Stable Diffusion. How: Define a deployment SDL, set resources + max price, providers bid; pay deployment fees in AKT/USDC from a Keplr/Leap wallet. CLI or Akash Console web UI.	AKT Akash's own Cosmos-SDK / Tendermint app-chain (also accepts USDC via IBC)	Open source	Permissionless provider marketplace secured by a PoS L1; AKT staked by validators/delegators; reverse-auction price discovery. Positions ~60-85% cheaper than AWS/GCP/Azure (project claim — verify). Among the longest-running compute DePINs; Blackwell GPU integration announced.
Render Network Mainnet Decentralized GPU rendering + compute marketplace RenderingAI inference	GPU-accelerated 3D rendering (OctaneRender, Redshift, Blender Cycles) and increasingly generative-AI / compute jobs. How: Submit render/compute jobs through Render client/partner integrations; node operators pick up work; pay in RENDER.	RENDER (renamed/upgraded from legacy RNDR) Solana (migrated from Ethereum in late 2023; legacy ERC-20 history)	Partial	P2P marketplace ('Airbnb of GPU compute') with Burn-and-Mint Equilibrium (BME) tokenomics — job payments burned, operator rewards minted, tying supply to usage. Originated from OTOY; one of the earliest GPU DePINs. Integrations with Runway, Luma, Stability, Black Forest Labs.
io.net Mainnet Decentralized GPU cloud / DePIN (AI clusters) GPU rentalAI inferenceAI training	AI-ready GPU clusters (up to ~10,000 GPUs) aggregated from independent suppliers; inference + training. How: Open IO Cloud, specify GPU type/count/region/duration; protocol assembles a cluster. Submit Ray jobs as if on a private cluster. Pay in USDC/fiat/crypto (converted to IO); suppliers paid in IO. Run IO Worker to supply a GPU.	IO (SPL token on Solana) Solana (settlement); orchestration via the Ray distributed-computing framework	Partial	DePIN aggregating underutilized accelerators worldwide; Ray-native orchestration abstracts multi-owner geography; supplier rewards in IO. Claims ~50-70% below AWS pricing (project claim — verify). Ray (UC Berkeley RISELab) is the orchestration core.
Bittensor Mainnet Decentralized machine-intelligence network (incentivized subnets) Intelligence marketAI inference	A market for machine intelligence — subnets offer inference, training, data, analytics, serverless AI compute, etc. (not a raw GPU-rental marketplace; you buy/sell intelligence/services). How: Consume services via a subnet's API/apps; miners contribute models/compute, validators score quality (Yuma consensus / proof-of-intelligence). TAO emissions split miners/validators.	TAO (+ per-subnet dynamic TAO / alpha tokens) Subtensor — Bittensor's own Substrate-based L1	Open source	Specialized incentivized subnets; 'dynamic TAO' routes emissions by subnet demand; quality measured by ML output, not raw FLOPs. Flagship 'decentralized AI' token; distinct model — incentivizes intelligence production rather than renting GPUs by the hour. Frame carefully: overlaps but is NOT a like-for-like GPU marketplace.
Golem Network Mainnet General-purpose decentralized compute marketplace General computeGPU rentalRendering	CPU and GPU compute for arbitrary tasks (rendering, batch compute, GPU rental); requestor describes task, provider executes. How: Run/connect to the Yagna daemon; requestors post tasks (CPU/RAM/GPU spec, deadline), providers bid, execute, return results, paid on-chain pay-as-you-go.	GLM (ERC-20; fixed 1,000,000,000 supply, no inflation) Ethereum L1 + L2 (Polygon, Arbitrum) for cheaper settlement	Open source	P2P marketplace with optional Byzantine-fault-tolerant redundancy (run a job on N providers, trust the majority) to catch faulty/dishonest nodes. One of the oldest (2016-era) decentralized-compute projects; multi-chain settlement; fixed-supply token.
Spheron Network Mainnet Decentralized compute (GPU + CPU) marketplace GPU rentalAI inferenceGeneral compute	On-demand GPU/CPU — 'Secure GPUs' (data-center providers) and 'Community GPUs' (consumer hardware via Fizz Nodes); LLM/AI workload deployment. How: Buy SPON credits, deploy containerized workloads to Provider Nodes (enterprise) or Fizz Nodes (community); smart contracts formalize leases and settle payment on Base.	SPON Base (Ethereum L2); launched 25 Jul 2025	Open source	Two-tier supply: Provider Nodes (data-center grade) + lightweight Docker-based Fizz Nodes on personal machines; rewards proportional to utilized resources. Fizz-node model directly enables home/SMB hardware participation. Newer token (TGE Jul 2025) — higher volatility (FLAG). ~25,000+ Fizz Nodes reported early 2026 (verify).
Gensyn Mainnet Decentralized ML training/compute protocol AI trainingGeneral compute	Coordination + verification for distributed ML training across untrusted hardware (off-chain execution, on-chain coordination/verification via a Reproducible Execution Environment). How: Submit training/verification workloads; providers execute off-chain, checkpoints/rewards settle on-chain. First app 'Delphi' (AI-settled prediction markets) live on mainnet since Apr 2026.	AI ($AI; TGE 29 Apr 2026, 10B max supply) A custom Ethereum rollup dedicated to ML	Partial	Off-chain execution + on-chain coordination; persistent on-chain identity for reputation; cryptographic verification of untrusted computation. Compute mainnet and the Delphi app (AI-settled prediction markets) went live 22 Apr 2026; token TGE 29 Apr 2026 — reached production mainnet from a Mar-2025 public testnet. Backed by the Protocol Labs ecosystem.
