IA ouverte communautaire — réseau GPU P2P gratuit pour 1 million d'utilisateurs d'ici 2030

Porteur : Fernando

Statut : en mobilisation

Portee : Mondial - Lieu : Online / Worldwide

Cree le 21/06/2026

Probleme identifie

Accessing powerful artificial intelligence is expensive: ChatGPT Plus at $20/month, Claude Pro at $20/month, OpenAI API at $0.01-0.06 per request. These services are proprietary: your data belongs to them, their models are closed, and their terms change whenever they want. Open source models exist (Llama 3, Mistral, Qwen, Gemma) — free, auditable, modifiable — but they require powerful GPUs that 99% of people don't have. A model capable of reasoning (70B parameters) requires 40 GB of VRAM — that's 2 high-end graphics cards ($3,000). Yet in every neighborhood, there are dozens of GPUs sitting idle: gamers who don't play 24/7, designers' workstations, small business servers, repurposed mining rigs. If 6 people share their GPUs on a local network, they can collectively run a model that no one can afford alone. It's the same principle as carpooling: the resource exists, it's just poorly distributed.

Solution proposee

Create a community network of shared GPUs to run open source AI models. Each member contributes their GPU when they're not using it (nighttime, work hours) and in return, gets free access to powerful AI. The model is partitioned across network nodes (like Petals or Exo Labs): each GPU computes a few layers of the model, and together they form a "supercomputer" capable of running Llama 3 70B, Mixtral, or any open source model. The network is regional (same city, same fiber Internet provider) to minimize latency. ChatGPT-style web interface but 100% free and local. Your data NEVER leaves the network — no American servers, no tracking, no censorship. The economic model: GPU contribution = free access. Non-contributor = symbolic community rate ($0.01/request). Zero profit. Community governance: contributors vote on which models to host, what usage rules to apply, and how to allocate resources.

Methode de calcul du seuil

Mobilize GPU owners (gamers, designers, researchers, businesses), ML/infra developers, legal experts (GDPR, privacy regulations), translators (multilingual models), and testers (students, researchers, nonprofits). The network starts with a total VRAM threshold (e.g., 200 GB of shared VRAM) before activation — guaranteeing that at least one 70B model can run continuously. The project is complementary to Nextcloud (model hosting), LoRa (P2P discovery), and ecovillage setups (community GPU clusters).

Actions (16)

Voir le projet sur Wikiaction

Explorer tous les projets