A proposal by David Carrero, co-founder of Stackscale, has sparked a key debate in the tech ecosystem. On his X profile (formerly Twitter), he suggested:
“And wouldn’t it be time to create an AI like torrent to have something free, decentralized, shared, and that is an artificial intelligence that doesn’t really depend on anyone?”
The idea, inspired by peer-to-peer (P2P) systems such as BitTorrent, seeks to port that model to artificial intelligence: a shared, sovereign, community-owned AI. This visionary thought quickly met a more technical, tangible reply: the AI:// protocol, an attempt to create a minimal standard for AI to access internet content in an optimized, lightweight way.
Using this as a thread, we examine how that reflection connects with real-world efforts in the field of decentralized AI, including initiatives such as Sahara AI, Nous Research, Bittensor, SingularityNET, AIArena, NodeGoAI, and more — and how protocols like IPFS or AI:// are signaling a clear trend towards a future of more open, democratic intelligence.
1. Decentralized AI: What’s the Proposal?
Distributed Artificial Intelligence (DAI) is a long-standing field that emerged as an alternative to centralized systems. It focuses on solving complex problems by distributing computation and coordination among multiple autonomous agents. It is used in multi-agent systems, parallel problem solving, and large-scale social simulations.
In the current context, the idea has evolved: we now talk about AI that doesn’t live solely in massive data centers, but rather in networks that are collectively managed, with no single owner.
2. Real Projects: Decentralization in Motion
Sahara AI
Founded in 2023, Sahara AI has raised $43 million USD with backing from Samsung NEXT, Pantera, and Binance Labs, among others. Its mission: to create a blockchain platform that rewards users, data sources, and model trainers. The aim is to address the lack of transparency and compensation in centralized generative AI models.
Nous Research
A Web3-based project that raised $50 million USD in 2025, reaching a $1 billion valuation. It is developing the Psyche Network, a blockchain-based (Solana) collaboration network for AI training. Its approach: distributed computing and collaborative models, resistant to censorship and power concentration.
Bittensor
An open-source blockchain-based platform promoting a peer-to-peer marketplace for intelligence. Nodes train models and are rewarded for their contributions, fostering quality and censorship resistance.
SingularityNET
Led by pioneer Ben Goertzel, SingularityNET aims to build an open AI network using Ethereum smart contracts. Its goal: interoperability between intelligent agents and autonomous evolution, with a community and ethics-first approach.
AIArena
An academic proposal that offers a blockchain-based decentralized AI training platform on Base. Participants contribute models and compute capacity, receiving on-chain rewards for their input.
DIN (Decentralized Intelligence Network)
A theoretical framework using federated learning over blockchain, where data stays local to each node, and only model parameters are shared. It also integrates cryptographic rewards and auditable incentives.
NodeGoAI
A decentralized network to monetize idle compute capacity, especially for AI and spatial workloads. It implements a blockchain protocol plus dedicated hardware to create a P2P ecosystem for distributed computation.
3. The Infrastructure: Protocols and Standards
For this vision to work, more than willpower is required: we need protocols that enable technical cooperation and structure distributed AI access.
AI:// Protocol
A direct response to Carrero’s comment. Initiated by Raúl Illana, this protocol defines a lightweight, AI-optimized method for language models (LLMs) to consume web content without distractions. It uses Markdown, TLS security, and minimalist design — a technical foundation for an “AI-friendly web”.
IPFS (InterPlanetary File System)
A robust and already-operational P2P file distribution network with content-based addressing. It can serve as storage infrastructure for data used by decentralized AI, offering redundancy, anonymity, and resilience.
Gemini + MCP
While not inherently decentralized, they are relevant to cooperative agent environments. Gemini is Google’s multimodal LLM, and the Model Context Protocol (MCP) allows orchestrating functions and agents, enabling technical collaboration between AIs.
4. Advantages and Incentives of Decentralization
A torrent-style AI model offers several potential benefits:
- Data sovereignty and control: users keep their own data, contribute voluntarily, and receive recognition or compensation.
- Censorship resistance: if each node holds part of the AI, shutting it down would require targeting multiple sources instead of one central server.
- Inclusion and diversity: AI access becomes more democratic, reaching those unable to afford licenses or centralized infrastructure.
- Transparency and traceability: operating on blockchain allows auditing who contributed which model or dataset and what improvements were made.
5. Challenges Still Ahead
Of course, this technological utopia faces significant hurdles:
- Scalability: training or running large models requires GPU resources that not all participants can provide.
- Security and trust: an open system may be vulnerable to manipulation, biased models, or malicious code.
- Standardization: without technical agreements, there’s a risk of fragmentation into incompatible networks and models.
- Regulation: decentralization can complicate compliance with privacy laws and responsible data use.
6. From Visionary to Feasible
Carrero’s vision pushes a powerful concept, and existing initiatives — Sahara, Bittensor, SingularityNET, Nous Research, AIArena, DIN, NodeGoAI — validate it as more than idealism: it’s an emerging field in active development.
Projects like AI:// and IPFS contribute concrete building blocks: the former improves how AI consumes content, the latter distributes that content. Together, they could form layers to build a torrent-style AI: shared, resilient, and open.
Research on models such as DaiMoN (since 2019) already anticipates many of these functionalities: rewarding improvements to AI models, distributed ledgers, and verification without full data access.
7. Towards a Community-Owned Artificial Intelligence
While Silicon Valley continues to concentrate power, the decentralizing trend pushes toward another model: AI as a commons, managed by the community and oriented toward collective, not corporate, interests.
As Illia Polosukhin —co-author of the transformer architecture and now a strong advocate for “user-owned AI”— has stated, this approach aims to prevent oligopolistic control of AI and promote transparency, ethics, and open access.
FAQ
1. What is decentralized AI?
It’s an approach where data, computation, and models are not hosted in a single center but distributed among independent participants who collaborate collectively and transparently.
2. Which projects already exist?
Sahara AI, Nous Research, Bittensor, SingularityNET, AIArena, DIN, and NodeGoAI are real-world examples implementing the decentralized vision.
3. How does AI:// fit into this ecosystem?
AI:// helps LLMs read web content without the noise and complexity of HTML, ideal for lightweight, distributed AI.
4. Is a torrent-style AI viable?
Technically, it’s in pilot stages. The main challenges are infrastructure, robust protocols, governance, and adaptive security.