Nosana Mainnet Decentralized GPU compute marketplace (AI inference focus) GPU rentalAI inferenceGeneral compute	On-demand GPU compute optimized for AI inference (and training/rendering); containerized job submission. How: Package workloads into standardized containers, submit to the marketplace; Solana smart contracts match supply/demand and pay providers in NOS on verified completion. GPU owners register rigs to supply.	NOS (SPL token on Solana) Solana (smart-contract matching + payment settlement)	Open source	P2P grid of geographically dispersed GPUs for low-latency inference; NOS staking by providers; jobs verified before payout. Mainnet launched 14 Jan 2025; ~985k jobs processed in 2024; inference-cost positioning ~2.5x cheaper (project claim — verify). Solana-native, fits the 'consumer GPU contributes' narrative.
Aethir Mainnet Enterprise decentralized GPU cloud (DePIN) GPU rentalAI inferenceCloud gaming	Enterprise-grade GPU containers (NVIDIA H100/H200/GB200) for AI training/inference and cloud gaming; plus Aethir Edge home devices. How: Book compute with ATH; GPU Cloud Hosts supply hardware; Checker Nodes verify service quality and earn ATH; enterprise contracts for sustained workloads.	ATH Arbitrum (Ethereum L2) — FLAG: some aggregators list 'Ethereum' loosely; ATH contract is on Arbitrum (verify at build).	Commercial	Pooled GPU supply (data centers + individual Edge operators) matched to AI/gaming demand; 90k+ Checker Nodes police quality; revenue-split with providers. Largest-scale enterprise compute DePIN positioning (440k+ containers, 94 countries, ~$166M ARR Q3 2025 per project — verify figures). Enterprise contracts distinguish it from pure spot marketplaces.
Theta EdgeCloud Mainnet Hybrid decentralized edge + cloud GPU marketplace GPU rentalAI inferenceAI training	Containerized GPU compute (AI inference/training, video transcoding) blending cloud GPUs (A100/V100) with community edge nodes. How: Deploy containerized workloads via EdgeCloud; orchestration routes between cloud GPUs and 30,000+ edge nodes; node operators paid in TFUEL (monthly payouts).	THETA (governance/stake) + TFUEL (gas/payments) Theta blockchain (with Pulse subchain for EdgeCloud transparency)	Partial	Hybrid: enterprise cloud GPUs + community edge nodes under an orchestration layer; dual-token (THETA stake / TFUEL pay). Evolved from decentralized video streaming/CDN into AI GPU compute; large existing edge-node base (10k+ contributors). Heritage = video, now pivoting to AI.
Flux Mainnet Decentralized cloud infrastructure (general compute) + GPU/AI marketplace General computeApp hostingGPU rental	CPU/RAM/storage app hosting on FluxNodes (Docker via FluxOS); GPU/AI/ML/rendering via FluxEdge marketplace. How: Deploy Docker apps to FluxNodes through FluxOS; use FluxEdge for compute-heavy GPU jobs; node operators lock FLUX collateral to provide capacity.	FLUX Flux's own PoW blockchain (FluxOS orchestration layer; collateral-locked nodes)	Open source	Permissionless collateral-staked node network (~10,000 nodes, 66+ countries, 560+ operators) eliminating single points of failure; tiered node collateral. One of the most decentralized by independent-operator count; long-running PoW + cloud hybrid. Strong fit for the sovereignty narrative (run your own node, no AWS).
Prime Intellect Emerging Decentralized AI training protocol (peer-to-peer) AI trainingGPU rental	Globally distributed model TRAINING across heterogeneous, permissionless compute; framework + orchestration (workers, validators, ElasticDeviceMesh). How: Contribute heterogeneous compute permissionlessly to training runs, or use the platform ('Lab') to train your own models; orchestration coordinates workers/validators across the internet.	None (no official token as of mid-2026; any 'PRIME' listings on DEXs are unofficial) Protocol coordination layer (on-chain identity for compute contribution; verify specifics).	Open source	Fault-tolerant globally distributed training (ElasticDeviceMesh manages dynamic process groups over the internet); anyone can join a run. Proof points: INTELLECT-1 (10B, first decentralized LLM training run at scale) and INTELLECT-2 (32B globally distributed RL). Backed by OpenAI founding members / Founders Fund. No official token as of mid-2026; runs a live provider-aggregating compute marketplace while the decentralized protocol remains pre-launch (Base Sepolia testnet).

Sources : la documentation/les dépôts officiels de chaque projet (URL de source complète par ligne dans le CSV/JSON). Aucune donnée de prix stockée. Le complément « louer décentralisé » à la base de données matérielle AI-GPU (posséder le silicium) et aux runtimes IA locaux (auto-hébergement). Fait partie des piles souveraineté numérique et IA souveraine.

Produits, réparations et guides connexes

Dernière révision: 27 juillet 2026.