Banyan is a managed distributed inference network: it runs batch and overflow AI workloads on ordinary, capable GPUs — the ones already sitting in homes and small racks — at roughly 90% below the price of a big cloud, paying the people who contribute that hardware in real money.
The companies building AI cannot get compute fast enough. Hundreds of billions of dollars are committed, but the physical world can't keep up — the chips, the memory, the high-voltage transformers, and the electricity are all constrained at the same time. New data centres are delayed for years by parts that cost a rounding error of the total. Meanwhile, enterprises waste most of what they already have: independent research puts average GPU utilisation around 5%. So the bottleneck isn't only supply — it's mismanagement of the supply that already exists.
Banyan is the coordination layer that turns a chaotic supply of idle, mismatched hardware into something that behaves like reliable infrastructure. It benchmarks each machine, routes work to the right one, checkpoints long-running jobs, verifies the results, and pays contributors in fiat for verified output.
A big cloud is the flagship store — full price, newest hardware, white-glove, one brand. Banyan is the factory outlet: the same models at a fraction of the price, no frills, many brands under one roof. Nobody expects white-glove service at an outlet — that's exactly why it's cost-efficient, and why the trade-off is understood and accepted.
Every job run through Banyan adds to the only dataset in existence mapping which model runs best on which hardware at which cost. A chip maker can only benchmark its own silicon; a cloud only what it stocks; a model lab has no diverse fleet. Only a neutral aggregator with mixed hardware can answer that question — and the dataset compounds with every new model and every new chip added to the pool.