Prime Intellect

Prime Intellect operates what it calls the "Open Superintelligence Stack": an integrated platform spanning GPU compute, reinforcement learning environments, evaluations, sandboxes, hosted training and inference, intended to let organizations train, deploy and continuously improve their own models and agentic systems rather than depend on closed frontier labs. The product surface includes Prime Lab (post-training of agents), an Environment Hub with more than 2,500 open-source RL environments, hosted evaluations covering 100+ open-source models with a public leaderboard, managed large-scale training for agentic workflows, and dedicated, serverless and LoRA-based inference for custom models. Secure sandboxes provide isolated code execution for large-scale RL.

On the compute side, the company offers on-demand access to 1–256 GPUs across clouds with SLURM and Kubernetes orchestration, InfiniBand networking and Grafana monitoring, plus "Liquid Reserved Clusters" that solicit quotes from more than 50 datacenter providers within 24 hours and allow customers to resell idle GPUs into a spot market. Listed hardware includes H100, H200, B200, B300, GB300 NVL72, GH200, A100, A40 and RTX Pro 6000, with published hourly rates. Access is via the open-source `prime` CLI and Python SDK, alongside the company's Verifiers library for RL environments and evals, the prime-rl asynchronous RL training framework, and prime-agent, a self-improving agent for coding and long-running autonomous tasks.

Earlier work included the Compute Exchange GPU marketplace, the PRIME framework for globally distributed training, INTELLECT-1 (a 10B-parameter model trained across three continents), METAGENE-1 (a bio-foundation model for pandemic detection), INTELLECT-MATH, the GENESYS and SYNTHETIC-1 distributed RL and synthetic reasoning dataset efforts, and TOPLOC, a locality-sensitive hashing scheme for verifiable inference.

Founding story

Founded in 2024 with the aim of enabling organizations to train their own agentic systems without relying on frontier AI labs, a goal management frames as extending model-training capability beyond a small number of San Francisco labs to every enterprise and nation state.

Business model

Prime Intellect monetizes a modular, marketplace-style platform: customers pay for hosted versions of its tools and for GPU compute. Coverage describes modular access so buyers can select individual components rather than commit to the full stack.

Revenue comes from hosted product usage — GPU compute rented by the hour at published per-GPU rates, reserved clusters, hosted training runs with capacity- and token-based pricing, hosted evaluations, and inference sold as dedicated deployments, pay-per-token LoRA serving and serverless APIs.

Traction

The company reports about 6,000 customers spanning AI startups, neolabs and enterprises, and annualized revenue exceeding $100 million as of July 2026. The Environments Hub hosts over 2,500 community RL environments and hosted evals cover 100+ open-source models. Open-source repositories show substantial adoption: prime-agent ~19,100 stars, Verifiers ~4,600 stars, prime-rl ~2,000 stars and the prime CLI ~306 stars. Ramp publicly describes training a small RL-trained subagent ("Fast Ask") on Prime Intellect's Lab that it says beat frontier models on accuracy at lower cost and latency.

Latest developments

In July 2026 Prime Intellect announced a $130M Series A led by Radical Ventures with Nvidia Ventures, Intel Capital, Dell Technologies Capital, Iconiq and existing investors, reported at a $1 billion valuation. Proceeds are earmarked for larger compute clusters, larger RL runs, and infrastructure for agentic training, inference and continual learning, including long-horizon agents, recursive language models, automation of AI research and science, and continual learning.