Prime Intellect Raises $130M Series A So Enterprises Can Train Their Own Agents
Prime Intellect announced a 130 million dollar Series A at a 1 billion dollar valuation on July 8, led by Radical Ventures with Nvidia Ventures, Intel Capital, Dell Technologies Capital, Iconiq, and angels including the founders of Perplexity, Box, Harvey, Cognition, and Mercor. Founded in 2024, already at a 100 million dollar annualized revenue run rate, with customers like Ramp and Zapier.
The pitch: stop renting your agents from a frontier lab. Prime Intellect sells the full stack for building your own β compute, a reinforcement learning framework, and evaluation tools, packaged as a modular marketplace. One customer claims their self-trained agent beat the frontier models on accuracy while running faster and at a fraction of the cost.
Put this next to Bespoke Labs' 40 million dollar Series A five days ago and the thesis is now unmistakable: the money is moving to the training grounds, not the models. Bespoke sells environments where agents learn safely; Prime Intellect sells the compute and RL loop to train them. Two rounds, one week, 170 million dollars, same bet β that enterprises will want to own the agent, not just the subscription.
The interesting tension: this bet only works if the harness-plus-RL recipe keeps closing the gap with frontier models on narrow tasks. The published research (harness optimization, on-policy distillation, async RL) says it does. Nvidia and Intel both writing checks into the same round says the silicon vendors believe it too.
More: https://techcrunch.com/2026/07/08/prime-intellect-raises-130m-series-a-to-help-enterprises-build-their-own-ai-agents/
← Back to all articles
The pitch: stop renting your agents from a frontier lab. Prime Intellect sells the full stack for building your own β compute, a reinforcement learning framework, and evaluation tools, packaged as a modular marketplace. One customer claims their self-trained agent beat the frontier models on accuracy while running faster and at a fraction of the cost.
Put this next to Bespoke Labs' 40 million dollar Series A five days ago and the thesis is now unmistakable: the money is moving to the training grounds, not the models. Bespoke sells environments where agents learn safely; Prime Intellect sells the compute and RL loop to train them. Two rounds, one week, 170 million dollars, same bet β that enterprises will want to own the agent, not just the subscription.
The interesting tension: this bet only works if the harness-plus-RL recipe keeps closing the gap with frontier models on narrow tasks. The published research (harness optimization, on-policy distillation, async RL) says it does. Nvidia and Intel both writing checks into the same round says the silicon vendors believe it too.
More: https://techcrunch.com/2026/07/08/prime-intellect-raises-130m-series-a-to-help-enterprises-build-their-own-ai-agents/
Comments