Solana just became the backbone of decentralized AI training. Here is the plan.

Nous Research announced the next phase of Psyche, its open infrastructure project for decentralizing AI training. The new phase introduces a peer-to-peer coordination layer built on the Solana blockchain. This enables training jobs to be distributed across underutilized GPUs from gaming rigs to idle cloud instances without a central orchestrator. The announcement positions Psyche as a direct challenge to the cluster-as-a-service model that dominates large-scale pretraining, a landscape where most teams rely on a handful of providers, as previous analysis of GPU hopping costs shows.

The core design: a Solana smart contract acts as a decentralized job board. Node operators submit GPU availability, pricing, and performance attestations. Training requesters submit job specifications covering model architecture, batch size, and expected wall time along with a SOL deposit. The contract matches jobs to nodes, escrows payment, and releases payment only when the training epoch is attested to the contract's satisfaction via a proof-of-contribution mechanism. The architecture is inspired by Nous Research's earlier DisTrO project, which focused on communication efficiency. Psyche extends it to economic coordination, echoing the approach of other decentralized infrastructure plays like Ollama's $88 million bet on open-weight models.

Phase 1 of Psyche launched in late 2024 and focused on the communication layer: how to shard gradient exchange across a Byzantine-fault-tolerant network. Phase 2 adds the Solana-based coordination layer and a lightweight attestation protocol that does not require repeated on-chain verification for every training step. The team reports a test deployment with 48 nodes using 192 consumer-grade GPUs achieving 85% of the throughput of an equivalent cluster with all nodes in the same rack. The 15% penalty comes from network latency and attestation overhead, a reminder that counting human hours also means accounting for coordination slack.

The practical significance: training large models currently requires access to dedicated clusters or spot instances from a few cloud providers. Psyche's model would let an academic lab aggregate idle cycles from a university's computer science department without building a shared filesystem or trusting a central scheduler. Payment in SOL tokens creates incentive alignment: node operators earn for contributed compute, requesters pay only for successful training steps. This shift from centralized GPU economies to open networks is part of a broader push, similar to how Kimi Sheets turns plain language into formulas rather than just explaining them, a move from promise to execution.

Risks abound. Decentralized training at scale has never been demonstrated for models larger than 7B parameters. Communication overhead, variable node latency, and malicious node detection are unsolved at 1000+ node scale. The Solana blockchain can handle high transaction throughput but at a cost that may escalate if the coordination layer becomes a bottleneck. Nous Research positions Psyche as research infrastructure, not production-ready. The team invites third-party audits of the attestation protocol. These are the same kinds of scaling questions that vibe coding faces when moving from prototype to production, the hard part is shipping at scale.

Still, the ambition marks a new front in the decentralization of AI infrastructure. If Psyche works at scale, it redefines who can train a large model.