Prime Intellect

Prime Intellect is a San Francisco based GPU cloud and distributed training platform that positions itself as an "open superintelligence stack," combining an on-demand and reserved GPU marketplace with an in-house training and inference framework. Rather than owning a fleet of datacenters like CoreWeave or Nebius, the company aggregates capacity from more than fifty third-party datacenter operators and resells it through a single API, competing on price, geographic breadth, and open research artifacts rather than hyperscale contracts.

The company drew attention in late 2024 with INTELLECT-1, a 10 billion parameter model trained across up to 112 Nvidia H100s split between five countries and three continents using its PRIME framework, an implementation of the DiLoCo distributed training approach with a custom int8 all-reduce that cuts communication bandwidth by roughly 400x. That research pitch, that useful models can be trained across heterogeneous, non-colocated clusters, is Prime Intellect's core differentiator against the "one giant campus" story told by CoreWeave, Crusoe, and Nebius.

Its customer roster leans toward AI-native application companies and open-model labs, including Ramp, Zapier, Character.ai, Goodfire, Browserbase, and Standard Intelligence. Backers include Founders Fund, Radical Ventures, Nvidia, and Intel, alongside angel investment from Andrej Karpathy and John Schulman. Prime Intellect competes at the long-tail and mid-market end of the neocloud pack rather than for hyperscaler training contracts.