Decentralised compute landscape: peers building distributed GPU training & inference networks
The 27 June 2026 All-In Podcast discussion on distributed (decentralised) computing — converting idle and surplus GPUs into shared, lower-SLA training and inference pools — maps onto a fast-growing cohort of projects beyond the names raised on the show. The panel's central caveat applies across all of them: distributed training degrades sharply once nodes are not physically adjacent, while distributed inference tolerates higher latency and is the nearer-term opportunity. The companies below are building the protocols, markets and training methods that make this viable.
Gensyn. UK-based decentralised machine-learning compute protocol founded in 2020 by Ben Fielding and Harry Grieve, connecting the world's idle GPUs/CPUs into a single verifiable training network — its core innovation is cheaply proving that a remote device executed an ML task correctly. Raised a $43M Series A led by a16z crypto in 2023, with CoinFund, Canonical Crypto, Protocol Labs and Eden Block.
Prime Intellect. San Francisco lab (founded 2023 by Vincent Weisser and Johannes Hagemann) running a Compute Exchange that aggregates global GPUs into a distributed marketplace, plus globally-distributed training — it trained INTELLECT-1, a 10B-parameter model, across three continents. Raised ~$20.5M total: a $15M Series A led by Founders Fund in March 2025 (Menlo Ventures; angels Andrej Karpathy, Clem Delangue, Emad Mostaque, Tri Dao), after a $5.5M seed.
Nous Research. Builds the Psyche network for decentralised LLM training over the open internet, using its DeMo/DisTrO optimisers to cut inter-node data transfer by orders of magnitude and coordinating runs on Solana — it ran what it bills as the largest pretraining conducted over the internet to date (a 40B-parameter model).
io.net. DePIN-style decentralised GPU network aggregating distributed compute supply into an on-demand cloud for AI and ML workloads.
Render. Distributed GPU network (token: RENDER) that pools idle graphics cards for rendering and, increasingly, AI/compute workloads — one of the earliest large-scale examples of monetising idle GPUs through a permissionless market.
Read more: All-In Podcast (distributed-compute segment, 1:21:46) · Decrypt · Prime Intellect · Nous Research
