The Idea

In early 2024, Vincent Weisser saw a canyon running through AI. On one side, university researchers with ideas and no compute. On the other, a handful of Silicon Valley labs training frontier models behind locked doors. There was no middle ground, he told Fortune.

Weisser was an unlikely person to bridge it. He came out of crypto and decentralised science, co-founding the longevity DAO VitaDAO. His co-founder, Johannes Hagemann, was an AI research engineer who had worked on distributed training at Germany's Aleph Alpha. Together they held a heretical thesis: you didn't need one giant data centre to train a serious model. You could stitch together idle compute scattered across the world and train across it.

The sharper insight came later. Pre-training had concentrated AI inside a few labs, but the part that actually made a model useful for a specific job, the post-training loop, the reinforcement learning that tunes a model to your workflow, was a black box nobody outside could touch. The labs owned the optimisation loop. Everyone else just rented the output.

Prime Intellect decided to hand that loop back. This was the beginning of Prime Intellect.

The Execution

• April 2024: Weisser and Hagemann raised a $5.5M seed led by Distributed Global and CoinFund, and launched a platform that pooled surplus global compute for open model training.
• October 2024: They shipped INTELLECT-1, the first 10B-parameter model trained across geographically distributed compute, with Hugging Face and Arcee among the contributors donating GPUs. Proof the decentralised thesis actually worked.
• February 2025: A $15M round led by Founders Fund pushed total funding past $20M. Andrej Karpathy came in as an angel.
• August 2025: Launch of the Environments Hub, an open, community-run library of the RL environments where agents learn. Distribution first: it earned trust from the exact engineers who would later sign off on procurement.
• Early 2026: The pitch hardened into an attack on the status quo. RL breaks the frontier open, the company argued, letting any enterprise own its model optimisation loop instead of renting an API. It pulled in teams already sick of vendor lock-in.
• July 8, 2026: A $130M Series A at a $1B valuation, led by Radical Ventures with NVIDIA Ventures, Intel Capital and Dell Technologies Capital, taking total funding past $150M.
• Today: Over 6,000 customers and a $100M annualised revenue run rate, with Ramp, Zapier and Arcee on the logo wall. Ramp trained its own subagent on the stack and, by its own account, beat the frontier models on accuracy at a fraction of the cost.

The lesson?

Prime Intellect never sold the technology. It sold the control problem its customers already felt. And it sequenced the proof: open-source credibility earned the developers' trust, sharp positioning pulled in the teams frustrated with lock-in, named logos like Ramp and Zapier closed the deals, and only then did the institutional money arrive. Six thousand customers first, then the cheque. Most infra startups pitch the demo. The ones that win pitch the pain the buyer can't stop feeling.