1 Startup · $130M · Open AI Training Stack

What they're building

Prime Intellect sells the capabilities of a frontier AI lab as modular infrastructure: compute, reinforcement-learning environments and sandboxes, post-training tooling, inference, and continual-learning deployment, so enterprises can train and own custom models and agents for their specific workflows instead of renting closed frontier models forever. The platform works like a marketplace — customers take only the pieces they need. Ramp used it to train a small RL subagent that finds answers inside spreadsheets, which the company says beat frontier models on accuracy at higher speed and a fraction of the cost; Zapier is also cited as a customer. The next bets are explicitly stated: long-horizon agents and Recursive Language Models that manage their own context and coordinate sub-agents where today's models break down.

Why this matters

This is the infrastructure-layer counter-bet to closed-lab dominance, and the cap table telegraphs it: three chipmaker venture arms (NVIDIA, Intel, Dell) benefit directly when training is distributed across many buyers rather than concentrated in a few labs. Radical's partner argues Prime Intellect uniquely packages top-tier-lab capability as a one-stop shop that's affordable at the frontier. CEO Vincent Weisser's stated conviction is that model-training capability shouldn't belong to a few people in a glass tower in San Francisco — it should be available to every enterprise and nation-state. The traction backs the rhetoric: roughly $100M annualized revenue within two years of founding, a reported $1B valuation, and reference customers (Ramp) publicly saying their trained-own-model beat the frontier alternative. If enterprises keep choosing to build rather than rent, this is one of the defining infrastructure companies of the cycle.