Jake Brukhman, founder of crypto venture firm CoinFund, has issued a stark warning about the increasing centralization of artificial intelligence development, arguing that recent events show how governments can effectively choke access to cutting‑edge models. The trigger was Anthropic’s compliance with US export controls, which forced the AI lab to disable its Fable 5 and Mythos 5 models for all foreign nationals, regardless of location. Brukhman pointed to this as proof that the real bottleneck is no longer code or data but the concentrated physical GPU clusters that train frontier models.
In response, a wave of distributed AI projects is building infrastructure to pool under‑utilized global GPU resources. Brukhman named several teams – Gensyn, Prime Intellect, Bagel, Pluralis Research, Nous Research, Macrocosmos and Covenant – that are actively working on algorithms capable of coordinating decentralized training at scale. CoinFund has already backed Prime Intellect, which raised a seed round in 2024, and continues to invest in the sector.
Pluralis Research is pushing the idea furthest by tokenizing AI model ownership. In this model, weights are split among network participants, creating a fragmented structure where no single entity controls the entire system. Tokenized models could align incentives for researchers, compute providers and users, while making shutdowns or censorship far harder to enforce. However, Brukhman acknowledged that decentralized training still faces gaps in latency, reliability and verification compared to hyperscaler clusters, though several teams are racing to close that gap.
Storage is another critical piece. Even if distributed training succeeds, it will require decentralized data pipelines that cannot be switched off by a single government order. Projects like Filecoin illustrate how existing Web3 storage infrastructure might slot into this stack, even if they are not yet tightly integrated. The broader question, Brukhman and others argue, is whether AI development will remain trapped within a few well‑funded labs subject to state‑level coercion, or whether permissionless networks can replicate what currently demands enormous centralized coordination.
