Training a frontier AI model typically requires tens of thousands of GPUs humming in unison inside a single data center owned by a company with a market cap larger than most countries. Prime Intellect thinks that’s a problem worth solving.
The startup, founded by Vincent Weisser and Johannes Hagemann, is building a decentralized platform designed to aggregate computing power from around the globe and channel it into training, evaluating, and deploying large language models. Think of it as turning the world’s idle hardware into one giant, distributed supercomputer, except the machines don’t need to be in the same building, the same country, or even the same continent.
What Prime Intellect has actually built
Plenty of projects promise decentralized AI infrastructure. Most of them are still at the whitepaper stage. Prime Intellect has already run two major training experiments, which puts it meaningfully ahead of the curve.
The first, called INTELLECT-1, produced a 10-billion-parameter language model trained entirely on globally distributed compute. The company describes it as the first open-source decentralized LLM training run conducted at scale. For context, 10 billion parameters is not trivially small. It’s roughly the size of models that were considered state-of-the-art just a few years ago.
