Prime Intellect Raises $130 Million to Help Enterprises Build Their Own AI Agents

Prime Intellect, the startup betting that any company should be able to train its own AI rather than rent intelligence from a handful of frontier labs, has raised a $130 million Series A at a roughly $1 billion valuation. The round, led by Radical Ventures, vaults the two-year-old company into unicorn territory and puts a nine-figure bet behind one of the year's most consequential enterprise questions: who actually owns the AI agents doing your company's work?

The financing drew a striking roster of backers, including Nvidia Ventures, Intel Capital, Dell Technologies Capital, and Iconiq, alongside a long list of founder-angels who between them run some of the most closely watched companies in software: Aravind Srinivas of Perplexity, Aaron Levie of Box, Winston Weinberg of Harvey, Jeff Wang of Cognition, and Brendan Foody of Mercor. That the people building on top of frontier models are personally funding a company designed to help enterprises route around those same models is, itself, a signal worth reading.

From decentralized research lab to enterprise "one-stop shop"

Founded in 2024 by Vincent Weisser and Johannes Hagemann, Prime Intellect first made its name in the open-source research community, not the enterprise sales channel. Its INTELLECT-1 project trained a 10-billion-parameter model across GPUs scattered over five countries and three continents. It followed with INTELLECT-2, a 32-billion-parameter reasoning model trained through fully asynchronous, "globally decentralized" reinforcement learning across a permissionless swarm of compute contributors. To make that work, the team built infrastructure from scratch, including PRIME-RL, a distributed RL training framework, plus components like TOPLOC to verify results from untrusted workers and SHARDCAST to broadcast updated model weights across the network. All of it was open-sourced.

That research pedigree is the foundation for the commercial pitch. Prime Intellect now sells what it describes as a "full-stack" for building AI agents: compute access, a reinforcement learning framework, sandboxed environments, and evaluation tools, offered in a modular, marketplace-style format so customers can take only the pieces they need instead of buying into an all-or-nothing platform. The rise of reinforcement learning, which rewards successful task completion and penalizes errors, is central to the thesis. It is what allows a company to become, in effect, its "own AI lab," refining models against the specific tasks its business actually runs on.

"They've stitched this together and built it in such a way that they're operating at the frontier in a way that's affordable," said David Katz, a partner at Radical Ventures, who framed the company as offering the capabilities of a top-tier AI lab as a "one-stop shop." The complexity is the moat: bypassing closed labs is technically possible today, but assembling the underlying infrastructure into a production-ready system is beyond most enterprises. Prime Intellect's wager is that it can sell the assembly.

The "build your own agents" trend has real revenue behind it

The numbers suggest the pitch is landing. Customers including Ramp, Zapier, and Flapping Airplanes pay for hosted versions of the tools, and the company says it has reached an annualized revenue run rate of $100 million, an unusually steep curve for a firm barely two years old. The proof points are specific: Ramp used Prime Intellect to build an agent for finding answers inside spreadsheets. "The result beat the frontier models on accuracy while running at faster speeds and a fraction of the cost," Ramp co-founder and co-CEO Karim Atiyeh said in a statement.

That last line captures why "build your own agents" is graduating from ideology to procurement decision. For a growing set of narrow, high-value tasks, a smaller model tuned with reinforcement learning against a company's own data can outperform a general-purpose frontier model, and do it cheaper and faster. The other driver is risk. Enterprises are increasingly wary of pumping proprietary data into OpenAI or Anthropic, and of building critical workflows on models that can be deprecated or switched off with little notice, as happened recently when Anthropic pulled its Fable model.

"How do I know that I'm not working with a company that is going to try to replace me and generalize to what I'm doing," Katz said, summarizing the anxiety pushing buyers toward ownership. Weisser puts the mission in more populist terms. "It shouldn't just be a few nerds in a glass tower in San Francisco that have the capability to train AI models," he told TechCrunch. "It should be every enterprise, every nation state."

What to watch

The tension in Prime Intellect's story is the gap between its decentralized-training roots and its enterprise reality: most paying customers want a reliable hosted product, not a permissionless GPU swarm, and it remains to be seen how much of the open, distributed vision survives contact with enterprise procurement and SLAs. Watch whether that $100 million run rate holds as frontier labs push their own fine-tuning and agent-building tools, and whether Prime Intellect's "AI sovereignty" framing resonates beyond early adopters into regulated industries and the nation-state customers Weisser name-checks. The $130 million buys runway; the open question is whether owning your intelligence becomes a mainstream enterprise default or stays a specialist's choice.