Distributed AI computing gains traction as companies turn to homes for model training

A growing movement from open-source and commercial entities aims to rebalance AI’s heavy infrastructure demand by harnessing idle consumer hardware, promising cost savings, enhanced privacy, and resilience amid data centre shortages.

The rapid expansion of artificial intelligence is putting fresh pressure on the technology sector. More data centres are needed to train and run large language models, and that demand is already feeding shortages in components such as memory chips while pushing up power consumption. As the strain grows, a counter-movement is emerging from the open-source world: instead of concentrating compute in vast facilities, it is trying to build a distributed alternative from idle home computers and small servers.

That model is not new in principle. IBM’s World Community Grid and the volunteer project SETI@Home both relied on spare processing power from ordinary users for scientific work. What is changing now is the commercial logic. Rather than giving unused capacity away, some operators are proposing that owners rent out their spare GPUs and processors so that smaller AI developers can run inference and, in some cases, parts of training more cheaply than they could in a conventional cloud environment.

Several companies are now backing that idea. Evolving Edge, based in Austin, is in beta with a platform built around residential hardware. Far Labs says its FAR AI system will soon open to users from a waiting list, while Bless, Salad and Gradient are among the other names pushing distributed compute as a practical alternative to centralised infrastructure. Their aim is to assemble enough nodes to offer on-demand processing that is always available, even if no single device is powerful enough to rival a modern hyperscale data centre.

The pitch has both technical and economic appeal. According to Evolving Edge, spreading workloads across existing consumer hardware can reduce latency and costs by routing tasks to nearby devices. Far Labs says developers can use a simple API to run AI inference on its network, with setup taking only a few minutes and no installation fee. The company also says its software follows a “least privilege” model and that the source code has been released for community review. It claims users keep access to their own files and local software, while the network uses only the underlying hardware.

There is also a revenue angle. Far Labs says participants can earn money from unused capacity, although it notes that payouts depend on demand, uptime, pricing and network rules. The company gives one example of about $4.41 a month, while also suggesting a much higher return relative to electricity costs. Gradient, meanwhile, describes a three-layer system: a communication layer for fast peer-to-peer networking, an execution layer that breaks inference into smaller parts for parallel processing, and a learning layer designed to support reinforcement learning in real time.

Privacy and resilience are central to the sales pitch, but they are also the main test. The distributed model can, in theory, limit exposure of sensitive data by splitting workloads and keeping only the minimum necessary information on each node. It may also offer some protection against outages or attacks by avoiding a single point of failure. Yet it remains to be seen whether a network built from consumer machines can scale far enough, or prove reliable enough, to compete consistently with the large data centres now dominating the AI economy.