Decentralized AI

Decentralized AI refers to efforts to develop, run and govern artificial intelligence without relying on a few large companies that control the most powerful models, chips and data centers. The term covers several different approaches: openly released models that anyone can run and modify (see Open Weights), training models across computers owned by many participants, running AI on people’s own devices (see On-Device AI), learning from data that stays where it is (see Federated Learning), and blockchain-based networks that reward people for contributing computing power or models.

Why decentralize

Advocates argue that concentrating AI in a handful of firms – and the governments that can pressure them – creates dangerous concentrations of economic and political power (see Hyperscaler and Moat). Decentralization, in this view, protects freedom of thought and speech, prevents single points of failure and censorship, and lets communities build AI that reflects their own languages and values (see Sovereign AI and Low-Resource Language). The argument draws on classical-liberal suspicion of concentrated power and on the open-source software tradition; the Ethereum co-founder Vitalik Buterin’s “d/acc” – defensive and decentralized acceleration – makes a related case for technologies that spread rather than concentrate power.

Approaches

Open-weights models from DeepSeek, Alibaba (Qwen), Mistral, OpenAI (gpt-oss, 2025) and Meta (Llama and, since August 2026, open-weights Muse models, though its strongest models remain closed) have done the most to spread capability, letting thousands of developers fine-tune and run models independently (see LoRA and Model Merging). Distributed training is harder, because training a large model requires fast communication between chips. In 2024 the startup Prime Intellect trained a 10-billion-parameter model, INTELLECT-1, on up to 112 H100 chips in five countries on three continents, using a method from Google DeepMind, and in 2025 it used globally distributed reinforcement learning for a 32-billion-parameter model; its INTELLECT-3, however, was trained on a single centralized cluster. Nous Research developed methods to reduce the communication such training needs and launched its Psyche network in 2025. Crypto-based networks such as Bittensor pay participants in tokens for providing models or computation; skeptics long questioned how much useful AI they produce, but in March 2026 the Bittensor project Templar completed Covenant-72B, a 72-billion-parameter model trained by more than 70 permissionless participants – still, by one estimate from Anthropic’s co-founder Jack Clark, with roughly a thousandth of the computing power of a frontier run.

Challenges

Decentralized efforts face hard limits. Frontier training needs enormous, tightly connected clusters (see AI Supercomputer), so distributed projects lag well behind the leading labs. Coordinating many independent participants raises problems of quality, trust and security, including the risk that contributors poison data or models (see Data Poisoning). And decentralization makes oversight harder: no one can recall an open model once released, enforce safeguards or stop misuse, which is why some safety researchers and governments favor control points such as chips and large training runs (see Compute Governance). Supporters counter that centralized control has its own risks, including capture by states or companies (see Regulatory Capture).

Why it matters

The question of whether AI will be centralized or decentralized is one of the defining political choices of the technology. Decentralization spreads power and innovation but makes coordination on safety harder; centralization makes oversight easier but concentrates control. Many analysts expect realistic futures to combine both – and the balance will shape who benefits from AI and who decides how it is used.

Written with Claude, made by Anthropic, a Nasdaq-listed developer of closed frontier models that does not release model weights, supports controls on advanced chips, and is backed by Google, Amazon, Microsoft and Nvidia; its co-founder Jack Clark is quoted here.

Related

Open Weights · Federated Learning · On-Device AI · Open Source AI · Compute Governance · Sovereign AI · Public AI

Sources: Prime Intellect: INTELLECT-1 Release: The First Globally Trained 10B Parameter Model (November 2024) · Prime Intellect: INTELLECT-2: The First 32B Parameter Model Trained Through Globally Distributed Reinforcement Learning (May 11, 2025) · Implicator: Prime Intellect’s INTELLECT-3 – open-source ambition meets centralized reality (November 2025) · Nous Research: DisTrO (2024) · Yuma Rao: Bittensor: A Peer-to-Peer Intelligence Market · TAO Daily: Templar’s Covenant-72B becomes the largest decentralized LLM training run ever completed (March 2026) · TAO Daily: Anthropic co-founder highlights Covenant AI as largest decentralized training network (2026) · NYU Shanghai RITS: Meta hasn’t given up on open source as Muse Spark launches (April 2026) · Vitalik Buterin: My techno-optimism (November 27, 2023) · Campus Technology: Meta releases Muse Glimmer (August 26, 2026)

Leave a Reply

Your email address will not be published. Required fields are marked *