Prime Intellect raises $130M at $1B valuation to build open sovereign AI stack
Key Points
- Prime Intellect raised $130 million at a $1 billion valuation, reaching north of $100M ARR on less than $20 million in prior funding, with Nvidia joining as a strategic investor.
- The company operates ~15,000 GPUs and is scaling toward 30,000+, with compute scarcity through 2025 identified as the primary growth bottleneck for enterprise and sovereign AI deployments.
- Prime Intellect pitches its full-stack platform to enterprises and governments seeking to own their AI model infrastructure outright, building proprietary training signals over time rather than relying on API access.
Summary
Read full transcript →Prime Intellect raises $130M at $1B valuation to build open sovereign AI stack
Prime Intellect has closed a $130 million Series A at a $1 billion valuation. The company had raised just $20 million prior to this round, meaning it reached its current run rate — north of $100M ARR — on less than $20 million in total spend. Nvidia joined as a strategic investor.
Vincent Weisser, co-founder and CEO, says the company has since doubled its pricing and is now tracking above that $100M figure.
What Prime Intellect sells
The pitch is a full-stack platform for training, deploying, and continuously improving AI models, sold as an alternative to building that infrastructure in-house. Customers range from AI-native startups like Ramp to newer AI labs to traditional enterprises and sovereign buyers who want an end-to-end stack they can run on-premise.
Weisser's clearest product proof point is Ramp's result: within one week and under $50,000 in training spend, Ramp fine-tuned an open model that outperformed Claude Opus on automating spreadsheets and finance tasks, at a fraction of the cost of running Haiku and 30% faster. The underlying logic is that a large frontier model handles orchestration and planning, while specialized sub-agents trained on specific use cases handle execution — cheaper and more accurate than routing everything through a general-purpose model.
“We've announced that we've raised a 130,000,000 and 1,000,000,000 with Prime Analytics to build the open super intelligence stack. We doubled the pricing since we actually closed the fundraise to a 100 [ARR]. So basically, yeah, on track to grow much more from here on. Right now, we run like 15,000 GPUs and scaling to like over 30,000.”
Compute scale and GPU constraints
Prime Intellect currently runs ~15,000 GPUs and is scaling toward 30,000+. The bulk of the fresh capital is going toward securing more capacity. Weisser says compute is the primary growth bottleneck right now — GPUs are effectively sold out through the rest of 2025, particularly for US-based data center buildouts. The company abstracts cluster management away from customers, letting a startup with a $50K training budget access thousands of GPUs for a short burst without making a multi-million dollar commitment.
That abstraction was built from day one: Prime Intellect optimized for fault-tolerant training across distributed, geographically dispersed compute, which Weisser argues gives them an advantage in sourcing from every available pocket of supply globally.
The sovereign AI thesis
"Sovereign" in Weisser's framing covers both national governments and Fortune 500 companies that want to own their model stack outright. The motivation is compounding: enterprises that run continuous RL loops on their own agents accumulate proprietary training signal over time, building a data moat that generic API access can't replicate. Weisser compares the trajectory to Tesla's autonomy levels, applied to knowledge-worker agents.
Open source outlook
Prime Intellect is a partner in Nvidia's Nematron Alliance, and Weisser says he's closely involved with both Nematron and Aussie's Trinity models as the leading American open-source efforts. He expects a surge in frontier open model releases from US-based labs, partly driven by the policy risk around Chinese open-source models.
Weisser says the steady stream of open releases from China is no longer reliable — companies like Moonshot have already changed their licensing and pulled back toward closed models, and state intervention could accelerate that. He also points to Poolside as an example of a lab expected to stay closed that instead opened its models, and argues more neo-labs will follow as the economics of open releases become clearer.
Data center geography
Outside the US, Weisser is watching Australia, New Zealand, Armenia, and Kazakhstan as active new entrants in data center buildout, citing cheap energy, available land, and US alignment. Within the US, he sees Texas as the state leaning in hardest. He acknowledges domestic opposition to data center construction is real but expects both tracks — US and international — to expand, with compute scarcity remaining the binding constraint on Prime Intellect's own growth through at least the end of the year.
Every deal, every interview. 5 minutes.
TBPN Digest delivers summaries of the latest fundraises, interviews and tech news from TBPN, every weekday.